A few months ago I found myself working across two computers at once, four projects running in parallel. But here’s what that actually looked like: each project wasn’t just an agent doing a task. A research assignment would spin up its own team — sub-agents fanning out across different angles of the question, each pulling sources, testing claims, flagging contradictions — all reporting back to an orchestration agent that synthesized their findings into a research brief I could read and redirect. An analysis task meant one agent ingesting a document set, another extracting patterns, a third stress-testing the emerging conclusions from a contrarian point of view, the whole thing converging into a report with a summary of what held up and what didn’t.

I was reading finished deliverables, marking up what needed to change, and issuing new direction.

That’s when I realized I was doing something I’d done before. In a previous life, I was a creative director. My job was to review designers’ work, mark it up with red pen, and hand it back with clear direction on what to fix. I wasn’t designing. I was directing. Multiple designers, multiple projects, work coming to me rather than me doing the work.

That’s exactly what I was doing now. Except my team was made of agents.

The distinction most people miss — and why it’s hard to see

Most leaders I talk to who use AI constantly — who are genuinely productive with it — haven’t made this shift. And the reason isn’t that they’re not trying hard enough. The gap between where they are and where they need to get is harder to see than it sounds.

Here’s how most people use AI today: you paste context into a chat window, describe what you need, and work through the output together. The AI drafts something. You tell it what’s missing. It redrafts. You refine the prompt, steer the next iteration, copy the final result. You were in the room for every paragraph. What you produced, you co-produced through conversation.

That’s a chatbot. And it’s useful. But it’s not directing agents.

When you direct agents, you write a brief. You specify the outcome, the standards, the sources to consult, the criteria for good work. You assign it. You come back later to a deliverable — a research report with cited sources, tested conclusions, a summary of what the agent found and what it couldn’t verify. You weren’t there for any of the middle. A team of sub-agents worked through it, reported to an orchestrator, and the orchestrator produced something you can now review.

The difference isn’t subtle. In chatbot mode, the AI helps you do your work — and you are the ceiling. Your hours are the constraint. Every output requires your presence. In agent mode, agents do work you assigned and bring you the output for approval. The ceiling is how many agents you can brief and review, not how many hours you have. One of these scales. The other doesn’t.

Most people who think they’re working agentically are in chatbot mode. Extended conversations, iterative refinement, sophisticated prompting — it feels advanced. It is advanced. It’s just not the same thing.

What actually makes an agent autonomous

Here’s the question worth sitting with: why does chatbot mode require constant supervision while agent mode doesn’t?

It’s not a capability difference. It’s a context difference.

A human employee doesn’t need you to explain the situation from scratch every morning because they bring context with them. They know the project, the client, the company’s standards, the decisions that were made last week. They fill in gaps you didn’t specify because they’ve absorbed how your organization works. That embedded context is what makes autonomous work possible.

Without that context, an AI operates the same way a new hire on day one does — it needs everything explained. Which is exactly what chatbot mode is. You’re not directing; you’re supervising every step, because the AI only knows what you’ve told it in this conversation. The moment you close the window, the context disappears.

The AI Operations framework solves this by encoding context into the system itself. The workspace gives agents project history — what happened in previous sessions, what decisions were made, what the current state of the work is. The knowledge layer gives agents institutional memory — patterns discovered, standards that apply, how your organization thinks about problems. An agent working inside this structure can pick up a task, understand where it sits in the larger project, and execute without being walked through it.

This is the mechanism that turns babysitting into direction. Context is what creates agent autonomy. And context is something you can build.

Where most high-adopters get stuck

Here’s the uncomfortable truth: if you’re using AI primarily in chatbot mode, you’re working harder than you need to — and your individual gains almost certainly aren’t reaching your team.

The data confirms this is widespread. McKinsey’s 2025 State of AI report found 88% of organizations use AI in at least one business function — but only 23% are scaling agentic AI anywhere in the enterprise. BCG’s research found just 5% of companies qualify as genuinely “future-built” for AI — organizations where the operating model has been redesigned around it, not just where tools have been distributed. Three separate measurements, same direction: most organizations are stuck at the tool layer, nowhere near the operational layer.

The barrier is almost never the model. As one AI operations practitioner put it in this analysis of human+AI workflows: “AI adoption doesn’t fail because the models are weak — it fails because the operating model is vague.” Individual gains don’t compound into team results because what’s missing is shared context — a system for it. In a 2024 survey of software engineers, 40% cited “trouble finding context” as their most common productivity pain point. That’s what happens when everyone prompts from scratch, saves output somewhere private, and the learning disappears into the next task. The gap between you and your team isn’t a training problem. It’s an infrastructure problem.

What this looks like when it works — even in the messiest work

I want to say something about scope, because the obvious objection is that this works for structured, repeatable tasks but falls apart in ambiguous, judgment-heavy work. Mine isn’t structured or repeatable. I work in innovation. Every problem is distinct, every project has a different shape, and I rarely do the same thing twice. What I have instead of repeatable tasks are methods — research standards, analysis frameworks, ways of testing assumptions, approaches to building proof-of-concepts that don’t devolve into vibe coding.

This model works in that environment because it encodes methods and judgment, not just task templates. I have research agents that operate under strict validation guidelines so the output can be trusted. I have agents that pattern-match across large bodies of information and surface what I wouldn’t have found myself. I have agents that approach emerging conclusions from a contrarian point of view — built to challenge, not confirm. I have agents that build prototypes methodically, step by step, with validation at each stage.

My work spans deep research, pattern analysis, framework development, proof-of-concepts, and go-to-market strategy across R&D, engineering, and business. In that environment — non-repeatable, wide-ranging, genuinely ambiguous — this operating model has accelerated my outcomes 14.25X. Not because the AI is doing the thinking. Because the AI is doing the work while I’m doing the directing.

Where to start

Don’t buy another AI tool.

Think about the next time you’ll sit down to write a research summary, a status report, or a client brief. That’s the workflow. Document what good output actually looks like. Build a simple workspace that gives an agent the context it needs. Define the review step — what you check before anything goes anywhere.

Run it three times. Fix what breaks. Log what you learned. That’s the first layer, and every workflow you build on top of it starts smarter because the context from the last one is already there. It compounds.

The companies BCG identifies at the frontier — the 5% creating substantial AI value — aren’t running better models than everyone else. They’ve redesigned how work gets done: reusable agents, shared knowledge platforms, structured architecture, leaders who direct rather than just use. They’ve crossed from AI-assisted to AI-operated.

The leaders on the other side of that crossing don’t work more hours than you. They’ve redesigned what they do with those hours. Their job is direction. Their agents do the work. That’s not a productivity hack. It’s a different operating model — and it’s available to any team willing to build it.

If this was useful, share it with someone who’s using AI a lot but hasn’t quite made this shift. And if you have — I’d love to hear what made it click for you. Reply here or drop a comment.

Over 400,000 jobs eliminated in tech sector restructurings since 2022. More than eleven thousand people retiring every day, taking decades of institutional judgment with them. And a national hiring rate of 3.1%, matching the lowest point recorded during COVID lockdowns, without a lockdown.

These aren’t three separate problems. They are one problem with three symptoms. The economy is reorganizing around human judgment at the exact moment the infrastructure for moving human judgment has ceased to function.

Here is the specific shape of that failure. AI is automating the execution layer of work: the tasks, workflows, and repeatable processes that credentials used to certify. What’s left is judgment: contextual reading, taste, creative recombination, the ability to identify which problems are worth solving. The retirement wave is simultaneously draining the people who have the most of it. Companies are hemorrhaging the most valuable thing they have at the same time AI is making it the primary source of human economic value.

And the system meant to replace that judgment, to route it from where it exists to where it’s needed, is frozen.

Not from economic contraction. GDP is positive, job openings sit at 7.4 million. The freeze came from an arms race. Employers deployed AI to screen candidates. Candidates deployed AI to generate applications. The result is a doom loop: more automation producing less signal, a national hiring rate at COVID lockdown lows, and the average job posting now drawing 244 competing applications, up 111% from 2022. The stalemate isn’t a symptom of economic weakness. It’s what happens when an AI arms race meets infrastructure designed in 1955.

The system was never built to identify judgment. It was built to sort resumes. Those were different problems when execution was the value. They’re the same problem now.

The Fourth Wave

Every prior wave of disruption in hiring told itself the same story: new technology, better outcomes.

The internet digitized the resume. LinkedIn created passive candidate sourcing. ATS platforms promised efficiency at scale. Each wave changed the medium and left the underlying logic untouched. The resume was still the primary data structure. Keyword matching was still the core screening mechanism. The job description was still aspirational fiction dressed as a spec.

The numbers confirm the pattern holds. Cost-per-hire for executive roles is up 21% since 2022 and 113% since 2017. Fifty-three percent of job seekers experienced ghosting in the past year. The hiring process has lengthened from roughly 12 days in 2010 to more than 42 days today. More automation, worse outcomes.

AI is the fourth wave: automating the same flawed process at higher speed. The result isn’t better hiring. It’s faster noise. And the noise is accumulating at the exact moment the stakes have changed. What the system is failing to route is no longer interchangeable labor. It’s judgment. The consequence of that failure is no longer friction. It’s loss.

The Market That Never Got Built

To understand why four waves of disruption haven’t fixed this, you have to name what’s actually missing.

Every major market in the modern economy eventually developed three things: price discovery, transparency, and liquidity. Financial markets got them. Real estate got them, imperfectly but meaningfully. Advertising got them: programmatic buying is essentially real-time clearing for audience attention. Even commodity markets developed standardized grades and exchanges that made price discovery possible across enormous distances.

The labor market never did.

On the candidate side: a prose document optimized for human scannability in 1955, crammed with self-reported claims and unstandardized language. On the employer side: a job description written to a title rather than a function, “Senior Product Manager” meaning wildly different things at a twelve-person startup and a Fortune 500, both using the same words.

The opacity wasn’t accidental. Information asymmetry is a source of pricing power, and it historically favored employers. Recruiters built entire businesses on proprietary candidate databases, what one analysis called “information arbitrage.” LinkedIn partially dissolved that arbitrage by bringing profiles online, but preserved the underlying logic. Profiles are still prose. They still resist machine matching at scale. And opaque criteria don’t just create inefficiency: when no one can see what’s being evaluated, discrimination has nowhere to be contested.

This is what a market without infrastructure looks like: participants guessing, intermediaries extracting value from information gaps, and the most important asset in the exchange, judgment, invisible to the very system trying to price it.

Both Sides Are Flying Blind

The dominant narrative frames this as an applicant problem. It isn’t. Employers are equally lost, and their primary instrument fails in exactly the same way.

A job description is aspirational fiction. It describes the person who filled this role before, filtered through the assumptions of whoever wrote it, inflated with requirements that sound rigorous but aren’t. Thirty-five percent of “entry-level” postings currently require three or more years of relevant experience. One in five employers intentionally leave unfilled roles listed, maintaining the appearance of growth while controlling headcount.

The degree requirement is the most expensive broken proxy in the system. University attendance was always a signal for something else: the capacity to learn, to persist through complexity, to navigate institutional environments. A reasonable approximation when demonstrated skills were hard to verify at scale. But research on predictive validity is plain: education predicts job performance at r = .10. One percent of variance. Laszlo Bock, who ran People Operations at Google while building one of the most rigorous hiring processes in the world, said it without ceremony: “GPAs are worthless as a criteria for hiring, and test scores are worthless, no correlation at all.”

Credential requirements don’t just fail to predict performance. They lock out workers who have the skills but not the paperwork. There are an estimated 70 million Americans who’ve built expertise through alternative routes: work experience, trades, self-teaching, community college. They’re excluded not because they lack capability but because the system has no mechanism to see capability, only credentials. The skills-based hiring movement was supposed to fix this.

And it has, mostly, announced itself.

Harvard Business School and Burning Glass Institute studied a decade of hiring data and found that employers dropped degree requirements in nearly four times as many roles between 2014 and 2023. Net result: fewer than 1 in 700 actual hires benefited. Ninety-seven thousand people out of seventy-seven million annual hires. Nine leading firms (Apple, Walmart, Cigna, ExxonMobil, GM, Koch, Target, Tyson, and Yelp) saw nearly 20% more non-credentialed hires and retention that ran ten percentage points higher. But across the broader market: nothing changed.

Joseph Fuller, who led the HBS research, was direct: “Changing your hiring policy is, at best, the end of the beginning.” Without changing the screening infrastructure, not the checkbox on the form but the operating logic underneath, it’s virtue washing.

Strip hiring to its essence: it’s a matching problem with five requirements. Complete information. Comparable units. Efficient search. Signal quality. A clearing mechanism. The current system fails all five. Resumes are self-reported prose. Job descriptions are invented vocabulary. Search now generates 244 applications per posting, up 111% in three years. Degrees and titles are weak predictors. Ghosting runs at record highs.

This isn’t a broken system. It’s the absence of a system, papered over with conventions that everyone agrees to pretend are functional. And the thing that absence makes invisible, at exactly the moment it matters most, is judgment.

What the Infrastructure Actually Looks Like

What needs to change isn’t the surface of the hiring system. It’s what the system is designed to measure.

The resume was built to sort execution: to compress a track record of tasks and titles into something a recruiter could scan in thirty seconds. That was a reasonable design when execution was the value. The infrastructure that replaces it needs to surface judgment: not what someone has done, but how they think, what contexts they’ve navigated, which decisions they’ve made and what happened.

Think of it the way we think about a credit profile: not a letter about your financial character, but a structured, verified, queryable record that any authorized party reads the same way. When Fair Isaac introduced the standardized credit score in 1989, the Federal Reserve described the impact plainly: it “was necessary for the emergence of large-scale open-ended consumer lending.” Before standardized scores, banks made lending decisions through personal interviews and character assessments: subjective, slow, biased toward whoever the loan officer already knew. Standardization didn’t just make lending faster. It made mass consumer credit structurally possible.

The labor market equivalent is structured work history as metadata: verified employment, documented decision types and contexts rather than job titles, skills demonstrated rather than claimed, outcomes where measurable. Not self-reported prose, a structured record that creates comparable units. Two data sets finding functional overlap, rather than two documents hoping a human can intuit the connection.

The job description gets the same treatment. Functional requirements weighted by importance, described in standardized capability language rather than title vocabulary. This forces employers to articulate what they actually need, which turns out to be clarifying independent of the matching problem. When you can’t hide behind “strategic thinker with strong communication skills,” you have to name the decisions the person will actually make.

The people most harmed by the current system are those without the time, resources, or institutional support to craft polished prose documents, workers returning after gaps, career changers, those who built expertise outside formal credential paths. A voice agent that interviews candidates about their work history converts spoken experience into structured metadata. This isn’t “AI rewrites your resume”: that just outsources the existing format’s logic to a model. Converting spoken experience into structured data meets people where they are, not where the system requires them to be.

Real-time relative scoring with transparent criteria closes the loop. Upon application, both sides receive signal: fit score, the criteria driving it, relative standing. Am I in the top quartile? What’s holding my score down? Is there a role two steps over where I’d be a stronger match? This creates a feedback mechanism the current system lacks entirely: both sides can see what’s being valued and adjust accordingly.

The interview doesn’t disappear. It gets better. Currently, interviews are doing the work the application process should have already completed: extracting signal the resume and job description failed to provide. In a system where structured data has already established functional fit, the interview returns to what only humans do well: assess culture, build mutual trust, make judgment calls about potential. That’s a more human process than the current one.

The mechanism that makes all of this stick is the incentive flip. Skills-based hiring failed at scale not because the idea was wrong but because the incentives never changed. ATS vendors profit from application volume. Recruiters are measured on time-to-fill, not match quality. Removing a checkbox doesn’t touch any of that, and as AI tools make applying cheaper and faster, volume will only grow. The current model gets structurally worse on its own trajectory.

A system that measures on longevity of placement rather than speed-to-hire restructures the revenue model entirely. Tie revenue to retention outcomes, whether the match was actually right, and every participant’s incentives realign. The Workday lawsuit (Mobley v. Workday, №3:23-cv-00770, N.D. Cal., conditionally certified May 2025) signals the legal environment is already moving toward accountability for screening outcomes. Build ahead of it.

The governance piece is what separates labor market infrastructure from surveillance capitalism: criteria must be published, auditing must be mandatory. Visible and contestable criteria make errors correctable. Hidden scoring is just a more sophisticated version of the existing opacity.

The Competitiveness Stakes

The argument so far has been about the mechanism. This section is about the cost of getting it wrong.

Labor market fluidity, the rate at which workers and capabilities move to where they create the most value, correlates directly with economic dynamism. Davis and Haltiwanger documented that US job reallocation rates fell more than a quarter after 1990. The cost of that friction compounds over time: BCG estimated global skills mismatch erased $8 trillion in GDP in 2018 alone; the US-specific projection runs to $2.5 trillion in lost output over the next decade.

What’s different now is the structural shift in who’s hiring. The retirement wave and AI restructuring are creating conditions for a significant move toward smaller, more specialized economic units, micro-firms and fractional arrangements where judgment is the product and AI handles execution. MBO Partners documents 5.6 million independent US workers earning over $100,000 annually, up from 3 million in 2020. LinkedIn profiles mentioning fractional roles grew from 2,000 to 110,000 between 2022 and early 2024.

This matters because small organizations fail differently than large ones. A mis-hire at a fifteen-person company is potentially existential, it cannot absorb months of searching, cannot staff a recruiting function, cannot carry the cost of a wrong match. These firms are the emerging unit of the judgment economy, and they currently have no hiring infrastructure designed for what they need.

The retirement wave need not mean permanent knowledge loss. Portable, queryable work histories create a new category: fractional access to pattern recognition that took decades to build. Not job search, the availability of judgment, on demand, at the moment AI has made judgment the scarce resource. The retiring expert’s knowledge doesn’t disappear. It becomes searchable.

Countries that build this capability pull ahead. Denmark moves more than 25% of its workforce across jobs every year while ranking seventh globally in GDP per hour worked. High mobility and high productivity reinforce each other when the matching infrastructure works.

We built highways to move goods. We built the internet to move information. We have not built the infrastructure to move human judgment at the speed the moment requires.

The Call

Everyone is diagnosing. Few are designing.

The resume had a good run. It solved the problem it was designed to solve: compressing a person into a human-scannable format when information was scarce and decisions were made by individuals with thirty seconds. That era is over. The information isn’t scarce. The decisions don’t need to be made by individuals scanning prose. The constraints that produced the artifact have dissolved.

What comes next is not a better resume, a smarter ATS, or an AI that screens faster. What comes next is market infrastructure: structured, verified, portable work histories; transparent and auditable matching criteria; real-time clearing that gives both sides the signal they need to make good decisions quickly; a revenue model tied to placement quality rather than application volume.

The companies and platforms that build this, not as a feature, not as a product category, but as foundational architecture for how human judgment moves through the economy, will do for the labor market what exchanges did for financial markets and credit scores did for lending: create the transparency and liquidity that makes the whole system function better for everyone in it.

That is one of the largest unseized infrastructure problems in the modern economy. And the moment for building it is not coming. It’s here.

It was because I stopped looking at the bill like software spend.

I had spent about $537 in metered AI usage over four days. On its face, that number looks high if your mental model is “employee asks chatbot questions.” It looks very different if your mental model is “professional delegates work to a team of intelligence.”

That is the shift most leaders are still underestimating.

AI is not just a tool. A spreadsheet is a tool. A dashboard is a tool. A search bar is a tool.

AI reads, plans, drafts, critiques, delegates, revises, remembers, and synthesizes. That does not make it a person. It does not make it safe by default. It does not remove human accountability. But it does mean the management frame has to change.

We already spend on intelligence every day. We call it salary, billable time, utilization, leverage, margin, and delivery cost. Human intelligence has a price. AI intelligence has a price.

The strategic question is not whether intelligence should cost money.

It always does.

The question is which work should be done by human agency, which work should be done by AI agency, and how we know the difference.

That is where AI ROI actually begins.

The Wrong Question Is “What Did the Tokens Cost?”

Most AI reporting still starts with consumption.

How many users? How many prompts? How many conversations? How much model cost?

Those are useful operating signals. They are not ROI.

Nobody measures a consulting team by counting keystrokes. Nobody looks at a senior manager’s salary and says, “We need fewer sentences.” We ask what work got done, what quality came back, what cycle time changed, what risk was reduced, what client value was created, and whether the economics make sense.

AI should be measured the same way.

A token is not value. A token is a unit cost of applied intelligence. The ROI question is whether that intelligence produced useful work at a better cost, speed, quality, or scale than the alternative.

This is why “reduce AI spend” can be the wrong instinct. Sometimes the right answer is to spend less. Sometimes the right answer is to spend more because more of the work has moved into the system.

The real management question is more precise:

What did this intelligence cost, and what did it produce?

What Changed in My Four-Day Experiment

In my four-day work burst, the spend did not come from asking a chatbot a lot of clever questions.

The largest category was structured analytical production: decision maps, long-form synthesis, report writing, evidence shaping. That represented about 37 percent of the metered spend.

The second-largest category was context infrastructure: project memory, status logs, decision capture, and the machinery that lets agents work across projects. That was about 26 percent.

That second bucket matters more than it looks.

In ordinary chatbot work, the human carries most of the context. The model gets a prompt, maybe a file, maybe a summary, and answers inside that narrow window. The human remembers the client history, the last meeting, the open questions, the political constraints, the source quality, and the shape of the deliverable.

Agentic knowledge work flips more of that into the system.

The system has to read across documents, emails, transcripts, research, slides, spreadsheets, prior decisions, and partially stated preferences. It has to infer what matters. It has to leave some things out. It has to build enough context to act.

That costs more because the work is bigger.

Anthropic’s Claude Code cost guidance is useful context here. Their enterprise deployments average about $13 per developer per active day, and 90 percent of users stay below $30. But they also note that agent teams can use roughly seven times the tokens of standard sessions because each teammate maintains its own context.

That is the mechanism.

Agents cost more when they are not just answering. They are carrying context, coordinating steps, and doing work in parallel.

The Breakout Example: Research Synthesis

Here is the example that made the economics click for me.

One workflow in the four-day window was research synthesis and source review. In plain English, the work looked like this:

1. Build a research pack.

2. Run a deeper research pass.

3. Review market pressure and current-source evidence.

4. Create an evidence base.

5. Produce a source brief that could support a real argument.

The metered AI cost for that workflow was about $63.

Now compare that with human intelligence.

If a senior consultant did the same work manually, a conservative estimate is 24 to 36 hours. At a $150,000 salary benchmark, that is about $72 per hour. So the human-cost equivalent is roughly $1,700 to $2,600 before overhead, and much higher at a billable rate.

That does not mean the AI “saved” $2,500 by itself. That would be too easy.

The output still has to be reviewed. Sources have to be checked. The synthesis has to be challenged. The recommendation has to be shaped by someone with judgment.

But even if a human spends three or four hours reviewing and improving the output, the economics are still different. The human is no longer spending most of the time gathering, sorting, and drafting. The human is spending time deciding.

That is the point.

AI ROI is not only labor replacement. In professional work, the better frame is intelligence redeployment.

Take the work that humans do because someone has to grind through the context. Move the first pass into an AI workflow. Then move the human to the parts where human judgment is actually scarce: source skepticism, client implication, risk, taste, trust, and the final recommendation.

That is where the return shows up.

Not in cheaper words.

In better use of judgment.

This Is Also Why Humans Stay in the Loop

The best evidence for AI in consulting also explains why blind automation is dangerous.

In the BCG field experiment with researchers from Harvard, Wharton, and MIT, 758 consultants using AI completed 12.2 percent more tasks, completed them 25.1 percent faster, and produced work rated more than 40 percent higher in quality when the task was inside the model’s capability frontier.

That is the upside.

But on a task outside the frontier, consultants using AI were 19 percentage points less likely to produce correct solutions.

Same technology. Opposite result.

That finding should shape the operating model. AI agency is powerful when the workflow is legible, evidence can be checked, and quality gates are built into the process. Human agency matters most where the work is ambiguous, the stakes are high, the context is political, or the answer has to be owned.

The question is not “human or AI?”

The question is “what mix of human and AI agency does this workflow deserve?”

The Infrastructure ROI Question

This is where the AI Chief of Staff idea becomes less futuristic and more practical.

Imagine it takes two people three weeks to build the operating system around a leader: memory, preferences, work intake, project context, source rules, reusable workflows, specialist agents, review loops, and reporting.

That build has a cost.

At the same $150,000 salary benchmark, two people for three weeks is roughly 240 hours, or about $17,000 of human build cost before overhead. Add AI usage, governance time, and implementation friction, and maybe the real internal investment is higher.

Fine. Then measure it like any other investment.

If one repeatable research workflow costs about $63 in AI usage and redirects $1,700 to $2,600 of human effort before review, the payback question becomes concrete.

Use a rough formula:

Payback runs = infrastructure build cost / net value per workflow run

For the research synthesis example, assume the AI workflow costs $63 and a human still spends three to four hours reviewing the work. At the same salary benchmark, that review time is about $216 to $288. So the combined AI-plus-review cost is roughly $280 to $350.

Compare that with $1,700 to $2,600 of manual human effort.

Very roughly, that workflow redirects $1,350 to $2,300 of human capacity each time it runs.

Against a $17,000 build cost, the infrastructure pays back in something like 8 to 13 successful runs before overhead. If one consultant runs that workflow every week, the payback is measured in a quarter. If a team uses it across repeated client work, the payback arrives faster.

The exact number will vary. The method matters more than the decimal.

Now the conversation is no longer “AI is expensive.”

It is: “Which workflows pay back the infrastructure fastest?”

That is a much better question.

It tells you where to build first. It tells you where to keep work human. It tells you where a cheaper model is good enough. It tells you where a stronger model is worth it. And it gives leaders a way to compare AI infrastructure investment against workforce deployment.

A Consultant With an AI Chief of Staff

Picture a senior consultant on a Monday morning.

Today, that consultant may spend the first hour reconstructing the week: client emails, meeting notes, half-finished deliverables, status updates, research tabs, and the lingering sense that something important is hiding in yesterday’s transcript.

In the AI Chief of Staff model, that first hour changes.

The Chief of Staff has already assembled the brief. It knows the consultant’s active projects, what changed overnight, which client question is urgent, which deliverable is at risk, and which follow-ups are still open. It has updated the memory layer. It has flagged what needs human attention.

Then it starts routing work.

A meeting synthesis agent turns yesterday’s transcript into decisions, risks, and next steps. A research agent builds the source pack for a market question. A writing agent drafts the client-ready narrative. A validation agent checks claims, assumptions, and source quality. A workflow agent updates the project log and prepares the next status note.

The consultant is not out of the loop.

The consultant is finally in the right loop.

Instead of being the person who remembers everything, finds everything, drafts everything, and then reviews everything at 9 p.m., the consultant becomes the orchestrator and reviewer of a small intelligence system.

That role is not smaller. It is more leveraged.

It also changes staffing economics. Some work still belongs with people: relationship management, trust-building, hard judgment, escalation, original strategy, and anything where the cost of being wrong is high. Some work should move aggressively to AI: first-pass research, synthesis, meeting capture, draft creation, comparison tables, source briefs, status updates, and repeatable analysis.

The firms that learn this fastest will not simply cut cost.

They will increase output per professional.

The Market Is Moving, But the Model Is Still Open

This shift is already visible.

Deloitte expects 25 percent of companies already using generative AI to launch agentic pilots in 2025, rising to 50 percent by 2027. EY says its EY.ai Agentic Platform will initially integrate 150 AI agents supporting 80,000 professionals. KPMG says its Workbench platform has a network of 50 AI assistants, agents, and chatbots, with nearly a thousand in development.

But I would be careful with what that proves.

It does not prove that everyone has solved the future of professional work. Most of what is public still looks like defined workflow automation, task agents, or delivery platforms. Valuable, yes. Directionally important, yes. But not yet the full personal operating model: a persistent Chief of Staff that understands one professional’s work surface and orchestrates across the whole day.

That layer is still open.

OpenAI’s B2B Signals report points in the same direction. Frontier firms use 3.5 times as much intelligence per worker as typical firms, but message volume explains only 36 percent of the gap. The difference is not just more chat. It is richer context, deeper work, and more delegation.

McKinsey’s 2025 AI survey found that 23 percent of organizations are scaling an agentic AI system somewhere, but no individual function has more than 10 percent reporting scaled agent use.

The window is open because almost nobody has turned this into a scaled operating model yet.

The New Management Discipline

The next management discipline is not prompt engineering.

It is intelligence allocation.

For each workflow, leaders need to know:

1. What output are we trying to produce?

2. What human intelligence does it require today?

3. What AI intelligence could produce the first pass?

4. What review or judgment must stay human?

5. What did the combined system cost?

6. What outcome changed?

Ask that at the workflow level, not the chatbot level.

The answer will not be the same everywhere. Some tasks deserve high intelligence spend. Some deserve cheap automation. Some should stay mostly human. Some should become hybrid workflows with tight review gates.

That is the work now.

Not “How do we get everyone to use AI?”

“How do we redesign work around the right mix of intelligence?”

The firms that win will not be the ones with the most prompts or the lowest token costs. They will be the ones that learn to spend intelligence well.

Human intelligence where it matters most.

AI intelligence where it scales best.

And a measurement system that can tell the difference.

I’ve been thinking about interface design for twenty years. Last month, I drew a diagram on a whiteboard that made me question whether that was still a job.

Not in an “AI is taking everything” way. In a more specific, structural way — the kind that only makes sense once you’ve seen it, and then you can’t unsee it.

I was sketching the system diagram for a proof of concept — mapping the data flow from conversation to an agentic pipeline, to a database, to an MCP server, to…

And somewhere in the middle of drawing that last box — the one representing where the user actually works — I stopped.

I didn’t need an interface.

I mean, I was planning to design one. I would have been thinking through user flows, building screens, debating layouts, nudging pixels. But sitting there looking at the diagram, I realized that once the data reaches the work surface, all of that is optional. The user can just ask for what they want. However they want to see it.

And I don’t think most people building enterprise apps have registered it yet.

The Assumption That’s Breaking

For thirty years, interface design has been built on a quiet assumption: we know what the user needs to see, and our job is to build it for them.

It’s a good assumption. It gave us the GUI, it gave us mobile-first design, it gave us the entire discipline of UX research — the idea that if you understand your user deeply enough, you can design the perfect path through information.

Nielsen Norman Group recently named the current moment one of only three major interface paradigm shifts in computing history. The first was batch processing. The second was command-based interaction — sixty years spanning terminals, CLIs, and eventually GUIs. The third is what’s happening now: AI systems that reverse the locus of control entirely. Users don’t tell the computer how to do something anymore. They tell it what they want.

That framing gives this moment historical weight without overstating it. The GUI didn’t arrive all at once. It took years to become dominant. But at some point, the direction became obvious.

We’re at that inflection point now.

The Signals Are Already in Your Building

Here’s what I’ve been watching.

The executives in my organization are spinning up HTML files to present ideas. Not hiring a designer. Not submitting a creative request and waiting a month. They’re just doing it. They describe what they want, it exists. Before, that was a month-long project with a loaned designer from another team. Now it’s an afternoon. That’s not a productivity improvement — it’s a different model of how information gets made visual.

OpenAI ran an “Intelligence at Work” event this week. One number in the announcements deserves more attention than it got: knowledge workers now represent 20% of Codex’s weekly active users — and they’re growing three times faster than developers. The fastest-growing tasks: data analysis up 110% week-over-week, research up 37%, knowledge artifacts — reports, presentations, memos — up 36%.

Anthropic took Claude Cowork out of preview in April and pushed it to all paid plans. They built it after watching non-technical teams bypass the chat interface and work directly in Claude Code to get things done. Microsoft’s Copilot Workspace went generally available at Build this week. SpaceX secured a $60 billion option to acquire Cursor, with both companies working together to build “the world’s best coding and knowledge work AI.”

All of this in the last ninety days.

This isn’t a trend to watch. It’s a product category forming in real time — one that makes the traditional designed interface optional.

What a Work Surface Actually Is

Let me be specific, because there’s a lot of noise in this space.

Claude Cowork and OpenAI Codex are work surfaces. Not IDEs, not chat interfaces — surfaces. They take the raw power of coding agents and bring it into environments where knowledge workers operate. They connect to your data via MCP (the protocol that lets AI agents plug into external systems), they run multi-step processes, and they generate whatever display the user asks for.

In the innovation app we’re building, an idea goes through ten stages of processing. When the user wants to see how their idea stacks up against competitors, they ask for a comparison table. The work surface generates it. They want to see potential market value? They describe how they want it displayed, and it appears. The table’s missing a competitor? They say so — it updates. The chart doesn’t show the regional breakdown they care about? They redirect — it regenerates.

The interface isn’t fixed. It resolves from the data and the user’s intent.

I’ve done this with output from other teams too — taken HTML files they gave me and regenerated the interface to the same data in a completely different way, because what I needed to understand wasn’t what they needed to present. The display was malleable. The information was stable.

This is what designers have been trying to build for thirty years. Adaptive interfaces. Personalized to the individual. Responsive to context. The holy grail of UX — always theoretically possible, always practically impossible, because building it explicitly for every user state at every permutation was impossibly expensive.

AI didn’t move the goalposts closer. It changed the game entirely.

What Design Becomes

There’s a practical reason this matters beyond the obvious efficiency gains. A colleague in engineering ops put it plainly: the cognitive cost of transitioning between projects isn’t being accounted for in resource planning — nobody measures it, but it compounds. The work surface changes that equation. The information moves to you rather than requiring you to navigate to it.

Which brings the obvious question: does this mean designers are finished?

No. But the craft changes in ways the design community hasn’t started thinking about yet.

Nielsen Norman Group put it directly in their work on generative UI: “We’ll need to shift from designing interfaces to designing outcomes. Humans will need to provide guidance and constraints for generative UI.”

The pixel-nudging goes away. What replaces it is higher-order thinking that this craft was always building toward — if the mental model shifts.

Instead of “what should this screen look like,” the question becomes “what should information look like when a user is in a comparative analysis intent state?” You define what that mode delivers: what it must show, what it should show, what it never shows. You define the logic, not the layout. You think in intent states and visual models, not components and flows.

The design system doesn’t disappear — it loosens into something more like a skills library. Think `SKILL.md` files that define how a comparative analysis intent should display, what a financial summary module must include, which patterns are anti-patterns for research output. Code snippets for distinct display modules. JSON that encodes the brand constraints and hierarchy rules. The agent reads those files, interprets them, and renders accordingly. You’re not arranging pixels — you’re authoring the rules the surface follows.

For engineering and product leaders, the implication runs upstream of the design org: the roadmaps you’re setting and the products you’re shipping were designed for a world where the interface is fixed. That world has a shorter runway than most roadmaps assume.

Where to Skate

I showed a colleague the diagram from that meeting. He works in these tools every day — Cowork, Codex, all of it. His mind was blown. Not because the technology was new to him, but because he’d never assembled the signals into this picture.

That’s where most of the design, product, and engineering community is right now. The tools are in the building. The behavior is happening. The product category launched simultaneously from three major companies in the same week.

The question isn’t whether this shift is coming. It’s whether you’re thinking about where the puck is going.

Stop designing for the interface that exists. Start defining intent states. Map what information looks like in comparative analysis mode, financial analysis mode, research synthesis mode. Build the skills library — the `SKILL.md` files, the display rules, the encoded brand constraints — that an agent can interpret and apply. Think in intention and anti-pattern, not layout and flow.

The pixels aren’t going anywhere. But the job isn’t placing them anymore. The job is defining what the information should mean — and letting the surface figure out how to show it.

Is it still a job? Yes. A better one. But only if you’re willing to stop doing the old one.

It wasn’t. That $400 day was a swarm of agents setting up a database, wiring three integrations, and standing up a small application — work that, three years ago, would’ve been a six-week sprint with two engineers. I ran it from my laptop on a Tuesday afternoon.

They put me on a list.

Here’s the thing, though. The list has a name now, and it didn’t come from my company.

In April, The Information leaked the existence of an internal Meta dashboard called Claudeonomics — a leaderboard ranking all 85,000-plus Meta employees by token consumption. The top performer had burned 281 billion tokens in 30 days. Mark Zuckerberg didn’t crack the top 250. There were badges: Token Legend, Session Immortal, Cache Wizard. Meta killed the dashboard once the press got hold of it, but the concept stuck.

By those standards, I’m a rounding error. What flagged me wasn’t extraordinary by industry standards — it was extraordinary by the standards of a normal company.

The term went mainstream fast. Y Combinator built a whole podcast episode around founders using AI to do “the work of 400 engineers.” A developer named Sigrid Jin told Axios, with a straight face, that if you want to understand the future of AI you should “spend on your AI pricing as much as you pay for your own house rent.”

I’d been nodding at all of this from a distance without quite registering that I was already doing it.

So I pulled my own numbers. And then I did the math on what would happen if everyone in my company started doing the same thing. That math is the actual reason I’m writing this.

My Bill, And Yours

From May 1st through May 19th, thirteen working days, I’ve spent $1,677.05 on Claude. That’s about $129 a working day. Running a normal 40-to-45-hour week. Annualized, that’s somewhere around $31,000 a year.

For reference, Steve Yegge — a longtime engineer at Sourcegraph, ex-Google and ex-Amazon, who’s been writing publicly about agentic coding for two years — recommends companies budget exactly that: $80 to $100 per developer per day on AI tokens. Tomasz Tunguz, the VC who tracks SaaS economics for a living, calls it “the fourth component of engineering compensation.” I’m not an outlier. I’m just early.

Now scale me up. Take a 5,000-person company. Tech-heavy, mixed knowledge workers, fully loaded payroll of around $650 million — that’s roughly $130,000 a person, which is reasonable for a mid-cap tech-ish org. Now imagine everyone is using AI at my pace.

That’s about $168 million a year in tokens. Twenty-six percent of payroll. In CFO terms: roughly 1,290 jobs worth of cost, just in compute.

You read that and you think, holy shit, that’s a lot. And it is.

Until you look at the output side.

I’ve been tracking my own performance for a while, and the productivity gain doesn’t collapse into one clean number — it stratifies by what you’re measuring. Task-level execution, the actual work I ship, is running about 20x my pre-agent baseline. Breadth of capability, the range of things I can credibly take on in a week, is around 16x. Project oversight at my level — I run AD-level work — is closer to 6x, because management leverage has different ceilings than execution. Sustained overall output and velocity, the headline number, lands around 15x. And on a single best-fit project, where everything aligned, I hit 120x.

Pick a multiplier and run the math.

If 5,000 people produce only 1.67x — that’s the measured jump in merged pull requests per engineer per day at Anthropic after they rolled out Claude Code, and it’s a measured number, not a self-report — that’s 3,350 extra full-time-equivalents of work for 1,290 jobs of token spend. About 2.6 to one. Even the floor pays back more than double.

At a defensible middle of 3x: 10,000 extra FTEs for 1,290 jobs. About eight to one.

At my sustained 15x: 70,000 extra FTEs for 1,290 jobs. Fifty-four to one.

At the 120x peak — which won’t sustain across a whole org, but tells you where the ceiling is — 461 to one.

There is no plausible number on this curve where a CFO can rationally cut the AI budget.

The one wrinkle worth naming, because it changes the calculus for some companies: if you build and serve your own AI products, you don’t just buy the tokens. You buy the compute that produces them. For a hyperscaler, the choice isn’t “tokens versus headcount” — it’s “data centers versus headcount.” That’s part of why a company can post record revenue and still trim staff: the capex bill has to come from somewhere. If you’re the rest of us — buying inference, not building it — your math is simpler. But know which game your company is in before you read your own numbers.

Why The Numbers Got Big In The Last Six Months

If any of this had been possible in 2025, we’d have seen it then. We didn’t. So what changed?

A few things happened at once, somewhere between December 2025 and February 2026.

The models hit the inflection point on the hockey-stick capability curve everyone had been predicting. Sonnet 4 and Opus 4 weren’t just incrementally better than their predecessors — they crossed a reliability threshold where you could leave them running for a few hours on something multi-step and ambiguous, and trust the output enough to ship it. Claude 3.5 was a smart assistant. Claude 4 was a competent colleague.

The harnesses matured. Claude Code, a curiosity in February 2025, hit $1 billion in annualized revenue by November and $2.5 billion by February — the fastest product ramp Anthropic has had. Cursor’s agent mode became genuinely usable. Sub-agents, MCP, hooks, parallel runs — the boring infrastructure of agentic work all got good in the same six-month window.

The patterns spread. Boris Cherny, who runs Claude Code at Anthropic, started publicly describing a workflow where he runs five parallel Claude instances on five git checkouts, with another five to ten browser sessions going at once. “A few thousand agents overnight,” he says. He ships 20 to 30 pull requests a day. Three years ago that was a parody of how engineering works. Now it’s the template, and dev influencers have spent six months building tools — Straude, ccusage, agent farms — to help everyone else copy the pattern.

The result, for individual operators, is a step change that doesn’t show up in the older studies — because the older studies were run before it happened. Hold onto that timing. It matters later, when we get to the research that says none of this is real.

What The Headlines Get Wrong

Open any newspaper this year and you’ll read that AI is taking jobs. Then you’ll read that AI isn’t taking jobs. Then you’ll read a Nobel economist saying it’s actually a 1 percent bump to GDP over a decade. The takes cancel each other out, and most of them are missing what’s actually in the data.

Here’s what the data says.

Challenger Gray, the firm that tracks announced layoffs, reported 1.2 million job cuts in 2025. AI was cited as the cause for only 54,836 of them. The biggest single driver wasn’t AI — it was DOGE, with nearly 294,000. Tech led the private sector at 154,000. So if you’re looking for the white-collar bloodbath in the layoff numbers, it isn’t there. AI is a small minority of the cuts.

But layoffs are the wrong place to look. Look at hiring.

2025 was the lowest year of hiring since 2010. The BLS revised down employment for the year ending March 2025 by 991,000 jobs — nearly a million people who never got hired. Indeed Hiring Lab reports software developer postings sitting 36 percent below their February 2020 baseline. Stanford’s Canaries in the Coal Mine paper, published in August, found that 22-to-25-year-olds in highly AI-exposed jobs have seen a 13 percent relative drop in employment since late 2022. Entry-level software engineering and entry-level customer service are each down about 20 percent. Older workers in the same roles grew.

The single most important sentence in macroeconomic policy right now was published in the November 2025 Federal Reserve Beige Book. Almost no one outside Fed watchers has read it. Here it is:

More Districts reported contacts limiting headcounts using hiring freezes, replacement-only hiring, and attrition than through layoffs… A few firms noted that AI replaced entry-level positions or made existing workers productive enough to curb new hiring.

The Philadelphia Fed put it even more cleanly: AI is letting employers “skip a recruiting class of entry-level workers.”

That’s the story. Not a bloodbath. Not nothing. A long, quiet, structural freeze.

The “AI is cutting jobs” narrative is wrong because the cuts are mostly other things. The “AI isn’t cutting jobs” narrative is wrong because it’s only counting cuts. What’s actually happening is a generation of unfilled requisitions, evaporated entry-level cohorts, and headcount lines that quietly never get added back when somebody leaves.

A layoff is loud. An empty job slot is silent. We’re watching the silent thing.

What CEOs Are Saying Out Loud

The funny part is, the people running these companies aren’t being subtle about it. They’re saying it on stage, in memos, on earnings calls. We just keep missing it because we’re scanning for “we cut 8,000 people” instead of “we will not hire.”

Tobi Lütke, CEO of Shopify, sent a memo to his whole company in April 2025 that read, almost word for word: before you ask for headcount, prove that AI cannot do the job. He said he’d seen employees “getting 100X the work done.” That’s not a phrase you smuggle into a memo lightly. He went on the record.

Marc Benioff, on a February 2025 earnings call: Salesforce wasn’t hiring any new software engineers that year. He cited a 30-percent-plus engineering productivity lift from Agentforce, and reportedly committed around $300 million in Anthropic spend for 2026 to back the bet.

Mike Krieger — Anthropic’s CPO, the guy who co-founded Instagram — told Lenny Rachitsky that 90 to 95 percent of Claude Code is now written by Claude Code. Satya Nadella said in May 2025 that AI was writing up to 30 percent of the code on some Microsoft projects; Microsoft cut about 6,000 people that same month. And this spring, Meta cut 8,000 and quietly canceled another 6,000 open reqs— even as its most recent quarter brought in a record $56 billion. They called it “flattening middle management.”

IBM is the case worth studying. Its AskHR agent automated 94 percent of routine HR tasks and displaced hundreds of HR jobs. But the CEO also noted — and this is the part the bloodbath story leaves out — that the savings got redeployed into engineering, sales, and marketing. Total headcount grew.

The pattern is loud once you start listening for it: not “we fired everyone,” but “we stopped hiring, and the people we kept are doing more.”

The Honest Caveat

This would be an easier piece to write if I left out the parts that complicate it. I’m leaving them in, because if you walk away thinking “AI is a 15x miracle, fire your team,” you’ll deserve what happens next.

One thing to hold first: the step change I just described happened between December 2025 and February 2026. Almost every published study skeptical of AI’s productivity gains was collected before that window.

The single most-cited counter-evidence is the METR study from July 2025. A randomized controlled trial: 16 experienced open-source developers, working on their own codebases, with Cursor Pro and the best Claude models available. The result — they were 19 percent slower with AI than without it. And those same developers believed they’d been 20 percent faster. A 39-point gap between perception and reality. Critics of AI productivity claims wave this paper around constantly. It’s the silver bullet.

Here’s the wrinkle. The study was published in July 2025, but the data was collected earlier — spring 2025 at the latest. The developers were using whatever was best at the time: Claude 3.5 and 3.7 Sonnet, an early version of Cursor’s agent mode, no mature sub-agent patterns, no MCP ecosystem. They were running the pre-unlock tools. A year in this technology is the difference between a learner driver and an F1 pit crew. I’d bet a 2026 re-run — same developers, same repos, but Claude Code and Sonnet 4 with sub-agents and parallel runs — would invert the result. METR has said publicly it’s revisiting the study. We don’t have the new numbers yet, but the burden of proof has quietly shifted.

The other two big skeptical findings tell a quality-and-organization story, and both drew on pre-agentic data. GitClear’s analysis of 211 million lines of production code through 2024 found copy-pasted lines rising from 8.3 to 12.3 percent and duplicate code blocks up eight-fold — consistent with the worry that AI favors volume over the structural rework that keeps a codebase healthy. A Faros AI analysis of over 10,000 developers found AI users completing 21 percent more tasks and merging 98 percent more PRs, while organizational delivery metrics stayed flat. Both are real signals worth watching, neither is a verdict, and both measured a world that has already moved on. The 2025 DORA report‘s line is still worth pinning to the wall: “AI doesn’t fix a team; it amplifies what’s already there.”

And then there’s Klarna. In 2024 the CEO publicly claimed they’d replaced 700 customer-service agents with AI. By 2025 he was quietly reversing course — rehiring humans, admitting customer satisfaction had collapsed on anything emotional or multi-step. The most famous AI-replacement story in business has already boomeranged. The system they deployed was a 2024-era model, not good enough for the job they handed it. Whether a 2026 redo would land differently is the obvious question, and Klarna isn’t running it.

So what do I do with my 15x?

I keep it. And I tell you the truth: I’m a self-selected, agent-native operator whose work happens to map well to what these tools are now good at. I’m in the tail of the distribution. I’m also a self-interested narrator — I want this technology to be transformative because my career is built on it being transformative. Discount accordingly.

But the math demolishes anyway. Take Anthropic’s measured number — 67 percent more merged PRs per engineer per day, not a self-report — and the cost-output ratio in the 5,000-person scenario is still better than two and a half to one. You don’t need to believe my 15x, or Lütke’s 100X, or DORA’s 21 percent. At every published figure, in every study, the AI spend pays back many times over.

What’s Actually Happening

Here’s the model I keep coming back to.

A company has two main inputs: capital and people. AI doesn’t change capital much for most companies — tokens are cheap relative to almost everything else. AI massively changes what a person can produce. The output of any given headcount line has just been multiplied. Not 15x for everyone, but materially, measurably more.

That means every company faces a question it didn’t face two years ago: with each person producing more, do you ship more, or do you keep producing the same amount with fewer people?

The answer depends on what economists call demand elasticity. If your market can absorb more — if you can ship more software, sell more services, expand into adjacent products — the rational move is to keep everyone and ride the productivity wave. Stripe. Shopify. Anthropic. These companies are growing their effective output faster than their headcount because there’s more market out there to take.

If your market is saturated, the rational move is to cut. You don’t need 700 customer service agents if 200 of them can handle the volume. So you fire 500. Klarna’s original move was correct in theory; it just turned out the AI wasn’t yet good enough to do the job. The shape of the argument was right.

Most companies are in between. They have some expandable market, but not enough to absorb their current headcount at the new productivity ceiling. So they freeze. They stop posting new reqs. They let attrition do the work. They take credit when revenue per employee goes up. They don’t say “we replaced you with AI” — they say “we’re being disciplined about hiring this year.”

That’s a layoff with a soft edge. It’s still a layoff, in aggregate, across the economy. It just doesn’t show up in Challenger Gray’s monthly report, because there’s nothing to announce.

The Beige Book describes exactly this. The BLS quietly revised down a million jobs that never materialized. The Stanford paper found the freeze landing hardest on the people just entering the workforce. The macro picture is a million unfilled chairs, not a million empty desks.

What To Do About It

If you’re running a company with somewhere to grow: this is the biggest leverage opportunity of any tooling decision in twenty years. Find out who your tokenmaxxers are, point IT at the budget instead of at them, and measure people on outcomes per quarter, not hours in a chair. The economics aren’t subtle, and they hold at even the most skeptical multiplier. If your team isn’t spending Yegge’s $80-to-$100 a day per developer, your competitors will — and they’ll ship more.

If you’re running a mature business with limited expansion: the cost-savings math is real, but Klarna is the warning. Take the productivity gains. Don’t take them all the way to “fire the humans.” Customer satisfaction collapses, quality drops, you spend the next year crawling back. Use the savings to fund the next product line instead.

If you’re an employee: the work itself is changing faster than the headlines suggest. The skill appreciating most isn’t writing code or copy or shipping designs — those are getting commoditized. The skill appreciating is directing the agents that do that work. Being the person who runs the system, not the person whose work becomes the system. That’s where the multiplier lives.

And if you’re an entry-level worker in a knowledge job: the recruiting class you didn’t get hired into is not bad luck. It’s structural, it’s still in early innings, and the path forward is to learn to operate the tools that are otherwise eating your job. The cohort that figures this out will compound. The cohort that doesn’t will be the lasting damage from this transition.

One Last Thing

The IT email wasn’t an alarm. It was a leading indicator. Somewhere in your company, the first person is already tokenmaxxing. Maybe they got flagged this week, like I did. Maybe they’re about to file an expense report that’s going to surprise someone in finance.

Pay attention to that person. They’re showing you the future of the next budget cycle, the next reorg, the next earnings call where you have to explain why revenue per employee is up and headcount is flat.

The hiring freeze is the AI story. The layoffs are the headline. Don’t confuse them.

She noticed three months ago. The AI review queue she worked through every morning had dropped — 60% fewer cases needing her attention. Her manager hadn’t said anything. No context, no conversation, no acknowledgment that anything had changed.

She didn’t ask. She updated her resume instead.

Here is what nobody told her: the queue dropped because she was doing her job correctly. Every edge case she documented, every piece of reasoning she captured, every pattern she identified — the system had learned from it. She was mapping the AI’s boundary, and the boundary had moved. That was the job. That was success.

The problem wasn’t silence. The problem was that her job description still measured her by the old metric. Volume of cases reviewed. Throughput maintained. By every executor standard she had ever been evaluated against, a shrinking workload was a warning sign.

No one had given her a different standard to measure against. So she used the one she knew.

This is Stage 1 redeployment failing — not because the technology went wrong, not because the communication plan fell short, but because the job was never redesigned. The metric was never changed. And the person doing the work had no way to distinguish success from a slow exit.

The Executor Trap

Every professional’s instinct is calibrated to throughput. Handle more. Process more. Be indispensable by volume. The measure of a good week is a full queue efficiently cleared.

That instinct was built over careers where it was the right instinct. It is the wrong instinct for the co-pilot stage — and most organizations never change the framework to say so.

The co-pilot role, designed correctly (as Part 2 established), is not an executor role. Its purpose is to map where the AI’s reliable territory ends — and to do that mapping so precisely that the map transfers to the system itself. The co-pilot isn’t clearing cases. They’re defining the boundary. Every accurately documented edge case is boundary data. Every precisely reasoned intervention is training signal. Every week where fewer cases needed a human is a week where the work landed.

The executor’s success metric is volume maintained. The co-pilot’s success metric is volume declining.

Those are not the same job. But most organizations deploy the co-pilot role and leave the executor metric in place. The job description doesn’t change. The performance review criteria don’t change. The implicit signal in the culture — do more, handle more — doesn’t change. And so the person in the role optimizes for the metric they have, which is the wrong one.

The results are predictable. Exception rates plateau. Progression stalls. Stage 1 continues indefinitely because the people running it are calibrated — unconsciously, reasonably — to keep it running. Gartner’s prediction that 40%+ of agentic AI projects will be abandoned by end of 2027 is regularly attributed to technology failure or unclear ROI. One underexamined explanation is the metric: organizations that never redesigned the co-pilot role plateau at Stage 1 and conclude that the technology didn’t deliver. The technology delivered. The job design didn’t.

The Redesigned Role

What the job looks like when redesigned is not abstract. Here are three industries where the transition is in progress and the change is concrete.

Legal. The executor version: review 150–200 contracts per week, flag issues, pass to senior associates. The success metric is throughput — contracts cleared, turnaround time maintained.

The redesigned version: the job is to build the AI’s exception taxonomy. Not to review contracts en masse, but to identify the specific clause patterns the AI consistently misclassifies — the indemnity formulations it reads as standard when they aren’t, the liability caps it approves that fall outside acceptable range for this client type — and to document the reasoning behind each correct answer precisely enough that the pattern can be encoded. The output is a taxonomy, not a review count. The ABA’s 2025 task force declared that “AI has moved from experiment to infrastructure for the legal profession”; the firms that progress past Stage 1 are the ones where the co-pilot is building that taxonomy, not just clearing the queue.

The success metric: intervention rate declining week over week. The proportion of AI-flagged items requiring human correction should be shrinking. When it does, the boundary is moving.

Accounting. The executor version: test 400–500 journal entries per week, run reconciliations, perform vouching procedures. Throughput is the measure. This is the routine work that trained junior CPAs through sheer volume — and that the AI has largely absorbed.

The redesigned version: the job is to map the AI’s exception patterns. The AI will flag anomalies — not all of them are errors. Some are legitimate business events that look unusual to a pattern-recognition system but are valid in context. The redesigned co-pilot identifies which anomaly categories are genuine errors, which are legitimate events the AI hasn’t learned to recognize, and builds the decision logic that distinguishes them. The output is a decision framework. The AICPA’s Profession Ready Initiative, launched in February 2026, is the profession’s acknowledgment that the job is changing and the job description hasn’t caught up.

The success metric: the proportion of AI-flagged anomalies requiring human review should decline as the framework matures.

Software engineering. The executor version: review pull requests, fix bugs, implement features. Commit count, PR throughput, story points closed.

The redesigned version: agent supervisor. The job is to identify the architectural decisions AI coding agents consistently get wrong — where agent-generated code is syntactically correct but architecturally misaligned, where it introduces technical debt that won’t surface until a later integration — and to document those failure modes precisely enough to inform system-level constraints. Anthropic’s 2026 Agentic Coding Trends Report found agents now complete an average of 20 autonomous actions before requiring human input. What the engineer supervising those agents is actually doing is defining the boundary of those 20 actions — which decisions the agent can own, which it must escalate, and what the failure modes look like when it gets that wrong.

The success metric: not commits. Architectural decision quality. Reduction in integration failures across agent-generated work.

In all three cases, the new success metric runs in the same direction: intervention rate declining, the boundary sharpening. The work that looks like doing less is exactly the work the role was designed to do.

The Paradox, Named

Here is what nobody says out loud in most co-pilot deployments: the better you do this job, the less the job needs doing.

Not the less valuable you are. The less this specific function requires a person. Those are different things, and the failure to distinguish them is where the woman in the opening went wrong — and where most organizations fail to help people go right.

Every professional instinct built over a career points toward indispensability through volume. Be needed for more, not less. The co-pilot role inverts that at the task level. Your most successful days are the ones where the AI needed you less than it did last week, because that means the boundary you’ve been mapping is holding.

This creates a specific and largely unacknowledged psychological challenge. Most co-pilots don’t consciously resist the new metric. But they unconsciously optimize for the old one. They handle cases the AI could probably manage, because handling things feels like working. They document edge cases less precisely than they could, because the intervention is the visible work and the documentation is invisible. They don’t push the AI as hard as they should, because pushing harder means fewer interventions, which reads — through the old metric — as being needed less.

The effect is subtle. No individual decision looks like resistance. The aggregate is a stalled exception rate and a Stage 1 that never progresses. Not because the technology failed, and not because the people are disengaged. Because the metric was wrong and the job was never redesigned to say so.

This is the fixable part. But only if it’s named.

Why the Paradox Is Wrong

The success paradox only holds if you’re looking at the wrong window.

William Stanley Jevons observed in 1865 that when steam engines became more efficient, Britain didn’t use less coal — it used more. The efficiency gain reduced cost per unit. Lower cost per unit expanded demand. Expanded demand consumed more total resource than inefficient engines ever had. Jevons’ Paradox has held across energy economics for 160 years. It applies here.

When AI makes knowledge work output cheaper per unit — cheaper contract reviews, cheaper audit samples, cheaper code checks — volume doesn’t stay flat. Cost drops trigger demand expansion. The PE firm that could afford one due diligence report now commissions three. The mid-market company that couldn’t justify a legal review for routine contracts now orders one on everything significant. An accounting firm that once managed 80 client engagements can handle 400. Output multiplies because the unit cost fell.

Every additional unit of AI output is an additional demand on human judgment. Someone has to evaluate it, direct it, decide what to do with the result, and own the outcome. As output volume grows, that demand grows with it. The co-pilot who maps the AI’s boundary and exits Stage 1 doesn’t graduate into a world with less need for judgment. They graduate into a world where the demand for judgment — the work that can’t be automated — is expanding faster than it was before automation started.

BCG’s April 2026 analysis of approximately 165 million U.S. jobs found that 50–55% will be reshaped over the next two to three years, with only 10–15% fully displaced over a longer horizon. 68% of companies expect to maintain workforce size. And accounting offers the most precise version of Jevons already in motion: Client Advisory Services — the judgment-intensive work — grew at a median 17% rate compared to 9.1% for overall firm revenue. Net CAS fees per professional reached $156,250, up 29% since 2022. Practices project 99% growth over the next three years. The compliance work is automating. The advisory work is accelerating. Those are not two separate trends. They are the same dynamic: unit cost falling, total demand rising.

The co-pilot’s success is not self-elimination. It is the prerequisite for stepping into the work Jevons guarantees is expanding.

What the Judgment Jobs Actually Look Like

The redistribution thesis is often left at the level of “advisory roles” and “higher-value work” — accurate but not useful to someone trying to understand what their Monday morning looks like in eighteen months. Here is what the endpoint jobs actually look like in the industries where Stage 1 is progressing.

The former document reviewer (legal).

The Stage 3 job is not “senior reviewer.” It is strategic associate: managing 5–8 active matters simultaneously, synthesizing AI analysis into the 2–3% of issues that require attorney judgment, briefing the partner on exactly which decision points need their attention. The client relationship that needs a person on the phone. The novel argument on an unsettled legal question where there’s no good training data. The SDNY ruling in February 2026 confirmed the accountability floor: an attorney’s judgment, signature, and professional standing cannot be delegated to a system. That floor is the Stage 3 job description. Success is no longer measured in contracts reviewed. It’s measured in matters resolved, client relationships held, issues caught before they became problems.

The former transaction tester (accounting).

The Stage 3 job is Client Advisory: combining the AI’s financial pattern detection with the business context the AI doesn’t have, identifying the strategic implication of what the numbers show, advising on decisions the analysis informs. The PCAOB’s AS 1000 standards hold the accountability floor in audit; everything the AI handles below it frees the practitioner for the work above it. The CPA who can tell a client what a pattern in their financials means for their next capital allocation decision is doing work that requires the AI’s output, professional judgment to interpret it, and enough client context to make it actionable. Success is measured in advisory revenue and decisions informed — not tests run.

The former code reviewer (engineering).

The Stage 3 job is engineering lead for a system that includes AI agents as team members. The work is architectural: defining the constraints the agents operate within, handling the integration failures that cross agent boundaries, making the system-level calls that no agent can make because they require a view of the whole. Where Stage 1 was catching the AI’s edge cases in individual pull requests, Stage 3 is designing the systems that determine what edge cases the agents will encounter. Anthropic’s labor market study shows no systematic increase in unemployment for AI-exposed workers, but a clear hiring slowdown for 22–25-year-olds in highly exposed occupations. The front of the pipeline is thinning; the work at the top is not contracting.

In each case, the path from executor to judgment role runs through the co-pilot stage. Not around it, not despite it — through it. The boundary-mapping work of Stage 1 creates the taxonomy that defines Stage 2. The judgment-call inventory of Stage 2 defines the permanent floor that Stage 3 works within. The stages are not just organizational milestones. They are how the individual develops the judgment the Stage 3 job requires.

The Three Conversations

The manager’s job in the co-pilot transition is not primarily to manage anxiety. It is to redesign the job explicitly, install the new metric from day one, and name the endpoint in advance.

Anxiety is downstream of metric confusion. The woman who updated her resume wasn’t irrational — she applied a reasonable inference to the information available to her. The fix is not better reassurance. It is a job description that makes a declining queue recognizable as success.

At the start.

Not: “your role is to review AI outputs for quality.” That is the executor description. The redesign conversation sounds like this: “Your job is to define what right looks like in situations the AI hasn’t encountered yet. The more precisely you document the reasoning behind each judgment call — not just the outcome, but why that’s the right answer in this context — the faster the system learns. We’re going to track your intervention rate together. Over time, that number should decline. When it does, it means what you’re doing is working.”

That framing gives the person a way to measure their own success that is consistent with the stage ending. The redeployment is not a surprise if the co-pilot understood the stage from the beginning. Prosci’s research is consistent: 58% of employees want to hear about personal impacts from their direct supervisor — not from a town hall, not from corporate communications. The conversation has to happen at the manager level to land.

During.

Show the curve. The intervention rate data belongs on the co-pilot’s screen, not just the management dashboard. A declining line is motivating in a way that abstract progress updates are not — it shows that the boundary they were asked to move is moving. Gallup’s April 2026 research found that employees who strongly agree their manager actively champions AI use are more than twice as likely to use it frequently. Showing the curve is the simplest possible version of that.

At the end.

The advancement conversation requires specificity. Not: “your role is evolving.” Not: “we’re moving to the next phase.” Those statements communicate that something is ending without saying what comes next.

Specific looks like: “The class of cases you were handling is now handled. The intervention rate has held below threshold for six weeks. That means Stage 1 is done. Here’s the taxonomy you built — this is what Stage 2 works from. And here’s where we’re pointing you next, and why.”

The co-pilot built something. Tell them what it was.

The Arc

This series gave you the map (Part 1), the mechanics (Part 2), and now the personal transition.

The hardest thing to hold, for anyone in a co-pilot role, is this: I am doing the job right when I am needed less for this specific task. That inverts every professional instinct built over a career. The instinct isn’t wrong — it’s calibrated to the wrong window.

Zoom out, and the picture changes. WEF’s Future of Jobs Report 2025 projects 78 million net new jobs globally by 2030, with 170 million created against 92 million displaced. BCG’s 2026 analysis found 68% of companies expect to maintain workforce size. And Jevons tells you why: AI output volume is expanding, and every unit of AI output is an additional demand on the human judgment that evaluates, directs, and owns it.

The work that succeeds by shrinking is building toward work that doesn’t. The judgment jobs are on the other side of Stage 1 being completed, not abandoned.

The junior analyst from the opening would have stayed if someone had given her a new metric to measure herself against. Most people would. The technology automates the work. The job redesign automates the fear.

Both matter. Only one of them requires a manager who understood what Stage 1 was actually for.

Part 1 ended with four questions: Where does your work sit on the 2x2? Which stage are you in? How do you know you’re progressing? What does it look like when Stage 1 ends and Stage 2 begins?

They’re four questions with the same root cause — a single decision most leaders make without knowing they’re making it. It’s usually buried inside a job description. It’s usually made in the first week of deployment. And it cascades into everything else.

The decision is this: when you put a human in the loop alongside your AI, what is that person’s actual job?

The Wrong Design

The default co-pilot deployment looks like this. AI generates outputs. Humans check them. Errors get flagged and corrected. Leadership watches the accuracy metrics. The governance committee declares the deployment responsible.

It sounds reasonable. It is the design that keeps organizations frozen in Stage 1 indefinitely.

The tell is in the language. If your co-pilots’ job could be described as reviewing AI outputs for mistakes, the design is wrong.

Not wrong because human review is bad. Wrong because review, as typically structured, produces passive oversight — a process that catches errors in the moment but doesn’t systematically improve the system over time. The AI generates output. The human says yes or no. The correction happens in that case. The AI doesn’t get meaningfully smarter. The co-pilot role continues indefinitely, because it was never designed to end.

In July 2024, Gartner predicted that 30% of generative AI projects would be abandoned after proof of concept by end of 2025 — citing escalating costs and unclear business value. That deadline has now passed. Their more recent forecast is starker: 40%+ of agentic AI projects canceled by end of 2027, because the AI cannot autonomously achieve complex goals or follow nuanced instructions over time. The reasons are consistent across both predictions. These are not technology failures. They are design failures. Organizations supervise AI without systematically improving it, performance plateaus, and the conclusion drawn is that the technology was the problem. The organizations that avoid this outcome are the ones that diagnose the design early enough to change it.

Recon Analytics found that 44.2% of lapsed Microsoft Copilot users cite distrust of answers as the primary reason for stopping. The passive design produces exactly this. Errors slip through. No one was systematically feeding the pattern of failures back into an improvement loop. The human did their job. The AI didn’t get better. The human stopped trusting it.

The cost isn’t just stalled adoption. The American Institute of CPAs launched its Profession Ready Initiative in February 2026 precisely because the accounting profession has been deploying passive co-pilots for years. The routine transaction testing, reconciliations, and data review that used to train junior accountants — the work through which the profession transferred judgment from senior to junior practitioners — is now done by AI. Nobody designed the mechanism for capturing what was embedded in that work and replacing it. Now the profession is scrambling to design a replacement: simulation-based learning, AI role-play, scenario training, cross-level mentoring. The active learning principle applied to professional formation, a year after the passive co-pilot design revealed its cost. The skills pipeline hollowed out while the governance dashboards showed green.

The Right Design

The reframe is small on the surface. The cascading effects are not.

The co-pilot’s job is not to catch what the AI gets wrong. It is to define what right looks like in situations the AI hasn’t encountered yet.

That sentence changes what a good day looks like. In the catch-errors design, a good day is one where the co-pilot found the error before it went out the door. In the guide design, a good day is one where fewer interventions were needed than last week — because the ones logged two weeks ago were captured, structured, and learned from. The difference is not philosophical. It is operational.

Consider what changes. How the role is described to the person doing it: “review AI outputs” versus “define what right looks like when the AI hits a situation it hasn’t seen.” The first is a quality gate. The second is active instruction. These are not the same job. In the first, the person asks: is this correct? In the second: is this correct, and if not, what is the correct answer and why — in terms the system can learn from?

How interventions are logged changes too. In passive oversight, a flag means the output was wrong and a human corrected it. In a guide-framed deployment, an intervention is a structured record: what situation triggered the override, what the AI produced, what the correct answer was, and what category of gap this represents. The log is the product — not because it creates an audit trail, but because it creates training signal.

There’s a reason this works at a technical level. In machine learning, active learning means having humans label only the cases the model finds most informative — the specific situations it doesn’t know how to handle — rather than passively reviewing everything. The efficiency gains across empirical studies are significant; the principle translates directly. A guide-framed co-pilot is concentrating human expertise at the boundary of what the system doesn’t yet know. A catch-errors co-pilot is reviewing outputs for mistakes but not systematically feeding the edge-case structure back into any improvement loop. Same human, same time, fundamentally different output.

The guide design also has an end condition the catch-errors design lacks. A human will always be needed to catch errors, because there will always be errors. But when the intervention rate for learnable cases approaches a defined threshold, the guide’s job is done. The role was never permanent. It was designed to end.

Building the Measurement System

Design the role correctly, and the measurement follows.

Enterprise human-in-the-loop deployments already track a metric called the exception rate — the number of human interventions required per N automated decisions, expressed as a percentage. Standard practice in mature HITL operations: accounts payable automation, insurance claims processing, document review at scale. Emerging vendor benchmarks from automation contexts suggest best-in-class operations run below 10% exception rates, with rates above 60% typically signaling the system is being deployed on work it isn’t ready for. Formal industry-wide thresholds for professional services AI don’t yet exist — which is itself instructive.

The concept is not new. The application to professional services co-pilots is.

In law, accounting, consulting, and insurance brokerage, few organizations are tracking exception rates as maturity signals. GitHub measures suggestion acceptance rates across millions of developers — an industry average of roughly 30%. Enterprises run adoption dashboards. What they’re not tracking is whether the rate of human intervention is declining over time, and whether that decline signals Stage 1 progressing or frozen.

That is the governance dashboard versus training dashboard distinction. A governance dashboard tracks AI usage — how often the tool deploys, how many outputs are reviewed. A training dashboard tracks AI improvement — how often human intervention is required, whether that rate is declining, and what the trend line says about readiness for the next stage. Most organizations have built the first. Almost none have built the second.

The training dashboard needs four things:

• total interventions per week (raw count)

• intervention rate expressed as interventions per 100 AI actions, which normalizes for volume so a rising count from more AI usage doesn’t mask a declining rate

• breakdown by intervention type (more on this below — the real signal lives here)

• The trend line, which is where the signal actually lives. A declining slope means Stage 1 is progressing. Flat means frozen. Rising means something is wrong.

McKinsey’s analysis of agentic AI deployment is direct on this point: organizations that build risk management, auditing, and oversight capabilities early are the ones that scale AI fastest — not despite the governance infrastructure, but because of it. The governance structure is not the obstacle to speed. It is what makes speed possible.

Classifying Interventions

The breakdown by intervention type is the most operationally powerful output a well-designed co-pilot deployment produces. It is also the output almost no organization is building.

Every human intervention in a guide-framed co-pilot falls into one of two categories.

Learnable edge cases are situations the AI hasn’t encountered yet but, once shown the correct answer, will handle correctly next time. The AI got it wrong not because the decision requires human accountability, but because this specific configuration of facts hasn’t appeared in its training. The correct answer can be encoded. Show the AI what to do here and its intervention rate on this class of problem declines toward zero.

Permanent judgment calls are different in kind, not just degree. These are not cases the AI got wrong because it lacks information. They are cases where the answer requires contextual discretion, ethical reasoning, or the kind of ownership that cannot be attached to a system that has no professional license, no duty of loyalty, and no consequences for being wrong. Show the AI the correct answer here a hundred times — it may produce the right output more often, but it cannot own the decision. The accountability lives with the human because it has to. These don’t decline. They represent the permanent floor.

The practical test: if you showed the AI the right answer in this case and every case like it, would it handle them correctly next time — not just produce the right output, but carry the accountability for the outcome? If yes, it’s learnable. If the correct answer depends on non-encodable factors — client relationship history, regulatory interpretation, ethical judgment in a genuinely novel context — it’s a judgment call.

No existing published framework maps cleanly to this distinction. Microsoft’s taxonomy of failure modes in agentic AI addresses technical failures. Multi-agent system research classifies coordination breakdowns. Enterprise HITL systems route by AI confidence thresholds. The learnable/judgment taxonomy is the next layer: of the cases that trigger human review, which are confidence gaps and which are accountability gaps? It emerges naturally from the guide framing. You won’t build it from the catch-errors design because you’re not asking the right question.

What the ratio tells you: if most interventions are learnable, Stage 1 has significant progression ahead. If most are already judgment calls, Stage 1 may be near its natural ceiling — and Stage 2 needs to be designed now.

The judgment calls that remain when learnable cases approach zero are the Stage 2 job description. Not a vague statement about “higher-value work” — an actual taxonomy of the decisions that required a human, why they required one, and what seniority of practitioner would be needed to make them. The insurance industry is building toward this: McKinsey projects that up to 95% of P&C policies could eventually pass through underwriting without human involvement. The HITL layer handles what remains — coverage disputes, fault determination in novel circumstances, and decisions that require accountability. That remaining layer is not what’s left when the AI fails. It is the permanent judgment floor that years of active co-pilot deployment will map. In healthcare, Notable Health reports that MUSC Health applied the same logic — guide-framed authorization, AI handling routine cases, humans for exceptions — and reallocated over 1,300 hours per week to higher-value patient care. The hours didn’t disappear. They moved up.

The Advancement Decision

When the intervention rate for learnable edge cases approaches the threshold defined before deployment, Stage 1 is done. The co-pilot role — as designed — has completed its job.

Most organizations won’t know this has happened. Running governance dashboards, the signal won’t surface. Without intervention classification, there’s no way to separate the declining learnable cases from the stable judgment calls — the aggregate rate looks flat even as the internal composition has shifted entirely. And without an advancement threshold set in advance, there’s no trigger. Stage 1 continues because no one built the condition that would end it.

Waymo published its formal advancement framework — “Safe to Deploy” — in June 2025. Twelve acceptance criteria spanning hazard analysis, cybersecurity, scenario-based verification, on-road testing, and risk management. A layered governance structure: methodology leads evaluate domain performance, a Safety Framework Steering Committee aggregates findings, a Waymo Safety Board makes the final deployment decision. The formal standard: “absence of unreasonable risk” — not zero interventions, not perfect performance. Evidence-based determination that risk has been reduced to a level acceptable for deployment.

Two things worth lifting for professional services. First, the criteria were defined before deployment — not after reaching a performance level that prompted an informal conversation about whether it was time to advance. The threshold was set in advance. When performance reaches it, the governance structure convenes. The decision is not informal and the trigger is not subjective. Second, “absence of unreasonable risk” is explicitly not 100% certainty. Waymo does not require zero failures. It requires evidence-based determination that the risk level is acceptable. The advancement threshold for a professional services co-pilot will not be zero learnable interventions. Set the acceptable level before Stage 1 begins, not after.

In practice: define what intervention rate, sustained over what period, signals that Stage 1 has done its job. Define what evidence is required and what governance structure makes the call. Then, when Stage 1 ends, document the boundary before the co-pilots redeploy — the edge cases they mapped and the judgment calls they classified are the knowledge the organization built. The judgment calls that remain at that point are the Stage 2 brief. Stage 2 begins with clearer scope than Stage 1 ever had.

McKinsey’s framework names the target end state “human above the loop”: AI operating autonomously, humans monitoring at a supervisory level and able to intervene when anomalies are detected. That is what a successful Stage 1 is building toward. The human who started case-by-case — in the loop — migrates to system-level supervision. Not the elimination of human oversight. The elevation of it.

The Decision

Part 1 ended with the question: do you know which stage you’re in?

This piece adds a second: how did you design the role that’s supposed to advance you through it?

If the answer is “to catch errors,” the design is producing passive oversight. The AI is not getting systematically better. The intervention rate is not tracked. The edge cases are not classified. There is no advancement threshold. Stage 1 will continue until someone makes a deliberate choice to change it — or until the project is canceled and the conclusion is that the technology failed.

If the answer is “to guide the AI through situations it hasn’t encountered yet,” the design is producing active learning. The intervention rate is declining. The edge cases are classified. The learnable cases are being absorbed. The judgment calls are accumulating into a taxonomy that will become the Stage 2 brief.

The decisions cascade from one. Most leaders don’t know they’re making it.

The co-pilot stage was always supposed to end. Build it like it will.

Estimated read time: 15 min | Word count: ~3,700*

I’d booked an autonomous vehicle. I expected an empty car — the full robot-car experience. Instead there was a woman in the driver’s seat. Hands in her lap, not on the wheel. Eyes forward. Not driving, exactly. Just… there.

The feeling wasn’t relief. It was uncanny. Something about a human sitting in a driver’s seat while not driving felt more unsettling than no human at all. My brain had preloaded an expectation and reality contradicted it in the wrong direction.

What I didn’t know then: she was there because the system needed her — just not to drive. She was there for the moments the car couldn’t handle on its own. The unmapped construction zone that appeared overnight. The pedestrian in a high-vis vest redirecting traffic outside the app’s model of the world. The decision point where the trained boundary ends and a human has to define what happens next. Her job was the edge case. Every one she handled moved the boundary. Every one she logged taught the system something it hadn’t known before.

Waymo ran the same sequence in Dallas: human safety drivers from May 2025, fully driverless by November 2025. Same routes, same city — and eventually, no one in the seat. The transition took six to seven months. The job of sitting in that seat — catching edge cases at that level — was done. The next set of problems needed attention somewhere else in the operation, by her or by someone else. Either way, the human need had elevated to the next level of abstraction.

Most organizations see the co-pilot as a destination, not a stage in their transition. They deploy a co-pilot — some combination of AI tools and human oversight — and declare success. The usage dashboards go green. Leadership calls it AI adoption. And the progression stalls, indefinitely, at Stage 1.

This is a major misread of the moment. And it’s happening everywhere.

The Misread: Why Organizations Get Stuck

There’s a story most companies tell themselves about the co-pilot stage, and it goes like this: we’ve deployed AI, we’ve kept humans in the loop, we’ve balanced innovation with governance. The dashboard is green. The audit committee is satisfied. The transformation roadmap says “Year 2: Scale.” Good enough.

This is not a strategic choice. It’s a failure mode with a polished name — co-pilot.

The co-pilot stage is not a destination. It is a training mechanism with an expiry date built into its design. Sequoia Capital’s Julien Bek put it plainly in March 2026: “A copilot sells the tool. An autopilot sells the work.” A co-pilot augments a human. An autopilot replaces the workflow. Those are not the same thing, and the direction of travel between them is fixed. The only real question is whether you’re navigating that progression deliberately or riding along, passive, until it catches up with you.

Three traps keep organizations frozen at the co-pilot stage, and they tend to operate simultaneously.

The governance trap. Leaders see co-pilot deployment as risk management: humans are in the loop, errors get caught, liability is contained. That framing is accurate as far as it goes, but it mistakes a phase for a final state. Governance is the mechanism that makes the co-pilot stage work — active human oversight catching errors, feeding corrections back into the system, building reliability. When governance becomes the goal rather than the method, the progression stops.

The job security trap. Professionals in AI-exposed roles quietly read the co-pilot stage as proof they are still necessary. The AI needs me here. This reading is understandable — and wrong. The woman in the driver’s seat was not proof that Waymo needed a driver. She was the final stage of testing before they took her out. Every edge case she handled made her presence less necessary, not more.

The metrics trap. Usage dashboards measure adoption, not progression. When AI tools are in use, leadership calls the transformation successful. Nobody asks the harder question: is this stage advancing, or is it frozen? If the same co-pilot roles exist in exactly the same form they did twelve months ago, the organization has not been adopting AI. It has been institutionalizing Stage 1.

Underneath all three traps is a structural logic most people miss: the co-pilot stages are not ideological. They are load-bearing.

When AI systems chain multiple steps — which is what any real professional workflow requires — errors compound. To illustrate: an AI agent achieving 85% accuracy per step has roughly a 20% chance of completing a ten-step workflow without a failure. At 95% accuracy across twenty steps, that figure rises only to 36%. The numbers are illustrative, but the principle is not. You cannot deploy a four-in-five failure rate on professional work that matters. The co-pilot stage exists to close that gap. Humans catching edge cases, identifying failure modes, escalating exceptions — that process is not supervision for its own sake. It is the mechanism that makes autonomous deployment eventually possible. The stages are not conservative. They are mathematically necessary.

Bain Capital Ventures mapped this into a six-level autonomy framework — from AI assistant at level one to fully autonomous agent at level six. Their advice is direct: “Own where you are and explain the reasons why this represents the ideal tradeoff between human and AI work for your use case.” Most organizations cannot do that. They haven’t done the audit. They don’t know which level they’re on.

That is the first problem. Here’s the map.

The Map: Where Your Work Sits

To know your timeline, you need to know where your work sits. This is not abstract — it determines which industries and roles move fastest, and why.

Two axes.

The vertical axis runs from intelligence at the bottom to judgment at the top. Intelligence is the work AI already handles well: pattern recognition, synthesis, information retrieval, processing speed, anomaly detection in structured data. Judgment is different — it requires accountability, contextual discretion, ethical reasoning, navigating genuinely novel situations, and owning the consequences of decisions. Intelligence can be optimized. Judgment has to live somewhere a human can be held responsible for it.

The horizontal axis runs from outsourced on the left to insourced on the right. Outsourced work is already delivered externally — there’s an existing budget line, outcomes-based purchasing, and the organization has already accepted that this work can be done from outside. Insourced work is core to operations: headcount-based, embedded in the org structure, part of how the company sees itself.

The bottom-left quadrant — outsourced and intelligence-heavy — is where AI autopilots arrive first. This is not an accident.

Sequoia’s Bek made the economic logic explicit: for every dollar companies spend on software, they spend six on services. That $6 is the autopilot’s total addressable market — the work that can be replaced by an AI that delivers the same outcome faster and cheaper. Companies that already outsource have accepted that the work can be done externally. There’s a budget line. They’re buying outcomes, not presence. That is the path of least resistance for an AI that delivers those outcomes.

The global business process outsourcing market sits at $347.95 billion in 2025, with finance and accounting the largest single segment at roughly $74 billion. Legal, tax, and HR outsourcing are all above 50% adoption rates. The bottom-left quadrant is enormous — and it is moving.

Management Consulting: The Disruption Is Coming From Outside

Here is the thing about management consulting that the public narrative misses: the threat to the industry is not coming from inside the firms. It’s coming from the clients.

What consulting firms sell is judgment — the brand promise is that senior partners have seen this before, know how to frame the problem, and can guide organizations through the hard part. But what consultants actually do on any given Tuesday is mostly intelligence work: market sizing, benchmarking, financial modeling, research synthesis, slide fabrication. The ratio of intelligence to judgment is high, because junior staff handle the intelligence work while partners handle the judgment.

This makes consulting simultaneously among the most defensible and most vulnerable professional services.

McKinsey reports that its internal AI platform, Lilli, is used by 72% of the firm’s 45,000 employees and saves consultants up to 30% of their research time. The firm has deployed 12,000 internal AI agents for document analysis and task coordination. BCG separately reported that consultants using AI complete tasks 25.1% faster with 40% higher quality. The intelligence layer is being automated rapidly — inside the firms.

But private equity firms managing over $2 trillion in assets are, according to DiligenceSquared, already replacing $500,000 McKinsey due diligence reports with $50,000 AI-generated equivalents. The company raised a $5 million seed round in March 2026. A 90% cost reduction on the intelligence layer, reported in live production at some of the largest investment firms in the world. The clients aren’t waiting for the consulting firms to automate themselves. They’re going around them.

What survives is the judgment: problem framing, stakeholder alignment, guiding organizations through the political and human dimensions of change. Those are not things McKinsey’s Lilli automates — and they’re not things DiligenceSquared is touching either. They are the work that makes recommendations stick.

Legal: The Accountability Floor Is Written Into Case Law

Legal services are moving faster than most professionals in the field are comfortable admitting. 79% of law firms have now integrated AI for document review, legal research, and contract analysis. Junior associates who once handled the volume work are now supervising AI outputs and being redeployed to analytical tasks — case strategy, interpretation, advocacy that requires judgment rather than retrieval. The role did not disappear. It changed shape and moved up.

The judgment floor in legal is not just conventional — it has been set by case law. In February 2026, SDNY Judge Jed S. Rakoff ruled that documents generated using AI tools are not protected by attorney-client privilege or work-product protection: “An AI tool is not a lawyer. It has no law license, owes no duty of loyalty, cannot form an attorney-client relationship, and is not bound by confidentiality obligations.” The accountability boundary in legal is codified. The judgment layer lives with the attorney because the law requires it to.

The realistic endpoint is an AI that handles everything except the attorney’s signature, judgment call, and ethical obligations. That is a very large portion of what lawyers currently do — and the profession is moving faster than the public narrative suggests.

Accounting and Audit: The Intelligence Layer Is Already Automated

In audit, the progression is perhaps the most clearly defined of any professional service — because regulators have drawn the line for you.

AI now handles planning analysis, internal control testing, journal entry anomaly detection, and real-time reporting on control deviations. The intelligence stack is largely automated. What AI cannot do — and what regulation explicitly requires a human to do — is exercise professional judgment, sign, and be accountable.

The PCAOB’s new AS 1000 auditing standards (effective for fiscal years beginning December 15, 2025) were specifically designed to address technology-assisted data analysis. The PCAOB’s position is that AI is “a catalyst for improved audit quality” — but the human auditor remains accountable and must document evidence of human oversight. The auditor’s signature is not a ceremonial formality. It is a structural requirement.

The skills pipeline implication is real and is already being managed. The AICPA launched an initiative called “Profession Ready” in direct response to a recognized gap: much of the time-consuming work that used to train junior accountants — transaction testing, routine reconciliations, vouching — is now being handled by AI. An AICPA/CIMA survey found that 56% of respondents named generative AI as the area with the largest skills shortage. Nearly a third of accountants under 25 are considering leaving the profession within five years, at a time when the average CPA age is 46. Accounting is the one profession actively managing the skills pipeline problem that other industries have not yet admitted they have.

Insurance: The Same Industry at Both Ends of the Grid

Insurance is the clearest single-industry demonstration of how the 2x2 works — because the same industry has radically different automation levels at different ends of the value chain.

Simple personal auto claims are already at Stage 3–4 autonomy. AI-enabled carriers have cut claim resolution time by 75% and reduced cost per claim by 30–40%. Industry sources report straight-through processing rates of 70–90% for basic personal auto claims at leading insurers — no human touches the claim. The intelligence layer of claims processing is not being automated. It has been automated.

At the other end: complex commercial underwriting and brokerage judgment remain firmly human. Sequoia specifically cited insurance brokerage as a $140–200 billion AI autopilot opportunity. But the brokerage judgment layer — complex risk advice, managing corporate client relationships, negotiating coverage on specialty risks — is the more defensible end of the industry. Sequoia-backed Pace raised $10 million in Series A funding to automate the $70 billion insurance BPO market with agentic AI — targeting the intelligence layer of brokerage back-office operations, freeing brokers to focus on the relationship and advice work that sits in the top-right quadrant.

The gap between simple auto claims and specialty commercial brokerage is not two different industries. It is the same 2x2 showing both ends of the grid within one sector.

Software Development: When the Grid Moves Across the X-Axis

Software development is the outlier in this mapping. Unlike most of the industries described here, it is largely insourced — core to most organizations, headcount-based, embedded in how the company builds and ships. It sits on the right side of the horizontal axis, not the left.

But it is extremely intelligence-heavy. Anthropic’s March 2026 labor market study found computer programming at 75% observed AI coverage — the highest of any single occupation measured. Anthropic’s 2026 Agentic Coding Trends Report found that AI agents now complete an average of 20 autonomous actions before requiring human input — double what was possible six months prior. Developers at Rakuten tested Claude Code on a 12.5-million-line codebase; the system completed a complex implementation task in 7 hours with 99.9% numerical accuracy. TELUS teams created over 13,000 custom AI solutions while shipping engineering code 30% faster.

The engineering role is moving from writing code to supervising agents, reviewing outputs, handling integration failures, and managing architecture decisions. This is what it looks like when the grid moves across the x-axis rather than up the y-axis — the work is insourced, but the intelligence layer is being automated as fast as or faster than in the outsourced industries.

The Top-Right Quadrant: Where the Boundary Settles

At the top right of the grid — insourced, judgment-heavy — sits the work that AI can inform but not own.

C-suite strategy and capital allocation. Client relationships in professional services. Crisis response and novel situations without precedent. Ethical decisions with accountability attached. These are not intelligence tasks with a judgment label stuck on them. They are genuinely different work — work where the accountability, the relationship, and the judgment call cannot be separated from each other and handed to a system that cannot own consequences.

AI can synthesize inputs for these decisions, and it is increasingly good at it. BCG’s research found that consultants using AI complete some tasks traditionally classified as analytical judgment 40% more effectively. The boundary between intelligence and judgment is moving upward — AI is handling more sophisticated analysis, leaving humans with higher-order calls.

But the evidence for AI making final strategic decisions, owning client relationships, or managing the political and human dimensions of consequential organizational choices does not exist in professional services. What exists is the boundary moving. Where it eventually settles is where accountability has to live — and accountability requires a human who can be held responsible.

That is the top-right quadrant. Not everything that sits there is permanent. But nothing moves until the accountability question has an answer.

The Stages: How You Actually Get There

The 2x2 tells you where work sits. The stages tell you how it moves. Four stages, roughly sequential, observable in real industries right now.

Stage 1: Junior Co-Pilots

Remember the woman in the driver’s seat — there for the edge cases, not the driving. That is Stage 1, and it applies everywhere AI is being deployed into professional work.

The traffic cone is the clearest illustration of what an edge case looks like in practice. In San Francisco, activists discovered that placing a traffic cone on a Waymo hood would disable the vehicle entirely — it would stop and wait, indefinitely, unable to decide. The story spread because it was funny. It also spread because it was true. The system hadn’t been trained on deliberate sabotage by a traffic cone. It hit the boundary of what it knew, and it froze. That is exactly the moment a human co-pilot exists to handle.

Every error caught, every exception escalated, every situation where the human has to step in and define what the right answer looks like — that is the training data. The junior co-pilot is not doing quality control. They are defining the boundary of what the AI doesn’t yet know, and each time they do, that boundary moves.

As the boundary moves, edge cases shrink. When edge cases shrink, the manager notices spare capacity. Not through a corporate transformation program — through ordinary management doing what managers do. They see slack and they fill it. The worker gets redirected. The AI covers the base. No announcement required.

Where do they go? Not out — up. The work that opens up as AI absorbs the base layer tends to be the work that requires a human to be present: coordinating across teams, managing the handoff between AI systems that don’t talk to each other well, handling the client or colleague who needs a person to speak to. This is not consolation work. It is the connective tissue of any organization — the human coordination layer that AI cannot replicate — and it has always been underserved because junior staff were busy doing the intelligence work instead.

The Anthropic study shows the leading indicator already in the data: no systematic increase in unemployment for AI-exposed workers since late 2022, but hiring of 22-to-25-year-olds in highly exposed occupations has already slowed. Stage 1 redeployment is happening at scale, mostly invisible, in the hiring data rather than the unemployment data. The front of the pipeline is thinning, not the middle falling out.

Skills hollowing is the real risk here, and it is worth naming directly: junior professionals who passively approve AI outputs are not developing the judgment they will need for mid-level and senior work. The AICPA launched “Profession Ready” because the accounting profession had already recognized this failure mode was active, not hypothetical. The co-pilot stage only trains the AI if the human is actively catching errors, questioning outputs, and understanding why the system failed — not just approving them. Good managers design Stage 1 as active learning. Bad ones create a generation of button-pushers, and that problem compounds forward into the pipeline.

Stage 2: Mid-Level Escalation

With junior-level edge cases handled, harder ones surface. The role doesn’t disappear — it changes shape and moves up.

In software development, the shift is already past Stage 1 for most serious engineering teams. Autonomous coding agents now complete 20 actions before needing human input. Engineers at the cutting edge are not writing code — they are supervising agents, reviewing pull requests for logic errors and integration failures, and handling the architecture decisions and system-level calls the agent cannot make. The Gartner-reported 1,445% surge in multi-agent system inquiries from Q1 2024 to Q2 2025 reflects what this looks like organizationally: single agents giving way to orchestrated teams of specialized agents, supervised by a human engineering lead acting as orchestrator.

In legal, the same pattern plays out at the mid-level. Junior lawyers redeployed from document review to analytical work move the judgment floor upward — but the analytical work they move into is still intelligence-adjacent. As AI handles that tier too, the escalation continues upward. Senior associates and partners absorb the cases requiring courtroom advocacy, complex advice, client relationships, and the kind of legal judgment that the SDNY ruling confirms must attach to a licensed attorney.

The mid-level practitioner’s transition follows the same logic as the junior’s — but the destination is different. Where juniors absorb the human coordination layer, mid-levels move into deeper client work and advisory roles: the cases that require someone who can synthesize AI outputs, apply professional judgment, and translate the technical into something a client can act on. This is not a smaller job. In most professional services, it is a more valuable one. The practitioners who manage this stage well use the window deliberately — expanding into the work above them while AI consolidates the work below.

Stage 3: The Senior Judgment Layer

The genuinely complex cases are left. Senior practitioners co-pilot the hardest work — situations requiring accountability, context, and reasoning the AI cannot reproduce at the required level of reliability.

In audit, this is the most precisely defined version of the stage. AI handles the full intelligence stack: planning, anomaly detection, control testing, variance analysis, real-time dashboards. The CPA holds the accountability line. The PCAOB’s AS 1000 standards require documented evidence of human oversight — not because regulators are being conservative, but because the audit opinion is a professional representation to the capital markets. The auditor’s signature is where accountability lives. It is not optional.

In consulting, the DiligenceSquared disruption is eliminating the intelligence layer from outside the firm — a $500,000 due diligence report replaced with a $50,000 AI-generated equivalent, in live production at PE firms. What survives is the judgment layer: partner relationships, strategic framing on questions with no precedent, the ability to guide an organization through the human dimensions of a major transformation. Those are Stage 3 activities. They are not being automated.

At Stage 3, the senior practitioner’s role is not shrinking — it is clarifying. As AI absorbs more of the intelligence layer beneath them, what remains is the work that most directly requires seniority: the client relationship that needs repairing, the regulatory negotiation that depends on professional standing, the strategic call that draws on pattern recognition across decades of experience, the novel situation that has no precedent in any training dataset. In many cases, Stage 3 is the point at which senior professionals finally get to do the work they entered the profession for — unencumbered by the intelligence tasks that used to fill their days and their junior staff’s.

One point on regulated industries is worth making directly. When people say “full autonomy is impossible in audit or legal,” they are usually making one of two different claims. The first is that the judgment boundary is set by regulation and will not move — largely true and likely to remain so. The second is that the stages model therefore doesn’t apply — which is not true. “Full autonomy” in audit means AI handles everything except the professional sign-off. That is an enormous portion of what auditors currently do. The stages model works in regulated industries. The judgment boundary settles where accountability requires it to live, not where the technology runs out.

Stage 4: Full Autonomy

The AI handles the full stack for a defined domain. Humans have migrated — up the judgment ladder, into new roles, into work the AI still cannot do.

Supervision does not end. It migrates up a level.

When Waymo removed human drivers from its Dallas fleet, it did not eliminate oversight. Riders can contact support via in-car screens or the app. A control room exists. Exception handling moved from the vehicle to the system level. The supervision decoupled from the vehicle and moved to the infrastructure that manages the fleet. The same pattern applies in professional services. Stage 4 does not mean no humans. It means humans managing exceptions, handling novel situations, owning the judgment calls that arise when something falls outside the trained boundary — one level up from where they were.

This is not a theoretical endpoint. Waymo went from supervised to fully autonomous in Dallas in six to seven months. The timeline was not industry consensus or regulatory permission. It was evidence: edge cases shrinking to the point where the human in the seat was no longer load-bearing. The progression is real, and it plays out faster than intuition suggests.

What This Means Now

The model above is not a prediction. It is a description of what is happening in industries you can see right now. The question is what it means for where you sit.

First question: do you know which quadrant your work sits in? Most leaders don’t. They can describe what their organization does, and they can describe what AI they’ve deployed. The intersection — where the work actually lives on the intelligence-to-judgment and outsourced-to-insourced grid — is usually unanswered. Without that answer, you cannot know which stage you’re in, which work is most exposed, or where the progression is already moving without you.

Second question: are your junior co-pilots actively catching errors or passively approving outputs? The difference determines whether the co-pilot stage is doing what it is supposed to do. If your junior staff are rubber-stamping AI outputs, two things are true simultaneously: the AI is not getting better at the boundary cases, and your junior staff are not developing the judgment they will need. Gartner predicts over 40% of agentic AI projects will be abandoned by end of 2027. The most common cause is not technology failure. It is organizations treating the co-pilot as the destination and never building toward the next stage.

Third question: is supervision moving up your org chart, or stalling? If the same people are doing the same co-pilot work they were doing twelve months ago — the same review loops, the same error-catching, the same escalation patterns — the stage is frozen. The indicator of real progression is that the nature of what gets escalated is changing, getting harder, moving up toward judgment. If it isn’t, something is wrong.

Fourth question: what happens to the institutional knowledge embedded in the co-pilot stage when the stage ends? This is the question almost nobody is asking. The junior practitioners catching edge cases are not just doing quality control. They are accumulating an understanding of where the AI fails, what the boundary looks like, and what kind of judgment is required to handle what the AI can’t. When those practitioners redeploy, that knowledge has to transfer somewhere — into the AI’s training, into the documentation, into the mid-level practitioners who inherit the harder cases. Organizations that treat Stage 1 as a checkbox and move on without managing that transfer will find the next stage harder than it should be.

Knowing where you are is the first step. It is also, for most organizations, the step that hasn’t been taken.

The harder questions — and the more valuable ones — come next. How do you structure the co-pilot role so it is genuinely training the AI, rather than just providing the appearance of oversight? How do you know when you’re ready to move from one stage to the next — and what does that transition actually look like in practice? How do you manage the workforce shift at each handoff without losing the institutional knowledge the co-pilot stage built up? How do you prevent Stage 1 from becoming permanent?

These are not strategic questions. They are operational ones. They require answers at the level of the manager, the team, the workflow — not the boardroom slide. The map above tells you where your work sits and how the stages play out. Building through those stages deliberately is a different kind of work entirely: specific, sequenced, and almost entirely absent from the AI adoption playbooks most organizations are running from right now.

That is what the second part of this addresses.

Your company rolled out an AI tool six months ago. Maybe it was Copilot. Maybe it was a custom GPT wrapper built by some team. Maybe it was an enterprise licence for ChatGPT, with a new policy against putting customer data into it.

A few people — you probably know exactly who they are — went completely down the rabbit hole. They’re using it for everything. They come to meetings with AI-generated summaries, AI-assisted slide decks, and AI-drafted proposals. They talk about it constantly.

Everyone else installed it, tried it once or twice, and went back to their normal workflow.

Meanwhile, your LinkedIn feed looks like this: AGI is 18 months away. AI is coming for white-collar jobs. The companies that don’t adopt AI now will be gone in five years.

Something doesn’t add up.

If AI is everywhere, if the headlines are right, if the revolution is already here — why doesn’t it look like that from inside most organizations? Why does the rollout feel less like transformation and more like a new app nobody’s sure how to use?

The answer isn’t that the headlines are wrong about the technology. It’s that they’re measuring something completely different from what’s actually happening in most workplaces. And understanding that gap is the most useful thing you can do right now.

Two numbers that both tell the truth

McKinsey’s 2025 global AI survey found that 88% of organizations now use AI in at least one business function. That gets picked up everywhere. It becomes the headline. 88% of enterprises are using AI.

That same year, the US Census Bureau published data showing that 5.4% of American businesses use AI to produce goods or deliver services.

Both numbers are accurate. They’re measuring completely different things.

The McKinsey survey asks executives at large companies whether their organization uses AI anywhere, in any capacity, at any level of commitment. If the marketing team has a Copilot licence and one person uses it to brainstorm email subject lines twice a month, that counts. They’re in the 88%.

The Census survey asks whether the business is actually using AI in its operations — in production, in the work product, in what goes out the door. That’s a much harder bar to clear.

One number describes access and experimentation. The other describes integration. Both are real. But when people argue about where AI adoption “is,” they’re usually citing the first kind of number while thinking about the second.

There’s a name for what the Census is measuring: process-integrated AI. It means AI is embedded in the way work actually gets done — in repeatable, expected workflows, with the organization’s knowledge and support. Not someone’s personal hack. Not a pilot that’s been running for eight months without a decision. Not a tool that lives in one team’s corner of the business. The kind of adoption that shows up in how things are built, delivered, and measured.

That’s the number that matters. And it’s the one that barely gets discussed.

What adoption actually means

The confusion starts because “adoption” gets used to describe three fundamentally different things.

Tier 1 is access. You have the tool. You open it occasionally. It’s useful for polishing emails, summarising documents, and generating a first draft of something you’d have written anyway. Most people in most organizations sit here. It’s genuinely better than nothing. It’s also not transformative.

Tier 2 is integration. AI is woven into a repeatable workflow. It’s expected — not optional. The team has actually changed how it works, not just added a new tab to its browser. The process is documented, supported by IT, and part of how output gets measured. A minority of organizations and workers are here.

Tier 3 is transformation. The business model has changed because of AI. New products exist that couldn’t have existed before. Roles have been restructured. Core economics have shifted. This is genuinely rare.

The job displacement stories are about Tier 3. The headlines citing 88% adoption are describing Tier 1. And most enterprises, most workers, most industries — are moving slowly through the space between the two.

Here’s the number that puts this in focus: of McKinsey’s 88% of organizations using AI, only 7% describe themselves as “fully scaled” — meaning AI is deployed and integrated across the organization. 32% are still experimenting. 30% are in pilots. 31% say they’re scaling but haven’t finished [1].

The stat doing all the work in the scary headlines is the weakest version of adoption. The version that means “someone in the building has a licence.”

Why organizations get stuck

If the tools are good and the business case is clear, why aren’t more organizations at Tier 2 or 3?

Because technology has never been the hard part. It’s never been the hard part with any technology.

According to BCG’s research across 1,000 senior executives in 59 countries, 74% of companies have yet to show tangible value from their AI investments — only 26% have developed the capabilities to move beyond proof of concept. S&P Global found that organizations scrap an average of 46% of their AI projects between proof of concept and broad adoption.

These aren’t technology failures. They’re organizational ones.

The companies that do it well follow something McKinsey has documented in its highest-performing firms: roughly 10% of the effort goes to the algorithm, 20% to data and infrastructure, and 70% to the people, process, and cultural change required to make it stick. Most organizations get this exactly backwards. They invest heavily in the technology and assume the rest will follow.

It doesn’t.

This is visible in one of the starkest data points in recent enterprise research. When executives were asked whether their organization had successfully adopted generative AI, 75% of C-suite leaders said yes. When employees in those same organizations were asked the same question, only 45% agreed.

That gap — 30 percentage points between what leadership thinks happened and what the workforce experienced — is where most AI strategies go to die. The announcement was made. The licences were bought. The transformation was declared. The behaviour never changed.

There’s one more signal that underlines this. 78% of people who use AI tools at work are using tools their IT department didn’t sanction — personal accounts, consumer products, workarounds. They found their own solution because the official one didn’t work for them, or didn’t exist, or was too restricted to be useful. That’s individuals solving a problem the organization failed to solve for them. It’s resourceful. But it’s not adoption. It’s shadow usage, and it means the organization has no visibility into what’s happening, no ability to measure it, and no way to build on it.

Where we actually are

Strip away the noise, and the picture looks roughly like this.

The tech sector is close to AI-native. 85% of software developers now regularly use AI tools, with more than half using them daily. In a 2024 study by GitHub and Accenture across enterprise development teams, 90% of developers had committed at least some AI-suggested code to their codebase. This isn’t the future — it’s already the baseline for anyone writing software professionally. The tech industry will likely be the first major function to hit genuine 75% process integration, probably in the next year or two.

Marketing and sales are developing fast, with measurable improvements in campaign throughput and deal cycles for organizations that have made the investment in proper integration.

Finance is moving methodically, cautiously, with good reason — the regulatory stakes are higher and the tolerance for AI-generated errors is lower.

Manufacturing and healthcare are years behind, held back by the complexity of integrating AI with physical systems, legacy infrastructure, and heavily regulated environments.

Retail is last.

Across all of these, when you normalise the research to what it’s actually measuring, the honest estimate for process-integrated AI — AI that’s genuinely embedded in standard work processes — sits somewhere between 15% and 45% of the large-enterprise workforce. The range is wide because the measurement is genuinely hard and the surveys genuinely disagree on definitions. But 15–45% is the reality. Not 88%. Not 5%.

The people inside organizations who feel like AI adoption is happening slower than the headlines suggest aren’t wrong. They’re just measuring the right thing.

When does this actually change?

Here’s what the data suggests about when 75% of the enterprise workforce will have AI genuinely embedded in their standard work processes — not just access to a tool, but integration into how the work gets done [12].

Optimistic scenario: 2029. This requires companies to rapidly convert tool access into workflow redesign, and it assumes that agentic AI — systems that can actually execute multi-step tasks without human intervention — becomes widespread enough to lower the integration barrier for workers who aren’t proactively upskilling.

Most likely scenario: 2032. Current friction patterns hold. Platforms do some of the heavy lifting as AI gets baked into Microsoft 365, Salesforce, and the tools people already use. Progress is real but uneven.

Conservative scenario: 2040. The pilot trap persists. The skills gap widens. Laggard sectors — retail, manufacturing, parts of healthcare — drag the average out significantly.

For context: the personal computer took roughly 20 years to reach majority adoption. The internet took about seven or eight. Generative AI hit 50% “any use” in under three years — genuinely unprecedented in the history of technology diffusion.

But “any use” has never been the same as transformation. The spread of the tool has always been faster than the spread of the behaviour change it enables. Email reached most knowledge workers in the 1990s. Most organizations spent another decade printing and filing every email they received — the tool was there long before the behaviour changed.

The predictions that AGI is just around the corner and the 2029-to-2040 adoption window are not contradictions. They describe two different things. The model can be extraordinary. The organization adopting it is still run by humans, running on processes built over decades, managed by people whose incentives don’t always reward disruption.

What this actually means for you

Here’s the honest version: the transformation is real, and its pace is going to accelerate. The companies that solve the human problem first — that crack the 70% of change management that everyone else is skipping — will pull ahead fast. That gap between leaders and laggards is already visible and it will compound. If you work at one of the laggards, you will feel it.

But here’s what the data actually gives you: time, and a window most people aren’t using well.

The headlines are designed to feel urgent, because urgency gets clicks. The reality is a transition measured in years, with a curve that rewards people who start building genuine competence now over people who wait for the official training programme.

Most people are still at Tier 1. Opening ChatGPT to draft emails and run glorified Google searches. That’s fine as a starting point. It’s not a strategy.

The move is to go further. Take something that actually matters to you — a problem you’ve been putting off, a decision that’s genuinely complex, a workflow that’s always been painful — and actually work it with AI. Not just ask it to clean up your writing. Give it something hard and pay attention to what happens.

You’ll find places where it’s extraordinary. You’ll also find where it falls apart. Where it’s confidently wrong. Where it needs you to know what good looks like before you can tell whether the output is any good. Where the limitation isn’t the model — it’s the quality of the question you asked. Those discoveries are the education. That’s the fluency you’re building.

Take a course. Experiment at home. Try things that feel slightly beyond your comfort zone at work. The people who will be most valuable on the other side of this transition aren’t the ones waiting for HR to organise a workshop. They’re the ones who already know how to use the tools, know their limits, and know how to work with AI rather than just alongside it.

The window is real. The data says you have years, not months.

But it’s not indefinite. And right now, most of the people you’ll compete with aren’t using it.

— -

References

[1] McKinsey & Company, *The State of AI 2025*, McKinsey Global Survey, November 2025. https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai

[2] US Census Bureau, *Business Trends and Outlook Survey (BTOS): AI Supplement*, Working Paper CES-WP-24–16, March 2024. https://www.census.gov/library/working-papers/2024/econ/ces-wp-24-16.html

[3] BCG, *Where’s the Value in AI?*, October 24, 2024. https://www.bcg.com/press/24october2024-ai-adoption-in-2024-74-of-companies-struggle-to-achieve-and-scale-value

[4] S&P Global Market Intelligence, *Voice of the Enterprise: AI & Machine Learning, Use Cases 2025*, May 2025. https://www.spglobal.com/market-intelligence/en/news-insights/research/ai-experiences-rapid-adoption-but-with-mixed-outcomes-highlights-from-vote-ai-machine-learning

[5] Writer, *Enterprise AI Adoption Survey 2025*, 2025. https://writer.com/blog/enterprise-ai-adoption-survey/

[6] Microsoft, *Work Trend Index 2024*, May 8, 2024. https://www.microsoft.com/en-us/worklab/work-trend-index

[7] JetBrains, *State of Developer Ecosystem 2025*, October 21, 2025. https://blog.jetbrains.com/research/2025/10/state-of-developer-ecosystem-2025/

[8] Stack Overflow, *Developer Survey 2025*, December 2025. https://survey.stackoverflow.co/2025/

[9] GitHub & Accenture, *Quantifying GitHub Copilot’s Impact in the Enterprise*, May 13, 2024. https://github.blog/news-insights/research/research-quantifying-github-copilots-impact-in-the-enterprise-with-accenture/

[10] Worklytics, *2025 AI Adoption Benchmarks: Employee Usage Statistics*, 2025. https://www.worklytics.co/resources/2025-ai-adoption-benchmarks-employee-usage-statistics

[11] Deloitte, *State of AI in the Enterprise 2026*, January 2026. https://www.deloitte.com/us/en/what-we-do/capabilities/applied-artificial-intelligence/content/state-of-ai-in-the-enterprise.html

[12] Projection scenarios modelled from synthesised enterprise adoption data anchoring Deloitte [11], McKinsey [1], and Gartner GenAI deployment research. For full methodology see: research synthesis at articles/adoption/_research/deep-research-report.md

[13] Federal Reserve Bank of St. Louis, *The State of Generative AI Adoption in 2025*, November 2025. https://www.stlouisfed.org/on-the-economy/2025/nov/state-generative-ai-adoption-2025

My first instinct wasn’t frustration. It wasn’t “I’ll figure out a workaround.” It was: I guess I’m on vacation until March.

Losing AI access meant all work stops. Not some work. Not the AI-specific work. All of it. That probably sounds crazy to you. But that reaction told me something I hadn’t fully articulated yet. Somewhere in the last two months, AI stopped being a tool I use and became a team I work with.

That distinction matters more than anything else I’ll say in this piece.

How I Used to Work

At the end of 2025, I was working like most of you probably do right now. Take on a project, have some chats with ChatGPT or Claude along the way, get help brainstorming ideas or drafting a section of a report. Use image generation for visuals in a deck. Maybe let AI help outline a workshop.

It was useful. It saved time here and there. But the work was still fundamentally mine to do. AI was an assistant — a smart search engine with better manners.

I was the one doing the industry research. Writing the analysis. Building the workshop in Mural. Structuring the JTBD map. AI handed me ingredients. I did the cooking.

If you’d taken away my AI access back then, I would have been annoyed. But I would have kept working. Losing it would have been like losing a good calculator — inconvenient, not debilitating.

That’s not where I am anymore.

What Changed

In early 2026, I stopped chatting with AI and started working with agents.

The difference sounds subtle. It’s not.

Chatting with AI means you’re the worker and AI is your research assistant. You ask questions. You get answers. You do the work.

Working with agents means you’re the director and they’re your production team. You set the vision. You define the quality standards. You review and refine what they produce. But the actual production — the research, the drafting, the data gathering, the formatting, the iteration — that’s happening around you, often in parallel.

Here’s what my workflow actually looks like now. I work in innovation — Jobs to Be Done mapping, client workshops, industry research, identifying unmet needs, ideating new solution concepts, rigorous theory testing, proof of concept, prototyping. The kind of work that used to mean weeks of solo research, synthesis, and deliverable production.

Now I have agents for it. I run them through Claude Code — a command-line interface where I can orchestrate agents, define their roles, and build structured workflows. When I need to map a competitive landscape, a research agent digs through industry data, identifies patterns, and surfaces insights I’d have spent days finding manually. When I’m building a JTBD analysis, an agent structures the framework while I focus on the strategic judgment — which jobs matter, which are underserved, where the real opportunity sits. When I need to pressure-test a concept, an agent runs it through evaluation criteria while I decide what questions to ask next.

I direct the process, make decisions at each stage, and apply my judgment to the output. I’m not copy-pasting from a chatbot. I’m running a production operation.

It extends to everything. Presentations, visual assets, written deliverables, workshop preparation — all of it runs through agents with context about my work, my standards, and my preferences. They aren’t generic. They know what “good” looks like for my specific work because I’ve given them the rules, the templates, and the constraints.

This is the part people miss when they hear “AI-generated content” and picture slop. Output quality isn’t a function of AI capability alone — it’s a function of the *system* you build around it. Context, rules, guardrails, and human judgment at the right moments. When that system is well-designed, the output is frequently better than what I’d produce solo. Not because the AI is smarter than me, but because it can explore more options, maintain more consistency, and execute more thoroughly than one person working alone.

I’ve been tracking my output this month. Conservatively, I’m producing 10x what I could do manually. On the projects where agents handle the most production work — deep industry research, multi-layered analysis, concept development — it’s closer to 27x. Those numbers aren’t precise, but the direction is unmistakable. This isn’t a marginal improvement. It’s a different category of work.

The Director’s Chair

The best metaphor I’ve found for this is filmmaking.

A film director doesn’t hold the camera. They don’t build the sets. They don’t edit the footage frame by frame. But the movie is unquestionably theirs. Their vision, their decisions, their quality bar.

That’s what agentic co-work feels like. I act as director, editor-in-chief, CEO of my own projects. I am frequently not the doer, but the leader. The agents are my cinematographer, my production designer, my sound editor — specialists who execute at a level I couldn’t match individually, but who need my direction to produce something coherent and intentional.

Sam Altman described this exact dynamic at the Snowflake Summit: “You hear people that talk about their job now is to assign work to a bunch of [AI] agents, look at the quality, figure out how it fits together, give feedback.” He’s describing me. He’s describing a growing number of people. And soon, he’ll be describing the baseline expectation for knowledge work.

Here’s what people get wrong, though: they assume this means less work. It doesn’t. Directing is harder than doing, in some ways. You have to know what good looks like before you see it. You have to articulate standards you used to keep in your head. You have to make dozens of judgment calls per hour instead of just executing a task.

The skill set shifts from production to curation. From creation to orchestration. The effort doesn’t disappear — it concentrates on the parts that only a human can do.

But the leverage? The leverage is something else entirely.

The Numbers Behind the Shift

My experience isn’t an outlier. The data tells the same story at scale — with a split ending.

According to Gallup’s Q4 2025 workplace survey, 49% of U.S. workers have never used AI at work. Nearly half. Meanwhile, daily AI use has climbed to 12%, and in tech, 57% are frequent users.

This isn’t a bell curve. It’s a bimodal distribution. People are either in or they’re out.

Among leaders, frequent AI use has risen from 17% to 44% since mid-2023. Among individual contributors? From 9% to 23%. Leaders are pulling ahead — not just in adoption, but in the kind of work AI makes possible.

Gartner predicts 40% of enterprise applications will feature task-specific AI agents by the end of 2026. In 2025, that number was less than 5%. An eight-fold increase in a single year.

Here’s the number that should stop you: according to OpenAI’s State of Enterprise AI report, frontier workers send 6x more messages than the median employee — and in coding and data analysis, that gap widens to 16-17x. Workers who save more than 10 hours per week aren’t just using AI more often. They’re using it across more tasks, with more tools, in more integrated ways. The gap isn’t between people who use AI and people who don’t. It’s between people who’ve restructured their work around AI and people who treat it like a fancier Google.

Yet according to Deloitte’s Tech Trends 2026 research, only 11% of organizations have AI agents in active production — despite 38% piloting them. Most are still experimenting. Still writing strategy decks about whether to try.

And the tools are catching up to the ambition. Anthropic just launched Claude with co-work — a visual, user-friendly platform for running agents without touching a terminal. The creator of popular open-source agent tools has been hired by OpenAI, signaling accessible agent interfaces coming from every major player. The barrier to entry is dropping fast.

The gap isn’t just between individuals. It’s between people like me who’ve already built these systems and organizations that haven’t started. And soon, they won’t even have the “it’s too technical” excuse.

The Uncomfortable Part

Let’s be honest about what this means.

If you’re still in the “chat with AI sometimes” phase, you’re not behind — but the ground is moving under you. The people who’ve made the shift to agentic work aren’t a little more productive. They’re operating in a fundamentally different mode. It’s the difference between emailing your team and *having* a team.

I know this lands differently depending on who you are.

If this makes you skeptical — fair. “AI changed everything” is a tired headline. But I’m not making a prediction. I’m reporting what happened to me, with the data to back it up.

If this makes you uncomfortable — sit with that. Not because AI is coming for your job tomorrow, but because the people who figure this out will operate at a level that redefines what “productive” means. That changes what employers expect and what clients demand.

If this excites you — you’re the right audience for what comes next.

The Dependency Question

I know what you’re thinking, because I had the same thought staring at that API limit screen: Isn’t this dangerous? Haven’t you made yourself completely dependent on a tool?

Yes. And I’m fine with that.

We’re already dependent on electricity. On the internet. On cloud services. On our phones. Nobody calls those dependencies “dangerous” anymore — they’re infrastructure. The question was never whether to depend on powerful tools. It was always about whether to be the last person doing it the old way while everyone else compounds their advantage.

My “vacation until March” moment wasn’t a failure. It was confirmation that the system I’d built was working so well I couldn’t imagine going back. That’s not weakness. That’s what leverage feels like from the inside.

The risk isn’t dependency. The risk is irrelevance.

Where to Start

If you’ve read this far, you’re probably wondering how to make this shift. There’s no quick playbook, but here’s what I’ve learned:

Stop chatting, start building. The difference between using AI and working with AI is structure. Give it a role. Give it context about your work. Give it quality standards. Turn a conversation into a workflow.

Think in systems, not sessions. A single AI chat is useful. A set of agents with defined roles, shared context, and quality rules is transformative. The magic isn’t in any one interaction — it’s in the architecture you build around them.

Embrace the role shift. You’re not being replaced. You’re being promoted. The skills that matter now are judgment, direction, quality control, and taste. Those are human skills. They’re leadership skills. And they’re suddenly the bottleneck.

Start before you’re ready. I went from chatting with Claude to directing a team of agents in about two months. You don’t need to be a developer. You need to be willing to experiment. The tools are meeting you where you are — and they’re getting easier by the month.

The window is open right now. Not next quarter. Not when your organization rolls out an AI strategy. Now.

Here’s the thing about exponential shifts: by the time they’re obvious to everyone, the advantage belongs to the people who moved first.

I learned that staring at an API limit screen, realizing my entire working life had been quietly, fundamentally transformed — and I’d barely noticed it happening.

Don’t wait for your own “vacation until March” moment to see what’s changed.

---

If this resonated, share it with someone who’s still in the “chat with AI sometimes” phase. They need to hear this more than you did.

Kevin Roose put it perfectly last month:

“I follow AI adoption pretty closely, and I have never seen such a yawning inside/outside gap.”

On one side: people running multi-agent claudeswarms, building personal AI assistants, tinkering with open-source projects like OpenClaw — an autonomous AI agent — that rack up 145,000 GitHub stars in weeks. On the other: people still waiting for IT to approve Copilot in Teams.

Six months ago, the mainstream take on AI was that it was a bubble — overhyped, underdelivering. That narrative is dead. In 2026 alone, the four biggest tech companies are spending north of $650 billion on AI infrastructure. Not because they think this might work — because AI agents are already in production, already replacing workflows, already repricing entire industries. When a legal AI demo wiped 16% off Thomson Reuters’ market cap in a single trading session, that wasn’t hype. That was the market recalculating who does the work.

Same technology. Same moment. Parallel universes.

I’ve been watching this gap up close — building AI agents for innovation and brand governance work. But the clearest picture of it didn’t come from work. It came from my kitchen table.

The Kitchen Table Experiment

My wife sat down with Claude to do some data analysis for work. She’d heard it was good with spreadsheets. Real project, real data, real deadline.

Her first attempt wasn’t even close.

So she tried again. Closer this time, but it grabbed too much — pulled in a term that looked similar to what she needed but shouldn’t have been in the analysis. Wrong bucket. Close enough to fool a casual glance, specific enough to matter.

A few more rounds. Each time she refined what she was asking for, the output got tighter. She started reading the results more carefully — not just checking if it worked, but understanding why it worked or didn’t.

By the end, she had what she needed. And for the first time in months, she was genuinely curious about what I do every day when I head down to the basement office.

She played with it — failed, adjusted, played again. And in that process, something clicked. Not just about the tool, but about the possibilities.

That’s not a training story. That’s a play story.

88% Adoption. 7% Scaled. What’s Missing?

The data on AI adoption looks impressive until you read the fine print.

McKinsey’s 2025 State of AI survey found that 88% of organizations now report using AI — up ten points from the year before. Sounds like the gap is closing.

Except only 7% have actually scaled it. Two-thirds are still in pilot mode. And only 39% report any measurable impact on the bottom line.

Microsoft’s 2024 Work Trend Index tells a similar story from the individual side: 75% of knowledge workers say they use AI at work. But 78% of them are bringing their own tools — personal subscriptions, not company-provided. They’re using what they found by messing around on their own time.

Here’s the stat that should haunt every executive: only 39% of people using AI at work have received any training from their company.

Here’s the part that should make that stat sting: the adoption window is compressing faster than any previous technology shift. Railroads took twenty years to justify the investment. Fiber optics took ten. AWS took six. The current AI infrastructure cycle? Roughly eighteen months. The gap between early and late is getting shorter, and “I’ll figure it out next quarter” is a riskier bet than it sounds.

So the picture is: almost everyone has access. Almost no one has been taught. And the people who are actually good at it? They taught themselves — by playing.

The standard diagnosis for the adoption gap is access, training, or strategy. The data doesn’t support any of them. Access? Nearly universal. Training? Barely correlates with fluency. Strategy? Documents don’t make people better at prompting.

The real gap is between people who gave themselves permission to experiment and people still waiting for someone to tell them it’s okay.

Why Play Works (And Training Doesn’t)

This isn’t just anecdotal. The research on how adults learn novel technology is surprisingly clear — and it favors tinkering over instruction.

People who tinker with new software make more errors at first — but develop deeper mental models and transfer skills far more effectively than those who follow manuals. IBM researchers figured this out in the ’80s. It still holds. The messy early phase wasn’t wasted time. It was the foundation of fluency.

Nothing lights up the brain like play. That’s not metaphor — it’s neuroscience. Stuart Brown, a psychiatrist who’s studied play across 6,000+ life histories, explains why: when you’re playing, you’re in a state of open-ended exploration — trying things because you’re curious, not because you’re following a syllabus. You’re building intuition, not just knowledge.

The key insight: play reduces the cost of failure. When my wife’s first prompt didn’t work, she wasn’t stressed — even though it was for a work project. She was curious. Huh, that didn’t work. Let me try it differently. The task was work, but her approach was play. That distinction matters. If she’d been following a manual or sitting in a training session, that same failure might have been discouraging enough to quit.

Ethan Mollick at Wharton has been making this case for two years: there is no substitute for trying AI yourself. Reading about it is not enough. Watching demos is not enough. You need to use it, and you need to use it for things you care about.

Not things your company tells you to care about. Things you care about. That’s the difference between training and play.

The Permission Problem

If play is the answer, why isn’t everyone playing?

Because play requires something most organizations systematically crush: the safety to fail.

Amy Edmondson’s research found that psychological safety — the belief you won’t be punished for mistakes — is the single strongest predictor of high-performing teams. People experiment when they feel safe failing. They wait for instructions when they don’t.

Carol Dweck maps the same terrain from the individual side. The person who thinks “I don’t know how to use this yet” will play. The person who thinks “I’m not a technical person” will wait. The difference isn’t skill. It’s orientation.

Now look at how most organizations have approached AI. Risk assessments. Approved tool lists. Mandatory training. Compliance requirements. Every signal says be careful. Every process says wait for permission.

No wonder 78% of AI users are bringing their own tools. The tinkerers aren’t waiting.

This is the same pattern we’ve seen with every technology wave. Before Slack was officially approved, teams were already using it. Before ChatGPT was on the approved vendor list, people were pasting work problems into it at home. The people who adopt technology first aren’t more technical. They have a higher tolerance for acting without permission.

But here’s what matters for leaders: this isn’t just an individual trait. It’s an environmental one. Some organizations create conditions where play happens naturally. Others make it nearly impossible.

And there’s an equity dimension worth naming. Knowledge workers with autonomy can experiment without scrutiny. Front-line workers, those in regulated industries, those with less job security — they can’t afford to “play” at work. The permission gap maps to existing power structures.

“Just play with it” isn’t equally available advice. Which is exactly why creating the conditions for play is a leadership responsibility, not just an individual one.

Play Isn’t Reckless

I should be clear: play doesn’t mean chaos. And sandboxes aren’t just about protecting data — they’re about making play accessible to the people who can’t afford to experiment unsupervised.

The best organizations don’t choose between control and experimentation. They create bounded spaces where play is safe — tools with data guardrails, internal instances where a bad prompt costs nothing. As Tim Brown at IDEO has argued, play is not anarchy — it has rules, especially when it’s in the service of innovation.

And play alone isn’t enough. A Harvard/BCG study of 758 consultants found that people who blindly followed AI output actually performed worse than people without AI at all. That study is from 2023 — ancient in AI years. But the finding holds: tools outpace judgment unless you develop the judgment through experience. Play without reflection is just entertainment. The real learning comes from the full cycle: try something, notice what happened, adjust, try again.

My wife didn’t just retry the same prompt. She noticed what went wrong — the mismatched term, the too-broad dataset — and adjusted her approach. That’s play plus reflection. That’s how fluency develops.

Two Audiences, Two Actions

If you lead a team or organization:

Stop building training programs. Start building playgrounds.

Provide AI tools with appropriate data boundaries. Give people genuine unstructured time to experiment — not a workshop, real time. Reward curiosity, not just compliance. Make it safe to share failures. Stop treating AI adoption like a risk to manage and start treating it like a capability to develop.

The McKinsey data is clear: the organizations capturing value from AI have cultures that encourage experimentation and bottom-up adoption. Not the best strategy decks.

If you’re an individual:

Stop waiting for permission.

Open Claude or ChatGPT tonight. Not for work — for something you actually care about. Plan your next trip without opening ten browser tabs. Figure out where your money actually goes each month. Write a bedtime story for your kid where they’re the hero.

See what happens.

It will feel clumsy. Your first prompt will be too vague, or too specific, or just plain wrong. You’ll stare at the output and think that’s not what I meant at all. Good. That’s exactly what my wife felt at the kitchen table — and it’s the feeling right before the thing clicks.

The gap between your first prompt and your tenth is where fluency lives.

The people on the other side of Roose’s gap didn’t get there through training. They got there by playing. And the distance between those two sides is growing faster than most people realize — not in years, but in months. It always started the same way: one evening, one question they actually cared about, one screen glowing on a kitchen table. Tonight could be yours.

— -

I build AI agents and write about the human side of working with them. If you’re navigating this gap — from either side — I’d like to hear what you’re seeing.

A worker is given a task: review a stack of documents and produce a summary report. Somewhere around the fortieth page, things start to drift. The categories get fuzzy. A few details appear that don’t match the source material. By the end, the report includes plausible-sounding facts that were never in the originals — invented names, fabricated figures, delivered with full confidence. Nobody catches it until someone checks the sources.

Now here’s the second story. A worker is given a task: update a complex system based on requirements from three different teams. Somewhere around the third revision, things start to drift. The original constraints get fuzzy. A few decisions contradict earlier ones. By the end, the system includes features nobody asked for and is missing features everybody assumed were included. Nobody catches it until the demo.

One of these is an AI. One is a human. The failure mode is identical: when context disappears, systems — biological or digital — don’t stop working. They start making things up.

(It doesn’t matter which is which. That’s the point.)

The Problem Has a Name Now

We’ve been living with context loss for as long as humans have organized into groups. Every time a senior engineer quits and takes undocumented knowledge with them. Every time a project gets handed off and the new team rebuilds from scratch because the brief was a two-sentence Slack message. Every time a meeting produces action items that contradict the last meeting’s action items because nobody reviewed the notes. This is context loss. We’ve just never had good language for it.

AI gave us the language.

In large language models, context loss is measurable and precise. An LLM has a context window — a finite amount of information it can hold in working memory at any given moment. Research from Stanford has shown that when you put critical information in the middle of a long context, model performance degrades significantly — a phenomenon researchers call being “lost in the middle.” The model doesn’t flag that it’s struggling. It doesn’t ask for help. It just gets worse — confidently.

A 2025 study went further: even when a model can perfectly retrieve every piece of evidence in its context (reciting tokens with 100% accuracy), its ability to reason about that information still degrades as the context grows. The problem isn’t memory. It’s comprehension under load. The system has the information. It just can’t use it anymore.

Sound familiar? It should.

The Human Version

In 1975, Fred Brooks published The Mythical Man-Month and gave us a formula that should be tattooed on the forearm of every project manager: n(n-1)/2.

That’s the number of communication channels required when n people need to coordinate. Two people need one channel. Five people need ten. Ten people need forty-five. Fifty people need 1,225. The relationship isn’t linear. It’s combinatorial. And every channel is a potential point of context loss.

Brooks’s central insight — “adding manpower to a late software project makes it later” — has been validated across thousands of software projects over the following decades. The reason is always the same: new people don’t arrive with context. They have to acquire it. And acquiring context consumes the very resource (experienced people’s time) that was already scarce.

But Brooks only described the scaling problem. The deeper issue is what happens to context once it’s distributed across multiple minds.

According to research from IDC, Fortune 500 companies lose an estimated $31.5 billion per year to what researchers call “institutional forgetting” — or, less politely, corporate amnesia. Employees spend a full fifth of their work time searching for knowledge that already exists somewhere in the organization. Bloomfire’s 2025 research found that knowledge mismanagement costs companies an average of 25% of their annual revenue. The knowledge exists. It’s just distributed across so many memory banks — people’s heads, email threads, abandoned Confluence pages, someone’s notebook from a meeting in 2023 — that it might as well not exist at all.

The Mirror

The AI systems we’re building to solve these problems are developing the exact same failure modes. For the exact same structural reasons.

In 1967, Melvin Conway observed that ”organizations which design systems are constrained to produce designs which are copies of the communication structures of those organizations.” Conway’s Law has been a truism in software engineering for decades. It’s now becoming the defining law of AI agent architecture.

A company builds an agentic AI system. Customer support builds a support agent. Sales builds a lead qualification agent. Finance builds an invoice processing agent. Legal builds a compliance agent. Each agent is optimized for its domain, trained on its department’s data, governed by its department’s rules. The resulting system — a network of specialized agents that struggle to communicate across boundaries — is a pixel-perfect copy of the org chart.

The fault lines between agents follow the same boundaries as the fault lines between human teams. Agents owned by the same group share data easily. Cross-team agent collaboration is rare, brittle, and expensive. When an agent needs context that lives in another agent’s domain, the handoff is lossy. Information gets summarized, compressed, or dropped entirely.

Research on multi-agent systems from UC Berkeley quantifies this. Failure rates across seven major multi-agent frameworks range from 41% to 87%. These systems consume far more computational resources than theoretically necessary, not because the individual agents are inefficient, but because the coordination overhead is enormous.

The math is the same math Brooks described in 1975. Five agents create ten potential handoff pathways. Ten agents create forty-five. But it’s not just that more pathways exist — it’s that the consequences are identical. Each handoff is a point where context degrades, where nuance gets compressed into summary, where the receiving agent proceeds with slightly wrong assumptions.

As the Berkeley research team put it: ”Most ‘agent failures’ are actually orchestration and context-transfer issues.”

Replace “agent” with “team member” and you’ve described every failed enterprise project in history.

Context Rot

Context loss would be manageable if it were static — if you lost a fixed percentage of information at each handoff and could plan around it. But context doesn’t just disappear. It rots.

The term “context rot” emerged in mid-2025 to describe what happens when model performance systematically degrades as input context length increases, even when the underlying task remains simple. It’s not that the model forgets. Each piece of information slightly distorts the model’s processing of every other piece. The distortions compound. Speaker identities get confused. Timelines warp. Unrelated facts merge. The model doesn’t crash. It drifts.

A longitudinal study of multi-agent AI systems measured this drift precisely. Over extended interactions, task success rates dropped by 42%. Response accuracy declined by nearly 25%. The number of times humans had to intervene increased by 216%. And the pattern was consistent: the agent didn’t suddenly malfunction. It slowly became less reliable as ”each slightly wrong memory entry influenced the next summarization, which influenced the next recall, which influenced the next decision.”

Organizations experience the same compounding decay. It’s just slower, harder to measure, and easier to misattribute.

Every “let me loop in Sarah on this” is an expansion of the context window. Every set of meeting minutes is a lossy compression of the original conversation. Every status meeting where twelve people sit through updates relevant to three of them is noise being injected into the signal. Every time a project gets restructured and responsibilities shift, the tacit knowledge about why certain decisions were made — the context that never made it into documentation — drifts a little further from the people who need it.

Organizational knowledge doesn’t sit still. It degrades with every transfer, every translation, every new person who inherits a project and fills in the gaps with assumptions. Knowledge management is entropy management — slowing the inevitable diffusion of critical context into inaccessible corners of the system.

The stakes vary by domain. In knowledge work, context rot costs money — poor communication contributes to more than half of project failures , according to PMI research. In AI, it costs credibility — Gartner predicts 40% of agentic AI projects will be canceled by 2027. In healthcare, where communication failures are the leading cause of sentinel events, it costs lives — linked to an estimated 150,000–250,000 patient deaths annually. The mechanism is always the telephone game: information distorting as it passes through nodes, each handoff stripping away nuance that seemed unimportant to the sender but turns out to be critical for the receiver.

Different domains. Same rot. Same math.

Context Engineering

Companies are racing to deploy multi-agent AI systems — Gartner reports a 14x surge in multi-agent system inquiries from early 2024 to mid-2025 — while simultaneously failing to manage context in their human organizations. They’re building AI architectures that inherit all the communication pathologies of the teams that built them, then wondering why the AI systems exhibit the same coordination failures.

Jensen Huang at CES 2025: ”The IT department of every company is going to be the HR department of AI agents in the future.” If that’s true, then organizations deploying agents at scale are about to confront context loss at a speed and visibility they’ve never experienced before. When a human team loses context, the failure unfolds over months. People compensate. They ask around. When an AI agent loses context, the failure is immediate, measurable, and reproducible. You can’t explain away a 42% drop in task success in a quarterly review.

That visibility is an opportunity. Because it turns out the solutions are the same for both systems.

There’s a reason “context engineering” is becoming the defining skill of AI development. The people building the most capable AI systems have realized that the quality of your context matters more than the power of your model. You can upgrade to the latest frontier model, but if your context is garbage — irrelevant information, missing information, poorly structured information — you’ll get garbage output with better grammar.

The same is true of organizations. You can hire the most talented people in the world, but if critical context lives in one person’s head, if handoff protocols don’t exist, if institutional memory resets every time someone leaves, those talented people will spend their time rediscovering things the organization already knew.

Anthropic’s engineering team captured the underlying principle: ”The quality of an agent often depends less on the model itself and more on how its context is structured and managed.”

A mediocre team with excellent knowledge management will outperform a brilliant team drowning in silos. That’s always been true. We just didn’t have a precise way to say it until we started watching AI agents fail in exactly the same ways our teams do.

Five Defenses Against Context Rot

The disciplines that prevent context rot in AI systems are the same ones that prevent it in human organizations. Here’s how to apply them:

1. Bound the state

For AI: Effective agents maintain a compact cognitive state — goals, constraints, confirmed decisions — rather than trying to hold everything in context.

For teams: Keep a living “project state” document that fits on one page. Goals, current blockers, recent decisions, next actions. If someone asks “where are we?” the answer should take thirty seconds, not thirty minutes.

The test: Can a new person (or agent) get oriented in under five minutes? If not, your state is too diffuse.

2. Structure the handoffs

For AI: Free-text handoffs between agents are the leading source of context loss. The fix is schema-validated payloads — structured formats that force explicit transfer of key information.

For teams: Create a handoff template and actually use it. What’s the current state? What decisions have been made and why? What’s been tried and didn’t work? What does the next person need to know that isn’t written down anywhere?

The test: Could someone take over this project from your handoff document alone, without a synchronous conversation? If not, your handoff is too loose.

3. Externalize the memory

For AI: Retrieval-augmented generation pulls relevant information from external knowledge bases on demand rather than stuffing everything into the context window.

For teams: Build a knowledge base that people actually use. Decision logs. Post-mortems. A searchable record of why things are the way they are, not just what they are.

The test: When someone asks “why did we decide X?” — can you find the answer in under two minutes without asking a person? If not, your memory is too volatile.

4. Budget the attention

For AI: Context windows are finite. Every irrelevant document chunk you inject degrades the system’s ability to process what matters.

For teams: Attention is finite. Every person you add to a meeting, every “FYI” email you send, every status update that could have been async — it’s all noise competing with signal.

The test: For every piece of information you’re about to share, ask: does this person need this to do their job? If not, don’t send it.

5. Consolidate regularly

For AI: Systems that periodically compress and consolidate their accumulated context show significantly less drift over extended interactions.

For teams: Sprint retrospectives, post-mortems, lessons-learned sessions. The practice of regularly stopping to ask: “What do we know now that we didn’t before? What should we remember? What can we let go?”

The test: Does your team have a regular ritual for turning experience into institutional memory? If not, you’re relying on individual recall — which means you’re one resignation away from forgetting everything.

 — -

The Mirror Works Both Ways

The parallel isn’t perfect. AI context windows and human working memory aren’t mechanically identical — an LLM can “look back” at earlier parts of its context in ways human short-term memory can’t. The analogy is structural, not literal. But the structural similarity is what makes it useful. Both systems degrade when overloaded. Both lose critical information at handoff points. Both compensate for missing context by filling in gaps — humans with assumptions, AI with hallucinations. And both respond to the same interventions.

The organizations that take context engineering seriously for their AI systems will inevitably start applying those same disciplines to their human teams. Once you see the failure mode clearly in one domain, you can’t unsee it in the other.

And the organizations that don’t? They’ll build AI systems that inherit every communication pathology their human teams already have, then scale those pathologies at machine speed.

AI didn’t invent context loss. It just made it impossible to ignore.

Our team had something most companies dream about: years of global customer survey data, synthesized into six representative personas -- not rough sketches, but detailed profiles built from thousands of data points.

The opportunity was clear. If we could bring these personas to life as AI -- make them conversational, responsive, testable -- we’d have something unprecedented: the ability to sit across from our customer and ask them anything. Test a positioning before the campaign. Pressure-test a strategy before the investment. Simulate reactions before millions were committed.

We built it. Version one nailed every question. On-brand, on-topic, thoroughly grounded in the research.

When the team tested the personas, the reaction was immediate: “It’s good, but something is off.”

They were right. Everything the personas said was accurate. None of it was real.

What If You Could Talk to the Research?

We had deep, well-synthesized customer intelligence. Six personas built from real data. The problem wasn’t the quality of the research -- it was the format. Static documents don’t let you ask follow-up questions. Slide decks don’t push back on your assumptions. Persona PDFs don’t tell you when your idea doesn’t land.

What if you could talk to the research instead?

That was the real promise. Not just understanding your customer, but simulating them. Testing with them. De-risking decisions alongside them -- in real time, before the stakes were real. Months of planning. Millions in budget. And the customer’s actual reaction remains a guess until the money is spent -- unless you can simulate first.

But only if the persona feels real enough to trust. And version one didn’t. The responses were accurate. They just weren’t believable.

Version 1: The Personality Lens

Our first architecture used two layers, roughly parallel to the Atkinson-Shiffrin model (https://en.wikipedia.org/) of human memory -- though we arrived at it through practice, not theory.

First, we built a personality for each persona from the survey data: identity, life stage, time reality (how much of it she has and how it shapes her priorities), emotional baseline, primary tension, self-narrative, voice and tone, decision logic, and explicit behavioral guardrails -- the things she would never do and the things she always does. This wasn’t a demographic sketch. It was a full cognitive profile: who this person is.

Second, we created what I call context vectors -- three interpretive filters that shape how each persona processes any input:

1. Emotional context. How she feels internally -- the emotional state coloring every reaction. A customer anxious about a major purchase processes a sales pitch differently than one excited about an upgrade.

2. Physical context. What’s happening around her -- the environment and bodily reality framing her attention. A busy parent scrolling on a phone at school pickup interprets a product offer differently than someone browsing at home on a Saturday morning.

3. External influencers. Who or what is driving pressure -- the people, institutions, or circumstances shaping her decisions. A customer whose partner is skeptical of the category brings that tension into every interaction, even when it’s never stated.

Personality defines who each persona is. Context vectors create the lens through which she interprets. Together, they form what I think of as the personality lens -- the system prompt sitting between the AI and every conversation, filtering everything through a specific human reality.

And it worked. Sort of.

Ask Persona 3 about a pricing change and you’d get a thoughtful reaction grounded in value sensitivity. Ask Persona 5 the same question and you’d get a different angle -- focused on trust and brand promise. The personalities were distinct. The context vectors made their interpretations feel different.

But ask either of them something that should trip them up -- a question touching a real tension in how customers think about the category -- and you’d get the same smooth confidence. No friction. No uncertainty. No sign the question was harder than any other.

Real customers don’t work that way. They light up on topics that touch their experience. They hesitate on questions that surface unresolved tensions. They reference that time when -- specific moments that shaped how they see things now.

In cognitive science terms, our personas had semantic memory (https://www.annualreviews.org) -- general knowledge about what each customer type thinks -- but no episodic memory. Tulving drew this distinction in 1972 and later formalized it as a theory of mental time travel (https://www.annualreviews.org: episodic memory is the stuff of lived experience, the specific situations and moments that make a perspective feel inhabited rather than constructed. A persona with only semantic memory knows about the customer. One with episodic memory has been the customer.

Version 1 had personality and interpretation. It didn’t have experience.

Version 2: The Memory Breakthrough

For version 2, we gave each persona a long-term memory.

The insight was straightforward, even if the engineering wasn’t: the original survey data wasn’t just information. It was experience. Customer verbatims captured not just what people thought, but how they reasoned through reactions in real time. Open-ended responses showed how they weighed competing feelings, which analogies they reached for, what they considered and dismissed.

We took all of it and built a vector store for each persona -- a searchable archive of question-and-answer pairs identified as source material for that specific persona. Every piece of customer expression was embedded and indexed, not as raw data, but as retrievable experience. Years of research that had been sitting in decks and spreadsheets, suddenly alive and searchable. When someone asked a persona a question, it didn’t just consult its personality lens. It searched its memories.

Now the system had three layers. Short-term memory: the context window of the active conversation. The personality lens: the system prompt containing both the persona’s identity and its context vectors. And long-term memory: the vector store of real customer expressions, grounded in specific situations.

It turns out this tracks with what you’d expect from the research. The Stanford team behind the 2023 generative agents paper (https://arxiv.org/) found that believable AI personas need three retrieval dimensions: recency, importance, and relevance. Remove any one and believability collapses. Their agents didn’t just respond -- they remembered, reflected, and evolved. We’d landed on a similar structural logic. The personality lens sets the character. Memory makes it real.

The results were immediately different.

Responses started pulling in specific situations from the research. A persona would reference a particular experience or recall a particular challenge, explaining its reaction with the kind of situated detail that only comes from lived context. It didn’t just know what the customer thinks. It knew why they think it, traced back to real moments. The difference was immediately obvious in testing -- the same question that used to produce a smooth, confident paragraph now produced a response that felt like someone thinking through their own experience.

Then something unexpected happened.

During a test session, one of the personas produced a response that included a pause. A hedge. A moment of visible processing where it circled an idea before landing on it, qualified before committing, thought out loud in a way that felt less like output and more like... thinking.

The team’s first instinct was to flag it as a bug. A disfluency. An imperfection in an otherwise coherent response.

Their second reaction, a beat later: “Wait. That’s exactly what a customer would do.”

Why “Um” Changes Everything

Here’s what most people get wrong about AI-generated content: they assume polish is the goal. That the best output is the most fluent, most structured, most grammatically pristine version possible.

The research says otherwise.

Clark and Fox Tree published a landmark study in Cognition (https://www.sciencedirect.com) showing that “uh” and “um” aren’t errors. They’re words -- interjections with specific communicative functions. “Uh” signals a short expected delay. “Um” signals a longer one. They tell the listener: I’m thinking. I haven’t rehearsed this. This is genuine.

Follow-on research confirmed the pattern. Listeners perceive speakers who include natural disfluencies as more genuine and thoughtful (https://www.sciencedirect.com) than those delivering perfect fluency. And Sandy Pentland’s work at MIT on ”honest signals” (https://mitpress.mit.edu) showed that unconscious communication patterns -- including speech tempo and pauses -- predict trust better than the content of what’s said.

Now consider AI text. Research published in PNAS (https://www.pnas.org) found that people use a telling heuristic: text that is “too polished” or “too generic” gets flagged as AI-generated. Human writing that happens to be well-structured is more likely to be mistakenly labeled artificial. We’ve internalized an expectation that authentic communication is imperfect. Polish has become a signal of inauthenticity.

This is the uncanny valley of text. Not wrong enough to reject. Too right to trust.

Think about it practically: you’re testing a new product concept with one of your AI personas. One gives you a perfectly structured response -- clear objections, balanced analysis, neat conclusion. The other starts with “So this is tricky, because on one hand...” and works through a messier but more revealing reaction. Which one do you trust? Which one actually helps you de-risk the decision?

Psychologist Elliot Aronson captured the mechanism in 1966 with the pratfall effect (https://link.springer.com): a highly competent person becomes more likable after a minor blunder. Imperfection humanizes. But the effect only works when competence is already established -- a blunder by someone perceived as incompetent just confirms the impression.

That’s exactly what happened with our personas. Version 2 had established competence -- deep, well-grounded responses drawn from real customer data. The occasional disfluency didn’t undermine it. It completed it. The system had absorbed enough of real customer cognitive patterns -- the tendency to think out loud, to qualify before committing, to circle an idea before landing on it -- that the imperfection wasn’t random. It was characteristic.

We didn’t need perfect personas. We needed faithful ones.

What the Work Taught Me

Two breakthroughs came out of this project, and they work as a pair.

The first is the filter. Giving a persona the right information is table stakes. The breakthrough is building the right filter -- the combination of personality and context vectors that turns raw information into a characteristic reaction. Two customers walk into the same product launch. One is energized by novelty; the other is suspicious of hype. Same pitch. Opposite decisions. The difference isn’t knowledge -- it’s the filter between input and response. Get the filter wrong and your simulation gives you consensus where the real market gives you conflict.

The second is human texture. The filter alone gave us differentiated responses, but they still felt constructed. It was the long-term memory -- the vector store of real customer expressions grounded in specific moments -- that introduced the texture of lived experience. The hesitations. The specificity. The “um.” You can architect the filter perfectly, but without experiential memory feeding through it, the output stays clean in a way that real humans never are.

If your AI persona can’t demonstrate both -- a specific interpretive filter and the situated texture of lived experience -- it’s not a twin. It’s a chatbot wearing a name tag.

Try it yourself. Pick a customer segment you know well and grab a blank page.

First, write three things about who they are -- their identity, their values, their decision logic. What would they never do? What do they always do? This is the personality layer.

Then write three things about how they’re interpreting right now -- their emotional state, their physical context, the external pressures shaping their decisions. This is your context vectors layer.

Now look at the two lists side by side. You’ve just built the skeleton of a personality lens. The next step is grounding it -- find the real customer language, the specific situations, the actual reasoning from your research data. That’s what turns a character into a twin.

The Paradox

We started this project trying to make the personas better. More polished. More complete. More authoritative.

The breakthrough came when we stopped.

Building AI customer personas taught me something I now apply to everything I build: fidelity beats perfection. The goal isn’t the best possible response. It’s the most faithful one -- faithful to how a real customer actually thinks, hesitates, reasons, and reacts.

The de-risking promise is real. You can simulate before you commit. Test before you spend. Get closer to your customer’s actual reaction before the budget is locked and the campaign is live. But only if you resist the instinct to make those simulations flawless. The moment you polish away the hesitation, the specificity, the messiness of real cognition, you lose the thing that made the simulation worth trusting.

Perfection is easy. Humanity is hard. It takes memory that grounds the persona in lived experience, interpretive filters that shape how it processes the world, and enough fidelity to real cognition that the output feels earned rather than generated.

The humanity in the machine wasn’t something we added. It was something we stopped removing.

I’ve been a designer for 22 years. These days, I design agents.

Not the kind that book your flights — the kind that generate marketing assets at scale. Brand governance agents that enforce visual identity across thousands of touchpoints. Innovation agents that can explore more ideas in an afternoon than any brainstorm ever could.

Last month, I watched one of our systems produce more creative output in a week than my early teams produced in a year.

Almost none of it was great.

Not bad, exactly. Just… adequate. On-brand. Technically correct. The kind of work that checks every box on the brief and leaves the audience completely unmoved.

We’ve scaled production. We haven’t scaled taste.

The Gap Nobody’s Talking About

The numbers tell a contradictory story: [88% of marketers now use AI in their daily workflow](https://www.surveymonkey.com/mp/ai-marketing-statistics/). But [only about a quarter are generating tangible value from it](https://www.bcg.com/publications/2025/are-you-generating-value-from-ai-the-widening-gap).

Unprecedented adoption. Underwhelming results. Both happening at once.

I see this up close. Companies rush to implement AI-powered creative workflows, celebrate the efficiency gains, then quietly wonder why everything still feels… generic. The tools got better. The output didn’t.

The missing variable isn’t technology. It’s judgment.

This isn’t just my observation. [Research from Harvard Business School](https://www.hbs.edu/bigs/artificial-intelligence-human-jugment-drives-innovation) found that AI assistance actually *widened* performance gaps rather than closing them — high performers got better while struggling performers declined. The difference? Judgment. As researcher Rembrand Koning put it: “For anybody who’s using AI in their work, you need to think carefully about the person who’s using the tool. Do they have enough judgment for tasks that are required?”

As AI commoditizes the *making* of things, the *choosing* becomes everything. What to make. How to make it land. When to break the rules. These decisions — the ones that separate forgettable from remarkable — require something we’ve always valued but rarely examined: taste.

Now we need to figure out how to scale it.

What Taste Actually Is

“Taste” gets thrown around loosely, so let me be specific.

Taste isn’t preference. Preference is “I like blue.” Taste is understanding why a particular blue works in this context, for this audience, at this moment — and why another blue would fall flat.

Philosophers have wrestled with this. Hume pointed out that beauty lives in the mind of the observer — but some observers are better than others. There are “true judges,” he wrote, with “strong sense, united to delicate sentiment.” Taste can be developed.

Here’s why this matters for creative work: when you say “this headline doesn’t land,” you’re not just stating a preference. You’re making a claim — as if it should be obvious. But the copywriter who wrote it feels the same confidence about their version. That tension between personal reaction and universal claim? It’s why creative feedback is so hard. Everyone speaks as if there’s a right answer while knowing it’s subjective.

In practice, taste is discernment. The ability to navigate endless possibilities and make decisions that add up to something coherent. A designer with taste orchestrates hundreds of small choices — typeface, spacing, color, tone — around a central vision. They know what to leave out as much as what to include.

Here’s the part that matters: taste is a skill, not a gift.

It develops through exposure — seeing lots of examples and figuring out why they work. Through articulation — forcing yourself to explain your reactions instead of just having them. Through practice — making choices, getting feedback, doing it again. Through context — understanding how creative work serves goals beyond itself.

After two decades of design work, I can confirm: my taste at year one was embarrassing. It improved because I studied, practiced, failed, and paid attention. The designers I admire most are still refining theirs.

If taste were innate, we’d be stuck. But since it’s learnable, we can teach it. We can build systems around it.

Which raises the question: how do we actually scale it?

The Floor and the Ceiling

Most conversations about AI and creativity get stuck in a false choice. [Industry analysts frame it as efficiency versus quality](https://www.m1-project.com/blog/ai-vs-human-creativity-in-marketing-finding-the-balance), automation versus artistry — as if you have to pick a side.

Here’s a better frame:

AI creates the floor. It ensures minimum quality, brand consistency, and operational efficiency. It handles adherence so humans don’t have to.

Humans create the ceiling. They make the judgment calls that turn compliant work into compelling work — the kind that actually moves people, earns attention in an ocean of content, makes the brand matter.

You need both. And by handing the floor to AI, you free humans to focus on what they’re actually good at: raising the ceiling.

But here’s the catch: freeing up capacity doesn’t automatically develop capability. Just because creatives have more time doesn’t mean they’ll develop better taste. That takes intentional work.

Raising the Floor: Rules That Run Themselves

Brand guidelines have existed for decades. But something shifts when they become machine-readable.

I’ve built brand governance agents that enforce visual identity in real-time. They don’t check compliance after the fact — they guide creation as it happens. Color within range? Typography following hierarchy? Tone matching brand voice? The agent knows, and it intervenes before mistakes ship.

This is the difference between brand *guidelines* and brand *guardrails*. Guidelines sit in a PDF nobody reads. Guardrails are embedded in the workflow.

The payoff: consistency at scale. Every asset, every touchpoint, every market — all aligned without an army of reviewers. This is what AI should be doing. It raises the floor reliably and tirelessly.

But rules have a ceiling. They capture past decisions, not future judgment. They tell you what’s on-brand. They don’t tell you what’s remarkable.

Raising the Ceiling: Understanding What’s Actually Going On

This is the human domain — and it’s harder to systematize.

Contextual awareness means understanding the ecosystem your work exists within. What are customers actually feeling? What’s the cultural moment? What references will land and what’s been said too many times already?

When I think about what elevates creative work from competent to compelling, it’s almost always this: the creator understood something about context that others missed.

Think about what customers are actually trying to accomplish. They don’t just have functional needs — they have emotional ones. They’re not buying a product; they’re hiring it to make progress in their lives. Understanding that emotional dimension — the feeling they’re seeking, the struggle they’re escaping — shapes creative choices at every level.

Or think about cultural timing. Brands that tap into the moment authentically create outsized resonance. But [according to Edelman’s Trust Barometer](https://www.edelman.com/trust/2025/trust-barometer/special-report-brands), 73% of consumers say their trust in a brand increases when it authentically reflects today’s culture — and they can tell when you’re faking it. Most attempts at cultural relevance backfire because they’re calculated rather than genuine.

This kind of awareness resists automation. Context shifts constantly. What worked last quarter feels stale now. You can’t encode “understand the cultural moment” into an agent prompt.

But you can build organizations that value it — and develop it.

Scaling Taste Through Culture

Here’s the uncomfortable truth: most organizations aren’t set up to develop taste. They’re set up to develop compliance.

Review processes catch errors. Brand guidelines enforce consistency. Approval workflows ensure nothing embarrassing ships. All of these raise the floor. None of them raise the ceiling.

If you want to scale taste, you need to treat it like any other organizational capability — something you invest in, teach, and reward.

Teach It

For individuals:

- Articulate your reactions. Don’t just say “this feels off.” Force yourself to explain why. Developing language for your reactions develops your discernment.

- Move beyond like and dislike. Ask: does this serve its goals? Does it work for its audience? Would different choices land better?

- Dissect work you admire. Don’t just consume it — pull it apart. What makes it work? What did they leave out?

- Practice deliberately. Make choices, get feedback, iterate. The designer who’s made a thousand decisions has better taste than one who’s made a hundred.

For teams:

- Create space for judgment calls, not just error-checking

- Discuss *why* something works, not just whether it’s on-brand

- Celebrate contextual awareness, not just compliance

- Make taste development part of growth conversations

Try this in your next creative review: Spend 5 minutes on “what’s wrong” and 15 minutes on “what would make this remarkable.” Most reviews invert this ratio — they’re dominated by error-catching and compliance-checking. Flip it. You’ll be surprised how the conversation changes.

Build Systems That Reward It

- Codify principles, not just rules. Rules tell you what. Principles tell you why. “Always include the logo” is a rule. “Our brand should feel confident but approachable” is a principle. Principles transfer across situations; rules just constrain.

- Create feedback loops. What’s performing? What’s resonating? Feed this back into the creative process — not just as data, but as learning.

- Invest in customer understanding. Not just analytics — genuine insight into what customers feel, want, and struggle with. This is the raw material of contextual awareness.

Hire and Promote for It

If you only reward speed and compliance, you’ll build an organization of fast rule-followers. If you want taste, you need to value it explicitly — in hiring, in performance reviews, in who gets promoted.

The Human Role, Evolved

In the AI-augmented creative workflow, human roles don’t disappear — they elevate. Here’s what that looks like in practice:

Taste-makers set direction. They’re not producing every asset, but they’re defining what good looks like. When I brief our agents, I’m not giving them a checklist — I’m articulating a vision. What should this feel like? What would make it remarkable rather than just acceptable? The agent can execute, but someone has to know what we’re aiming for.

Curators select from AI outputs. When an agent generates twenty options, someone has to choose. That choosing is where taste lives. It’s not passive — it requires understanding why option seven works and option twelve doesn’t, even when both pass the brand guidelines. The best curators can articulate their choices, which makes them better over time.

Teachers develop taste in others. This might be the most important role. The ceiling rises when more people can make good judgment calls, not just a few senior creatives. A creative director who hoards taste creates a bottleneck. One who teaches it builds an organization that scales.

Sense-makers interpret context. What’s happening in culture? What do customers actually feel right now? What’s the moment demanding? These questions require human judgment — and they require humans who’ve developed the capacity to answer them well. The sense-maker notices that last quarter’s winning approach now feels dated, and knows why.

None of these roles existed in their current form five years ago. All of them require taste. And all of them can be developed.

The Stakes

We’re at an inflection point. AI has handed us the ability to produce infinite content. But infinite content without taste is just noise.

The brands that matter in the next decade won’t be the ones with the most sophisticated AI. They’ll be the ones that use AI for brand adherence while building organizations that cultivate judgment.

The creative professionals who thrive won’t be the ones clinging to production work. They’ll be the ones who develop taste so sharp they become irreplaceable — and who help others develop it too.

The floor is rising automatically. The question is who will do the harder work of raising the ceiling.

Taste is the moat. Start building it.

*22 years of design, now building AI agents. If you’re figuring out how to scale judgment alongside scale itself, I’d like to hear what’s working

.