事实正是如此。斯坦福数字经济实验室发现，在高度暴露于 AI 的职业中，22至25岁劳动者的就业水平，比按照低暴露同行的轨迹本应达到的水平低约19%；这个差距从2025年7月的15%扩大了。就绝对值而言，两个暴露度最高五分位中的这一年龄组就业率，在2022年11月至2026年6月间下降约11%，而暴露度最低的三个五分位增长约10%。调整通过减少招聘发生，而不是通过裁员。10
细分结果才是关键：
就业下降集中在 AI 使用倾向于自动化人类任务的职业。在 AI 更多用于增强劳动者的职业中，就业持平或上升，尤其是对经验更丰富的劳动者而言。10
2025年7月，METR 做了一项随机对照试验，让16名有经验的开源开发者在熟悉的仓库里完成246项任务。允许使用 AI 工具时，他们反而慢了19%。事前他们预测会快24%，事后仍相信自己快了20%。这项结果在一年间成了反驳生产率主张时被引用最多的证据，而且理应如此。11
METR 在2026年2月重新审视了它。2025年末的后续研究估计，回访参与者慢了18%，新招募者慢了4%，两组置信区间都跨过零；研究者得出的结论是：严重的选择偏差——最热情使用 AI 的开发者系统性地拒绝参加——意味着他们的数字“很可能代表真实生产率影响的下界”。他们现在直白地说：“我们认为，与2025年初的估计相比，开发者在2026年初很可能已经从 AI 工具中获得了更大的加速。”他们正在重新设计实验。12
另外两个发现仍然完整存在，并且它们不是削弱而是强化了这个论点。一项让128名知识工作者参加的随机、实验室嵌入实地实验发现，生成式 AI 一贯提升速度，但对质量的影响取决于任务：知识打包和知识创造的质量提高，知识获取的质量下降。13 Microsoft 365 Copilot 对56家公司、6,000多名员工进行的为期6个月的随机试验则发现，常规用户每周少花3.6小时处理邮件——减少31%——文档完成速度提高12%，而花在会议上的时间没有显著变化。14
Three years of asking which professions AI would replace produced the wrong map. The 2026 evidence points somewhere else entirely — at the boundary between professions, and at how few people it now takes to cross it.
The standard question about AI and work has been stable since 2023: which occupations get automated, and in what order. It produced a genre — the exposure study, the percentage-of-tasks estimate, the ranked list of doomed professions — and the genre has been useful. It has also been asking about the wrong unit.
The question worth asking in 2026 is narrower and stranger: what happens when one professional can cheaply command the software-building, research, data-processing, and verification labor that used to require a team, a department, or a vendor?
That is not a question about job loss. It is a question about organizational form. And the evidence that arrived over the past nine months — from OpenAI's own usage telemetry, from labor-market microdata, from engineering measurement platforms watching millions of pull requests — converges on an answer with more force than the occupation-ranking literature ever managed.
Start with what changed in how people use these systems, because the shift is documented with unusual precision.
In December 2025, 35.4% of active individual Codex users sent at least one prompt that would have taken an experienced human at least an hour to complete. By May 2026, that share was 70.2%. Over the same five months, the share sending at least one prompt representing eight hours of human work went from 2.1% to 25.6% — a twelvefold increase in the appetite for delegating a full working day.1
The internal numbers are more extreme. Between November 2025 and June 2026, the median OpenAI worker's monthly output tokens rose at least tenfold in every job function. The median employee in a legal role generated thirteen times more; the median researcher, more than fifty times. By June 2026, Codex accounted for 99.8% of the output tokens OpenAI workers generated across Codex and ChatGPT combined — a near-total migration from conversation to delegation inside a single company in seven months.1
Two smaller figures in the same paper matter more than the headline ones. More than 10% of users manage three or more concurrent agents at some point in a given week. And 26.6% use skills — reusable instruction sets that encode a workflow so it can be run again.1
Those two numbers describe something other than faster work. Running three agents in parallel is a supervision problem, not a typing problem. Writing a skill is not performing a task; it is building a machine that performs the task, and leaving it installed.
The progression has four rungs, and the telemetry shows where the population currently sits on it:
flowchart TD
T["<b>1 · Task</b><br/>Summarize this contract<br/><i>Human owns the whole workflow</i>"]
W["<b>2 · Workflow</b><br/>Compare 40 contracts,<br/>produce a spreadsheet<br/><i>A bounded job is delegated</i><br/>≥ 1-hour prompts<br/>35.4% → 70.2%"]
P["<b>3 · Process</b><br/>Every inbound contract:<br/>extract, compare, flag, file<br/><i>The procedure is delegated</i><br/>≥ 8-hour prompts<br/>2.1% → 25.6%"]
S["<b>4 · System</b><br/>Clause database, deviation<br/>detection, self-checks<br/><i>The professional owns machinery</i><br/>26.6% write skills"]
T --> W --> P --> S
Rung four is the one that matters. A professional standing on it is no longer using AI to do the work — they are manufacturing the apparatus that does the work, and the apparatus persists after today's question has been answered.
The boundary between occupations is dissolving first#
The more consequential finding landed in July 2026. OpenAI's economic research team classified a random sample of work-related messages from U.S. business users against O\*NET, the Labor Department's occupational database, sorting each one as generic, within-occupation, or cross-occupation.2
Strip out the generic work — writing emails, scheduling meetings, the connective tissue every job shares — and 43.5% of what remains concerns tasks historically belonging to a different profession. Not adjacent work. Another occupation's work.
The distribution across the eight occupation groups studied is where the argument lives:
Occupation group
Cross-occupation share of occupation-specific AI use
Customer experience
77%
Design
75%
Human resources
69%
Legal
56%
Marketing
53%
Sales
40%
Finance
40%
Engineering
28%
Engineering sits at the bottom. Engineers import the least outside work of any group measured — while engineering tasks travel more widely into everyone else's usage than almost anything else. Troubleshooting computer applications and systems ranks among the top three engineering-related tasks in every single non-engineering occupation group.2
That asymmetry is the whole thesis compressed into one table. The traffic runs in one direction: out of software, into everything else. Engineers already had the capability that agentic tools confer, so they have the least to import. Everyone else is acquiring it now, and using it to reach across a boundary that used to require hiring someone.
Why programmers are the leading indicator, not the largest beneficiary#
The instinct is to assume programmers capture the biggest gain, since coding agents are the most capable tools on the market. The instinct is backwards, and the reason is arithmetic.
Programmers already possessed software-production capability. Doubling or quintupling it is a large improvement on a base that was never zero. Consider instead the biologist, the litigator, the schoolteacher, the small-business owner — people for whom the practical capacity to build working software was, for all economic purposes, zero. Moving from zero to a functioning research pipeline is not a productivity multiplier. It is a change in category.
Anthropic's measurement work quantifies exactly this headroom. Its "observed exposure" metric compares what models can theoretically do in an occupation against what people actually use them for. In computer and mathematical occupations, theoretical capability sits at 94% while observed exposure reaches 33%. Office and administrative support shows 90% theoretical against a substantially lower observed share. Roughly 30% of workers — cooks, mechanics, lifeguards, dishwashers — have zero coverage at all.3
Microsoft's parallel study, built on 200,000 anonymized Copilot conversations and scoring 785 occupations, found the highest applicability in knowledge work that involves gathering, synthesizing, and communicating information: translators, historians, writers. Its authors are explicit that applicability is not displacement, and that the score should not be read as a forecast of job elimination.4
Both findings point the same way. The gap between what these systems can already do in a profession and what that profession currently asks of them is enormous, and it is widest outside software. The binding constraint has stopped being model capability. It is the professional's ability to specify, wire together, and check the work.
The Coasean turn: AI cuts the cost between occupations#
Ronald Coase asked in 1937 why firms exist at all, when markets could in principle coordinate every transaction. His answer was transaction costs: searching for a counterparty, negotiating terms, writing contracts, monitoring performance. When those costs are lower inside an organization than across a market, the organization absorbs the activity. The boundary of the firm sits wherever the two costs balance.5
Searching, summarizing, drafting, structuring handoffs, and checking work against rules are precisely the activities agentic systems perform at near-zero marginal cost. That does not merely make firms more efficient. It moves the boundary.
The historical pattern was a chain of internal transactions, each link a place where cost and delay accumulated. The emerging pattern removes the intermediate links entirely:
The claim is not that AI automates occupations. It is that AI collapses the cost of moving work across the seams between them — and the seams are where organizations put their people.
The usage data already shows the fingerprint. Task crossover is more common in smaller workplaces: 18.9% in workspaces of two to five seats against 16.3% in workspaces of a hundred or more, among users in the middle half by message volume.2 Where there are fewer colleagues to hand work to, the tool absorbs the handoff. That is the transaction-cost mechanism, visible in telemetry.
The founding data moves in step. The share of new startups on Carta with a single founder rose from 23.7% in 2019 to roughly 36% in 2025, up from 31% the year before — a doubling across a decade, with the sharpest move at the end.6
One qualification keeps this honest, and it cuts against the triumphant version of this argument. A shrinking minimum viable organization is not a promise of median success, and the two get conflated constantly — usually by people selling something. The United States had 30.4 million nonemployer businesses in 2023, taking in roughly $1.8 trillion between them: an average near $59,000 per firm, and that figure is revenue, before tax, before healthcare, before every other cost of operating without a payroll.7 Whatever the tools now permit, this is what the population of one-person businesses actually looks like.
Both propositions hold at once, because they describe different quantities. The floor is falling; the average is not. The number of people required to produce a given output has dropped, and that stays true whether or not any particular person turns out to be the one who clears the bar.
Generation got cheap; verification became the constraint#
The most valuable evidence for this thesis comes from the place it looks worst.
Faros AI compared engineering outcomes at low versus high AI adoption inside the same organizations, across two years of telemetry from 22,000 developers on more than 4,000 teams. Throughput rose exactly as promised. Then came the other column — and the shape of the whole finding is a queue moving downstream, not a gain evaporating:8
flowchart TD
GEN["<b>Generation — got cheap</b><br/>Tasks per developer +33.7%<br/>Epics per developer +66.2%<br/>PR merge rate +16.2%"]
VER["<b>Verification — the constraint</b><br/>Review time +441.5%<br/>First review +156.6%<br/>Unreviewed merges +31%"]
OUT["<b>What reaches production</b><br/>Bugs per developer +54%<br/>Incidents per PR about 3x<br/>Code churn about 10x<br/>Deployments −11.7%"]
GEN -->|"the queue relocates"| VER
VER -->|"what the queue lets through"| OUT
The review figures are medians, and code churn is the share of code rewritten within two weeks of being written — work that was produced, reviewed or waved through, and then thrown away.
GitClear's independent measurement of commit history finds the same erosion in the artifact rather than the process: refactoring fell from 25% of changed lines in 2021 to under 10% in 2024, cloned code rose, and the share of changes that revisit code older than a year has dropped 74% since 2023.9
Read as a verdict on AI coding tools, this is damning. Read correctly, it is the thesis confirming itself on schedule.
Production became cheap. Nothing else did. The queue did not disappear — it relocated, from the keyboard to the reviewer, and everything downstream of generation is now the constraint. A team that can produce fifty thousand lines overnight has not solved its hardest problem; it has purchased a much larger version of it. The scarce resource moved from making the artifact to specification, evaluation, verification, and accountability — and the organizations absorbing the productivity gains are the ones that treated review as infrastructure rather than etiquette.
This generalizes past software, and the generalization is the point. A lawyer can now generate two hundred arguments, a scientist five hundred hypotheses, an investor a thousand signals. In every case the same four questions decide whether any value survives contact with reality: What should be asked? What counts as good? What evidence would change the answer? Who is accountable when it is wrong?
The labor data says the same thing in a colder voice#
If the constraint has moved to specification and judgment, the labor market should reward exactly that — and punish work that is codified enough to hand over whole.
It does. Stanford's Digital Economy Lab finds employment among workers aged 22 to 25 in highly AI-exposed occupations running about 19% below where it would be had it tracked less-exposed peers, a gap that widened from 15% in July 2025. In absolute terms, employment for that age group in the two most exposed quintiles fell about 11% between November 2022 and June 2026, while the three least-exposed quintiles grew about 10%. The adjustment runs through reduced hiring, not layoffs.10
The disaggregation is what matters:
Declines are concentrated in occupations where AI usage tends to automate human tasks. In occupations where AI is used more to complement workers, employment is flat or rising, particularly among more experienced workers.10
And the sharpest cut of all: employment has fallen among young workers in occupations that rely on codified knowledge, and risen among experienced workers in occupations that rely on tacit knowledge built through practice and mentorship.10
Anthropic's analysis of the same period finds no systematic unemployment increase among highly exposed workers since late 2022 — the difference-in-differences estimate is indistinguishable from zero — alongside the same suggestive slowdown in hiring for workers aged 22 to 25.3
Put together: this is not an extinction event for professions. It is a repricing within them. Codified execution is losing value; specification, judgment, and verified accountability are gaining it. Which is precisely what a world running on cheap generation and expensive verification would look like.
Intellectual honesty requires naming the strongest counterevidence, and then reporting what happened to it.
In July 2025, METR ran a randomized controlled trial in which sixteen experienced open-source developers completed 246 tasks in repositories they knew well. Allowed AI tools, they were 19% slower. They had predicted a 24% speedup beforehand and believed they had gotten a 20% speedup afterward. The result was the most-cited rebuttal to productivity claims for a year, and deservedly so.11
METR revisited it in February 2026. Its late-2025 follow-up produced an estimated 18% slowdown among returning participants and 4% among new recruits, with confidence intervals crossing zero in both cases — and the researchers concluded that severe selection bias, with the most AI-enthusiastic developers systematically declining to participate, meant their figures "likely represent a lower-bound on the true productivity effects." They now say plainly: "we believe it is likely that developers are more sped up from AI tools now — in early 2026 — compared to our estimates from early 2025." They are redesigning the experiment.12
The strongest objection has been substantially withdrawn by the people who raised it. That is how good empirical work behaves, and it is worth more than a hundred confident predictions.
Two other findings survive intact and sharpen rather than undermine the argument. A randomized lab-in-the-field experiment with 128 knowledge workers found generative AI consistently improved speed while its effect on quality was task-contingent: quality rose for knowledge packaging and knowledge creation, and fell for knowledge acquisition.13 And the six-month randomized trial of Microsoft 365 Copilot across more than 6,000 workers at 56 firms found regular users spending 3.6 fewer hours per week on email — a 31% reduction — and completing documents 12% faster, with no significant change in time spent in meetings.14
That last null result is the most instructive number in this entire body of research. Email is individually controllable; meetings require other humans to agree. AI is dissolving computational friction — research, drafting, calculation, transformation, monitoring, testing — at a rate it cannot touch institutional friction: trust, authority, negotiation, accountability, permission, liability. Professions dominated by the first will compress. Professions dominated by the second will change on a schedule set by human institutions, not by model releases.
Traditional productivity asks how much work one person can perform. The question that fits the evidence asks how much reliable autonomous work one person can organize, specify, and verify.
That reframing changes what is worth measuring. Occupation-level exposure scores describe a world in which the job is the atom, and the OpenAI crossover data shows the atom splitting. The more informative indicator is organizational: the smallest number of people required to produce a given professional output. Track that per field, and the transformation becomes legible in a way that unemployment statistics — currently showing very little — cannot yet capture.
The professional of 2026 is not best described by the tasks personally performed. The description that fits is the system of human and machine labor designed, commanded, and held accountable — and the evidence says the people winning are not the ones producing the most, but the ones who can still tell whether any of it is right.
Building got cheap. Knowing got expensive. The distance between those two facts is where the next decade of professional work will be decided.
All figures current as of August 2026. This is a fast-moving evidence base — the METR reversal documented above is a reminder that the most-cited number in a field can be withdrawn by its own authors within eight months.
Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe, and Aaron Chatterji, The Shift to Agentic AI: Evidence from Codex, OpenAI, 25 June 2026 (arXiv:2606.26959). Usage figures derive from a 0.1% sample of individual Codex users who opted in to training-data sharing; internal OpenAI figures are self-reported by the company and should be read as such. ↩︎↩︎2↩︎3
Caroline Chin and Alex Martin Richmond, Work at the Frontier: How AI is expanding what people do at work, OpenAI, July 2026. Of all work-related messages, 61.5% are generic, 21.8% within-occupation, and 16.8% cross-occupation; rebasing to exclude generic work yields the 43.5% figure. The workspace-size gradient is reported for users in the middle 50% by message volume and is noted by the authors as non-monotonic among top-quartile users. ↩︎↩︎2↩︎3
U.S. Census Bureau, Nonemployer Statistics, 2023 reference year: 30,427,808 nonemployer establishments with combined receipts near $1.8 trillion, an average of roughly $59,200 each. Nonemployer businesses are those with no paid employees and receipts of $1,000 or more. Receipts are revenue, not profit, and the average conceals a highly skewed distribution. ↩︎
Faros AI, The AI Engineering Report 2026: The Acceleration Whiplash. Telemetry from 22,000 developers across more than 4,000 teams, comparing each organization's lowest- and highest-AI-adoption periods. Vendor research: Faros sells engineering-analytics tooling, and the framing serves that business. The measurements are drawn from workflow telemetry rather than surveys, which is the reason to take them seriously despite the source. ↩︎
Shifting Work Patterns with Generative AI, NBER Working Paper 33795. Six-month randomized field experiment, 6,000+ workers across 56 firms; roughly 40% of those granted access became regular users. ↩︎
