Tuesday, 9:14 a.m. Maya in support still owns the ticket — her name sits at the top of the thread — but the suggested reply is already waiting in the sidebar, polite and almost right. Two desks over, Devon pastes a stack trace into an assistant, gets a plausible explanation before the docs tab finishes loading, and mutters something that sounds halfway between thanks and suspicion. Down the hall, a product manager opens a half-drafted memo she asked for at 8:40, coffee still cooling, cursor blinking on a sentence she will rewrite three times before lunch.
Nobody cleared out overnight. The cubicles are full. Keyboards still click. What changed is quieter: the first pass arrives before the blank page does, and the job tilts toward review, taste, and knowing when the model is wrong.
That is the quiet version of the AI job story — not empty floors, just checklists rearranging while the unemployment rate still looks ordinary.
The shift you won’t see in the unemployment rate
The evidence we have so far points to task reorganization inside jobs more than wholesale disappearance of occupations. Tools are spreading quickly. Careful studies of wages and hours are still fairly muted. Productivity gains show up clearly on specific tasks. And some roles are quietly becoming more about review, taste, accountability, and knowing when the model is wrong.
That framing matters if you’re skeptical of the louder headlines. Mega-forecasts of white-collar wipeout often jump from “the model can do X” to “the payroll line for X vanishes,” and they skip how firms actually roll tools out. It’s wiser to lean on field experiments, linked survey-and-admin work, and Bureau of Labor Statistics projections that assume gradual adoption than on LinkedIn scare posts. The sections below stick to what has been measured — and flag what has not.
What AI already takes off the checklist
Start with production work, not demos. In a large staggered rollout published in the Quarterly Journal of Economics (2025), Brynjolfsson, Li, and Raymond found generative AI assistants for customer-support agents raised issues resolved per hour by about 14–15% on average. Gains were largest for less experienced agents. Top performers saw smaller or mixed effects. That is a real firm, thousands of agents, and a concrete task — not proof that headcount fell one-for-one with productivity.
Writing and inbox work look similar once you zoom into the task layer. A 2025 field experiment across 66 firms and 7,137 knowledge workers randomly assigned Microsoft 365 Copilot: among the roughly 80% of treated workers who used the tool, Outlook email time fell about two hours per week (around 17%), with less after-hours digital work — but little detectable change in meeting volume, document counts, or broader task mix (Dillon, Jaffe, Immorlica & Stanton; also NBER w33795). That is quieter than the older lab result where ChatGPT cut time about 40% and raised quality about 18% on occupation-specific writing tasks (Noy & Zhang, Science 2023) — a useful early signal, but one that lacked hard fact-checking and firm-specific context. In the wild, the first draft gets cheaper; the send button does not.
Coding assist is also quieter when you look at whole-job math instead of timed lab puzzles. Across three enterprise field experiments covering 4,867 developers at Microsoft, Accenture, and a Fortune 100 firm, pooled estimates associate Copilot use with about a 26% rise in weekly completed tasks (pull requests), with larger gains for less experienced developers (Cui, Demirer, Jaffe, Musolff, Peng & Salz, Feb 2025). That is a task speedup inside real workflows — not a claim that software jobs evaporate. BLS still projects software developer employment up 15.8% for 2024–34, arguing that demand and AI buildout may outweigh productivity drag on headcount (BLS Economics Daily, July 2026; MLR overview, Jan 2026).
Other checklist items show up as streamlining more than replacement: call summaries and sales notes, first-pass graphic and technical drafts, legal document sift, and research assist. Paralegals and legal assistants face a tighter BLS employment path than lawyers in those projections, partly because review capacity can stretch further — while lawyers still own advice and still have to catch errors. A 2025 preregistered evaluation of leading legal research AIs found Lexis+ AI and Thomson Reuters tools still hallucinated on roughly 17–33% of queries — better than raw GPT-4, nowhere near “hallucination-free” (Stanford HAI / Magesh et al.; Journal of Empirical Legal Studies, accepted 2025). Human review is not optional in that domain.
Usage data backs a task map more than a job apocalypse. Anthropic’s Economic Index has tracked Claude conversations mapped to O*NET tasks since early 2025. By the March 2026 update, coding still dominated, but use cases on Claude.ai had diversified; about 49% of jobs had seen at least a quarter of associated tasks performed with Claude, and augmentation (learn, iterate, validate) edged up relative to earlier snapshots (Anthropic Economic Index, Mar 2026). The authors still caveat hobby-versus-work blur and product-mix skew — and it’s still one of the better revealed-preference maps we have.
Put simply, here is what AI already eats versus what stays human:
- Suggested replies and support macros — the model drafts; humans escalate, judge tone, and grant policy exceptions.
- First drafts of writing — the model compresses blank-page time; humans own facts, voice, and the send button.
- Boilerplate code, stubs, and “explain this” — the model accelerates; tech leads still own architecture, security, and merge risk.
- Summaries, notes, contract sift, design variants — the model narrows the pile; humans still sign advice, brand fit, and filings.
Where humans got more expensive (in the good sense)
If AI is swallowing first drafts, why does judgment feel more valuable? Because someone still has to know which side of capability you are on — and put a name on the outcome.
Danish field evidence captures the quiet pattern especially well. Humlum and Vestergaard surveyed about 25,000 workers in eleven highly exposed occupations and linked those answers to administrative records through December 2024. Adoption was rapid. Difference-in-differences estimates show no clear differential effect on earnings or hours, ruling out effects larger than about 2% two years after ChatGPT. New AI-related work split roughly into content generation (~42%), reviewing AI outputs and compliance (~35%), and integrating AI into workflows — policies, fine-tuning, embedding tools (~26%). Most employers encouraged or allowed use; few banned it (NBER w33777; RFBerlin summary). Still waters on paychecks. Rapid currents in the task mix.
Knowledge work adds a sharper edge: the “jagged frontier.” In a BCG field experiment led by Dell’Acqua and coauthors — now published in Organization Science (2026) — GPT-4 helped consultants on tasks inside AI capability — more completed, faster, higher quality — but hurt performance on a task designed outside that frontier (about 19 percentage points fewer correct recommendations). Prompt guidance did not fix the miss (HBS / Organization Science PDF). Judgment, in practice, is simply knowing which side of the jagged line you’re standing on.
That is why some roles got pickier about humans — not softer:
- Lawyers and senior counsel — models sift and draft first-pass research; humans own advice, filings, and duty of competence when hallucinations show up.
- Editors and publishers — models spit headline variants and draft copy; editorial responsibility for accuracy, bias, disclosure, and the publish call cannot be outsourced to a model.
- Software tech leads — faster pull requests are not correct architecture; merge accountability, security, and risk still sit with people.
- Compliance, audit, and risk — models draft memos and flag anomalies; regulators still expect human oversight, sign-off, and residual-risk ownership.
- Teachers and trainers — models outline lessons and quizzes; integrity policy, pedagogy, and assessment design stayed human — and grew noisier.
The punchline is boring on purpose: someone still signs the brief, the merge, the publish button, and the client advice. Accountability didn’t get automated. It got louder sitting next to faster drafts.
How to stay useful without pretending you’re a chatbot
None of this requires becoming a prompt personality. It asks for calmer competence around the handoff — knowing what to trust, what to rewrite, and what to own.
- Own the handoff. Treat AI as a first-draft, search, and macro machine. Your job is to verify, adapt to context, and sign your name. The Danish review-and-compliance bucket, the legal hallucination literature, and the jagged frontier all point there.
- Map your tasks, not your job title. Which 20–40% of your week is draftable? Which parts require stakeholders, taste, or liability? Anthropic’s task diffusion is a better lens than job-title panic.
- Keep an “AI fails here” list. Numbers that must be sourced. Novel strategy. Regulated speech. Customer emotion. Security. Dell’Acqua shows people over-trust outside the frontier.
- Build integration skills. House style in prompts, eval checklists, logging when AI was used, light policies — the people who get asked to embed tools are doing the Danish “integrate” work.
- Keep craft AI compresses but does not replace. Domain depth — specialty nuance, local market, codebase history, brand voice — is what makes review fast and correct. Studies reward complementary skill plus tools, not tools alone.
- Document accept/reject decisions. A short note on why you kept or killed a model output becomes institutional memory and proof of accountability when something goes wrong.
- Watch entry-level pathways. Separate discussions flag early-career risk in exposed occupations; Danish chatbot-adoption firms are not clearly driving wage/hours damage yet. Seek roles that still train judgment on real stakes, not only AI babysitting.
- Measure outcomes bosses care about. Resolved tickets with low reopen rates. Features shipped without incidents. Drafts that need fewer edit cycles. Clients who stay. Tokens generated are not a résumé.
What the numbers don’t say yet
BLS employment projections for 2024–34 already bake in AI-related drag for some exposed roles — customer service representatives −5.5%, medical transcriptionists −4.9%, claims adjusters −5.1%, office and administrative support as a group −3.9% — while software developers still grow strongly (+15.8%). Those are projections, not observed mass layoffs (BLS Economics Daily, July 2026; MLR, Jan 2026).
Across the OECD, roughly one-quarter of workers have been classified as “exposed” to generative AI — meaning a sizable slice of tasks looked amenable to assistance. Exposure isn’t the same thing as displacement. Net employment impact mixes automation, new tasks, and productivity complementarity (OECD, Skills in the AI age, 2026).
Honest gaps remain. U.S. linked employer–employee evidence on wages and hours is thinner than the Danish work. Public data on headcount inside firms that rolled out support or coding AI is still weak — and productivity is not the same as layoffs. Early-career and apprenticeship effects are still contested. Enterprise API usage may look different from consumer chatbot mixes.
So the quiet story is reorganization. AI got loud on drafts. Humans stayed expensive where judgment, taste, and a signature still matter. Stay useful by owning that handoff — not by competing to sound like the chatbot that already took the first pass.