Jobs AI will replace

How this ranking works

Every occupation here is ranked primarily by the Anthropic Economic Index's observed exposure score — what share of an occupation's tasks actually show up in real Claude.ai conversations, weighted by how completely the model handles them. That is a measure of current behaviour, not a theoretical ceiling, which makes it the most defensible single signal for a page titled “jobs AI will replace.” Where it is useful, we cross-reference Eloundou et al.'s theoretical task-exposure score, which measures what a model could do rather than what people are doing with it.

Occupations are grouped by their Standard Occupational Classification major group, then ranked within each group by observed exposure. Every figure is pulled directly from the same dataset behind our searchable exposure tool, which covers all 245 occupations against all four major studies side by side. Nothing on this page is averaged across studies or interpolated to fill a gap — where a number is not available, it says so.

The 15 most exposed occupations

Computer Programmers and Customer Service Representatives sit at the top of the observed-exposure ranking, both above 70%. From there the list is dominated by two kinds of work: structured office and clerical tasks, and technical-analytical roles whose deliverable is text, code or a written recommendation.

Cluster 1: Office and administrative support — the biggest by headcount

This is the largest exposed cluster in the dataset by raw worker count: 10 of the 21 office and administrative support occupations clear 20% observed exposure, covering 10.7 million of the group's 14.6 million US workers. It is also the most automation-skewed cluster — when these roles show up in Claude usage at all, the model is usually doing the task rather than assisting with it.

The task lists explain why. Data Entry Keyers' single dominant task — “read source documents… and enter data in specific data fields” — scores a full 1.0 on Anthropic's task-penetration scale. Office Clerks' top tasks, scheduling and filing, score the same. These are close to the textbook case for a language model: unambiguous inputs, a defined output format, no physical component. Secretaries (70.6% automation), Medical Secretaries (72.0%) and Office Clerks (69.3%) are among the most automation-skewed occupations in the entire 245-occupation dataset.

The counter-evidence sits in the same rows. Usage share for most of this cluster is still small — Data Entry Keyers at 0.08% of all Claude conversations, Billing Clerks at 0.04% — which means the automation split for these occupations rests on a thin sample and should be read as indicative rather than precise. And the least-exposed tasks within the same jobs, like Office Clerks' “collect, count, and disburse money,” score zero: these roles are bundles of tasks, and the bundle is being narrowed, not deleted wholesale.

Cluster 2: Computer and mathematical occupations — smaller, but almost entirely exposed

Thirteen of the fourteen computer and mathematical occupations in the dataset clear 20% observed exposure — 4.18 of 4.20 million workers, effectively the entire category. It is a smaller cluster by headcount than office work, but it holds the single highest exposure score of any occupation: Computer Programmers at 74.5%.

This cluster is the strongest evidence that exposure has moved up the income scale. Computer Programmers earn a $100,390 median wage; Information Security Analysts, $129,180. Neither pay level offers protection when the core deliverable is written analysis or code — Programmers' least-exposed task is “assign, coordinate, and review work…of programming personnel,” the management layer of the job, not the coding itself.

A useful caution sits inside the same SOC group: Software Developers, a broader occupation that folds in requirements analysis and architecture, scores just 28.8% observed exposure — less than half of Computer Programmers' 74.5%, despite overlapping job titles in everyday use. How narrowly a role is defined changes the number enormously; treat any single-digit-precision exposure score as a description of a specific task bundle, not of everyone who might use that job title.

Cluster 3: Sales and information-delivery roles

Twelve of sixteen sales occupations clear 20% exposure, 8.5 of 13.1 million workers. Customer Service Representatives post the second-highest exposure of any occupation in the dataset at 70.1%, and Sales Representatives of wholesale and manufacturing products carry the single highest usage share of any occupation measured: 1.39% of all Claude conversations.

The pattern is dialogue that can be scripted. Customer Service Representatives' top tasks — “solicit sales of new or additional services” and “refer unresolved customer grievances to designated departments” — both score a full 1.0. Product knowledge lookup, objection handling and CRM logging are language tasks even when the job is nominally about a relationship.

Retail Salespersons make the counter-case inside the same major group: only 32.2% exposure despite superficially similar work. “Ticket, arrange, and display merchandise” scores full exposure, but “bag or package purchases and wrap gifts” scores zero. Floor retail is a bundle of physical and verbal tasks; phone and wholesale sales are closer to pure text. The medium the job is conducted in, not the job title, is doing most of the work here.

Cluster 4: Business, financial and HR analysis

Seven of fifteen business and financial operations occupations clear 20% exposure — 5.25 of 7.62 million workers — anchored by Market Research Analysts and Marketing Specialists at 64.8%, the third-highest exposure score in the whole dataset.

These occupations' output is a written recommendation, and the highest-scoring tasks read that way directly: Financial and Investment Analysts' “recommend investments and investment timing” and “evaluate and compare the relative quality of various securities” both score 1.0. Exposure reaching into Financial Managers (39.1%) and Marketing Managers (32.0%) — both six-figure median-wage roles — shows this is not confined to entry-level analyst work.

Management Analysts — management consultants, in practice — sit lower at 24.3% despite doing superficially similar work to the analysts above them. Their least-exposed task, “confer with personnel concerned to ensure successful functioning of newly implemented systems,” is interpersonal follow-through across an ambiguous, client-specific problem — a different shape of work than producing a standardised analysis against a defined framework.

How exposure is distributed across occupational groups

Zooming out from individual occupations to entire SOC groups makes the shape of this clearer: exposure is concentrated almost entirely in desk work. Groups built on physical presence — healthcare practitioners, transportation, construction, food preparation, installation and repair — show zero occupations above the 20% exposure threshold.

Notable exceptions: where exposure is high but automation isn't

A few occupations don't fit neatly into the clusters above, and they are instructive precisely because they break the pattern. Interpreters and Translators post 43.0% observed exposure and one of the highest theoretical task-exposure scores in the dataset (88% under Eloundou's model) — translation is close to a pure language task. But Writers and Authors, Editors and Graphic Designers, despite very high usage share (2.2%, 2.6% and 1.1% of all Claude conversations respectively — among the highest in the dataset), show comparatively modest observed exposure (24.6%, 24.6%, 36.7%) and augmentation-skewed splits (71.5%, 66.4% and 70.4% augmentation). People in creative and editorial work use AI constantly — they just mostly use it as a collaborator rather than a replacement, which is the opposite pattern from the office-clerical cluster above.

Where automation already outweighs augmentation

Exposure tells you where a model could help. The automation-to-augmentation split tells you something sharper: among occupations with at least a thin real usage sample (0.1% of conversations or more), these are the roles where, when AI shows up, it is already doing the task end to end rather than assisting a person.

That footnote matters more than it looks. Correctional Officers and Jailers shows 0% observed task exposure but a 97.5% automation share on the handful of conversations Anthropic's classifier mapped to that occupation. The honest read is not “AI is automating prison work” — it is that a tiny, noisy sample produced an extreme ratio. The same caution applies, at a smaller scale, to any occupation in this dataset with a usage share under roughly 0.1%.

Exposure is not automation, and automation is not job loss

Every ranking on this page measures exposure or observed AI usage. Neither is the same as a job disappearing, and conflating them is how a research finding turns into a headline.

Anthropic's sixth Economic Index report (June 2026) found real Claude.ai conversations in May 2026 split 51.4% augmentation to 48.6% automation — a narrow majority still has a person in the loop, and that balance has drifted toward augmentation, not away from it, since Anthropic began tracking it. Its separate labour-market analysis (March 2026) found no systematic rise in unemployment among highly exposed workers, with the clearest concerning signal being slower hiring of younger workers in some exposed occupations — a warning about entry points into a career, not about existing jobs vanishing. For what has actually been announced and attributed to AI, rather than forecast, see our AI job displacement tracker.

What automation has done before

The clearest historical case study is the bank teller. Boston University economist James Bessen's research found that ATMs cut the number of tellers needed per branch from 21 to about 13 — and full-time-equivalent teller employment in the US still rose, from roughly 500,000 in 1980 to about 550,000 in 2010. Cheaper branches meant more of them; urban bank branches grew 43% over the same period, and tellers shifted from cash handling to relationship and sales work. The task was automated. The occupation was not eliminated — it was redefined around what was left, and it took online banking, a different technology entirely, to eventually shrink it.

At the economy-wide level, research by MIT's David Autor and colleagues on the origins of new work found that most US employment in 2018 sat in job titles that did not exist in 1940, and that technologies which complement human output have historically generated more new task categories than automating technologies eliminated. The same paper is not an unqualified all-clear, though: it also found that the source of new-work creation shifted after 1980, from middle-paid clerical and production jobs toward high-paid professional roles and low-paid services — squeezing the middle of the distribution even while the aggregate number of jobs grew. That is a strikingly similar shape to the office-clerical exposure pattern on this page.

The strongest case against this page's own framing

MIT economist Daron Acemoglu's “The Simple Macroeconomics of AI” (NBER, 2024) takes task-exposure estimates like the ones ranking this page at face value and runs them through a standard growth-accounting model. His result: even generous assumptions about AI's task-level capability imply no more than a 0.66% increase in US total factor productivity over ten years — a small fraction of the double-digit GDP claims sometimes hung on exposure statistics. His argument is that exposure counts overstate real economic disruption for three reasons: many exposed tasks are a small share of the total value an occupation produces, cost-effective deployment lags lab-demonstrated capability by years, and human oversight persists inside workflows that get labelled “automated.” If Acemoglu is right, the ranking on this page is closer to a map of where AI deployment is economically attractive than a countdown to when those jobs disappear.

What the macro forecasts add

The occupation-level rankings above are the only part of this page that can say anything about a specific job. The widely quoted economy-wide numbers are useful only as context, and none of them counts jobs that have actually gone.

What this means if you're in an exposed cluster

• Look at your own task list, not your job title. Every occupation in every cluster above keeps a set of least-exposed tasks scoring at or near zero — that bundle is where your leverage sits, and it is usually judgement, accountability or in-person presence.
• Automation-skewed is the sharper warning than high exposure. A high score with an augmentation-heavy split (like Writers and Graphic Designers) describes a changed way of working. A high score with an automation-skewed split (like Secretaries and Computer Systems Analysts) describes work already being delegated end to end.
• Watch usage share, not just exposure. An occupation with a fast-growing share of real conversations and a rising automation percentage is the pattern to take seriously; a static or shrinking share suggests the exposure score is still mostly theoretical.
• Treat granularity as information. Computer Programmers (74.5%) and Software Developers (28.8%) sit in the same broad field with wildly different scores because the SOC system splits narrow, well-specified coding from broader systems work — a reminder that 'AI is coming for programmers' is too coarse a claim either way.

Method, sources and limitations

This page draws on the same dataset and the same sourcing rule as our full exposure tool: every figure is pulled directly from a primary source and stored under the study that reported it, never averaged or blended across studies.

• The Future of Employment (Frey & Osborne, Oxford Martin School, 2013): Modelled probability that an occupation could be computerised over one to two decades, based on nine O*NET bottleneck variables. Predates large language models entirely. Primary source
• GPTs are GPTs (Eloundou, Manning, Mishkin & Rock (Science), 2023-2024): Share of an occupation's O*NET tasks where an LLM, or software built on one, could cut the time to complete the task by at least half without lowering quality. Exposure, not automation. Primary source
• Anthropic Economic Index (Anthropic, 2026): Observed behaviour, not potential: millions of real Claude conversations mapped to O*NET tasks, then split into automation (the model does the task) and augmentation (the model helps a person do it). Primary source
• GDPval (OpenAI, 2025): Whether frontier models can produce real occupational deliverables — briefs, models, decks, care plans — well enough that an experienced professional prefers them to a human expert's work in a blind comparison. Primary source
• Employment and median wage: BLS Occupational Employment and Wage Statistics, May 2025 national estimates, released 15 May 2026.

Employment and wage data are US national estimates. Observed exposure and the automation/augmentation split describe one AI vendor's traffic (Claude), skew toward technical and professional users, and get noisy at low usage shares — flagged explicitly in the automation-skewed table above. None of the figures on this page are a prediction of what any specific employer will do.

Why we built this

Planetary Labour builds an autonomous go-to-market engine — software that runs marketing work itself, publishing content and building distribution without a team operating it. We sit on the automation side of this question, which is exactly why we would rather rank occupations by real usage data and show our sourcing than publish a single alarming or comforting number.

The pattern in the clusters above matches what we see in our own product: the tasks that move first are the repetitive, well-specified ones with a clear input and output, and what is left is judgement, taste and accountability for the result. That is the argument in our manifesto and across the rest of our future of work writing.

Frequently asked questions