Quick Answer: Asking which jobs will AI replace is asking the wrong question, AI is not replacing occupations, it is absorbing tasks inside them. Anthropic's Economic Index finds that in 49% of occupations AI already covers at least a quarter of the tasks, and the covered tasks require more education than the economy-wide average, not less. Exposure is not displacement: Anthropic's labour-market research finds no systematic unemployment increase among highly exposed workers since late 2022. The one measurable pressure point is the entry ramp, not the profession.
Nobody is going to send you a letter saying an AI took your job. They will simply stop advertising the version of your job that was mostly tasks.
How to Read These Numbers Without Panicking
Most "AI will replace X jobs" headlines fail on the same four reading errors. A strong reading is not a more optimistic one, it is a more precise one.
| What you're reading | Weak reading | Strong reading |
|---|---|---|
| Coverage vs displacement | "74.5% exposure means three-quarters of that job is gone" | Coverage measures how much of the task set models touch, not whether the employer removed the role. Exposure β displacement. |
| Theoretical vs observed | One percentage per occupation, treated as fact | Two numbers exist: what models could do, and what people actually do with them. In some groups the gap is an order of magnitude. |
| Task vs occupation | "My job title is on the list" | Occupations are bundles of tasks with different fates. A heavily covered occupation can still have an irreplaceable remainder. |
| Aggregate vs individual | "This is my personal risk score" | These figures describe millions of task instances. Seniority, sector, regulation and your employer's adoption curve move you off the average. |
Hold those four distinctions and every number below is readable. Drop them and they turn into a horoscope.
The First Inversion: Occupations Aren't Automated, Tasks Are
Anthropic's Economic Index measures which tasks people actually bring to AI models, then maps them back onto occupations. The result is not a list of doomed professions but a map of which parts of jobs are already absorbed.
In 49% of occupations, AI now covers at least a quarter of the task set. Meanwhile 30% of workers show zero measured exposure, not low exposure, none.
That pattern is not noise. When a task sits fully within current model capability, 68% of real usage lands on it; when it is theoretically infeasible, usage falls to 3%. And 97% of observed tasks are ones the models are rated capable of doing. People find the capability boundary and work right up against it.
So "will AI replace my job?" is malformed. The answerable question: which of my tasks sit inside that boundary, and what is left when they go?
Your job title is a billing label. Your task list is the thing under measurement.
The Second Inversion: AI Took the Harder Tasks
This is the finding that breaks most people's mental model, and the one headlines get backwards.
Across the economy, the average task requires 13.2 years of education. The tasks AI covers require 14.4 years. The absorbed work sits above the economy-wide average, which means, arithmetically, the work left behind sits below it.
The automation story we were sold ran the other way: machines take the simple work first and humans climb the skill ladder. Here the ladder is being climbed from the top.
The performance figures agree. Current models deliver roughly a 9Γ speedup at high-school-level tasks (12 years of education) and 12Γ at college level (16 years), while success rates barely move with difficulty: 70% on simple tasks, 66% on college-level ones. Harder work, more advantage, almost no accuracy penalty.
Two of Anthropic's examples make the consequence concrete:
- Travel agents lose complex itinerary planning (13.5 years of education), keep routine ticketing (12.0 years). The role gets simpler: deskilling.
- Property managers lose administrative bookkeeping (12.8 years), keep stakeholder management and negotiation. The role gets harder: upskilling.
Same mechanism, opposite outcomes. Which one you get depends on the residual tasks, not on how exposed your occupation looks.
Who Is Actually Exposed
The named occupations with the highest observed exposure. Read each figure as "how much of this task set already runs through AI", not as a countdown.
| Occupation | Observed exposure |
|---|---|
| Computer programmers | 74.5% |
| Customer service representatives | 70.1% |
| Data entry keyers | 67.1% |
| Medical records specialists | 66.7% |
| Market research analysts and marketing specialists | 64.8% |
| Sales representatives (wholesale & manufacturing, non-technical) | 62.8% |
Where an occupation name is linked above, it points to the nearest CV example profile in our library, not to a page about that exact O*NET title, medical records specialists to the medical assistant example, sales representatives to the sales manager example. The value is seeing how the residual work is evidenced, not a title match.
The other end. The least exposed occupations, named in the same research: cooks, motorcycle mechanics, lifeguards, bartenders, dishwashers, dressing room attendants.
That list settles who this technology pressures. It is not the manual economy. Workers in high-exposure occupations are far more likely to hold graduate degrees, 17.4% against 4.5%, and earn a 47% premium. Their average age is 42.9. The exposed population is educated, mid-career and well paid.
It is also the population that spent twenty years being told credentials were the hedge.
The Third Inversion: Employment Hasn't Moved
This is where honest analysis parts company with clickbait, so state it plainly.
Anthropic's labour-market research finds no systematic increase in unemployment among highly exposed workers since late 2022. Three years into broad deployment, across the most-exposed occupations, the displacement signal is not in the employment data.
Not "not yet, but soon." Not measurable now.
There is exactly one measurable exception, and it is tightly bounded. For workers aged 22-25 in exposed occupations, the job-finding rate is down 14% relative to 2022, roughly 0.5 percentage points, a result Anthropic states barely reaches statistical significance. Treat it as an early signal, not a verdict: adjustment appears to run through hiring rather than firing, concentrating the effect on people trying to get in rather than people already inside.
One forward-looking figure fits: in official occupational projections to 2034, each additional 10 percentage points of task coverage is associated with a 0.6 percentage point lower growth projection. A slow bend in the hiring curve, not a cliff.
A shrinking intake is not a layoff. It is a layoff you only notice five years later, in the shape of the org chart.
The Real Clock: The Capability-Adoption Gap
If employment hasn't moved while half of all occupations are a quarter covered, something is absorbing the shock. That something is adoption lag, and it is enormous.
| Occupational group | Theoretical coverage | Observed coverage |
|---|---|---|
| Computer & mathematical | 94.3% | 35.8% |
| Business & financial operations | 94.3% | 28.4% |
| Management | 91.3% | , |
| Office & administrative support | 90.0% | 34.3% |
| Legal | 89.0% | 20.4% |
| Architecture & engineering | 84.8% | , |
| Arts, design & media | 83.7% | 19.2% |
| Life & social science | 77.0% | , |
| Sales | 62.0% | 26.9% |
| Education & library | 61.7% | 18.2% |
| Healthcare practitioners | 59.9% | , |
| Community & social services | 50.5% | , |
| Food preparation & serving | 16.9% | , |
| Construction & extraction | 16.9% | , |
| Farming, fishing & forestry | 15.7% | , |
| Transportation & material moving | 12.1% | , |
| Building & grounds maintenance | 3.9% | , |
Architecture and engineering is the sharpest illustration in the dataset: 84.8% theoretical capability, only about 5% of it realised in actual use. Sales realises roughly 43% of a far lower ceiling. Computer & mathematical and office & administrative support carry the highest observed coverage of any group, so the disruption you can already see belongs to the fastest-adopting groups, not the largest ones.
If you want a concrete reading of what a group's residual work looks like on paper, one representative profile per group is enough: the financial analyst CV example stands in for business and financial operations, and the registered nurse CV example for healthcare practitioners. Each is one role inside a wide group, not the group itself.
McKinsey sees the same gap from the employer side: 92% of firms are increasing AI spend while roughly 1% call themselves mature in deployment, and executives estimate 4% of employees use AI regularly where employees report 13%.
The strategic reading: where adoption badly lags capability, disruption is pending, not absent. A group at 5% realisation has not been spared. It has not started.
The front line is visible too. As of November 2025 consumer usage splits 52% augmentation to 45% automation; on the API, work wired into systems rather than typed into a chat window, automation is 75%. The task categories that doubled are unglamorous: sales enablement, B2B lead qualification, cold-email drafting, market operations.
Why You Cannot Assess Your Own Exposure
Self-assessment fails here in a documented direction.
In Anthropic's survey work, only 10% of workers rate their own job loss as likely, while 37% or more put a junior colleague's probability above 60%, and 35% or more expect AI to handle most or nearly all of their own tasks within a year.
Read those together: people believe the capability is arriving imminently, believe it will displace someone, and believe that someone is not them.
The bias has a shape. Workers with 15 or more years of experience rate their own exposure about 10 percentage points lower than comparable colleagues. Experience feels like insulation, while correlating with exactly the high-education task bundle that sits above the 13.2-year line.
There is a real effect underneath it: six or more months of tenure is worth about +4 percentage points in task success under full controls. Institutional context is genuinely productive, just a smaller edge than its holders believe.
One mechanical finding is worth internalising. The correlation between prompt sophistication and response sophistication is r > 0.92, output quality tracks input quality almost exactly. The scarce skill is not "using AI"; it is knowing the problem well enough to specify it. That is a judgment skill, and it is the one most CVs fail to evidence.
Your 7-Step Plan
- Stop searching your job title. List your tasks. Twenty lines, verb-first, honest about how your week actually goes. Exposure is measured at this level and nowhere else.
- Sort each task by capability, not comfort. Could a current model with full context produce a competent first version, not a perfect one? The 66-70% success band is the calibration.
- Calculate your residual, not your exposure. What decides your next five years is what remains once the absorbable tasks go, harder (property manager) or easier (travel agent) than before.
- Check your sector's realisation gap, not its ceiling. High theoretical coverage with low observed use means the change is queued, not cancelled.
- Shift weight toward judgment, accountability and relationships. These are the residual categories in every upskilling case in the data: decisions under uncertainty, work someone must sign for, work that needs a person across the table.
- Rewrite your CV to evidence decisions, not throughput. "Handled 200 tickets per month" describes a covered task. "Owned the escalation threshold; cut repeat contacts by re-scoping the policy" describes a residual one. Same job, different read.
- If you are early-career, treat the entry ramp as the live risk. Compete on demonstrated judgment, not on task competence that is now cheap.
The Short Version
Occupations are not being automated. Task bundles are being rearranged.
The absorbed tasks are the educated ones. The residual is not.
Employment has not moved, except at the door, for people aged 22 to 25.
The gap between what these systems can do and what organisations have deployed is the real countdown. TalentVP was built around that gap: what a CV must evidence changes the moment task competence stops being scarce.
Exposure is not a sentence. It is a reading assignment about your own week.
Related Guides
- Is my job at risk from AI? The four-fates task audit, the method behind the residual calculation above.
- New jobs created by AI, and how to position a CV for them, the advertised roles the capability-adoption gap is generating.
- AI and the German job market, for international professionals, why DACH diverges from this baseline.
- CV examples by role, how decision-level evidence reads on the page.
- For senior professionals, positioning when your advantage is judgment, not throughput.
Sources
Data vintage: Anthropic Economic Index reports through June 2026, and the labour-market analysis published March 2026. Anthropic publishes quarterly; figures are reviewed each quarter.
All figures in this article come from published research. The underlying data is Anthropic's, drawn from real usage mapped onto the US Department of Labor's O*NET occupational taxonomy.
- Labor market impacts of AI: A new measure, Anthropic. The observed-exposure measure and the occupation-level figures.
- Anthropic Economic Index: Economic primitives, task complexity, success rates, the education-per-task analysis, deskilling and upskilling examples.
- Anthropic Economic Index: Learning curves, the tenure effect and the fastest-growing task categories.
- Anthropic Economic Index: Cadences, the worker survey.
- Scenarios for our Economic Future, Anthropic Economic Futures, the 2030 projections.
- Future of Jobs Report, World Economic Forum, where cited.
Anthropic's employment analysis is based on US labour data. Where this article discusses the German or wider European market, that is structural reasoning about how those figures translate, not measured German data.
