The short answer
The jobs most exposed to AI are clerical and administrative: data entry, typing, bookkeeping and payroll clerks, finance and insurance clerks, and contact-centre sales. Every major exposure study agrees on that.12 The least exposed are physical: trades, cleaning, care assistance, driving, and most work done with hands rather than documents.45
But exposure is not replacement, and the researchers who build these indices say so in plain words. Three years after ChatGPT, United States labour market data show no economy-wide displacement.7 What they do show is fewer entry-level hires in exposed occupations,6 and official projections of modest decline in a handful of roles, such as customer service, alongside growth in others that are also highly exposed, such as software development.10
This guide defines the terms that get confused, then breaks sixteen representative occupations into their tasks: what is changing, what is not, and what a person in each role can do.
Six words that mean different things
Most bad articles about AI and jobs fail in the first sentence, by treating these as synonyms. They are not.
| Term | What it means | Example of a source that measures it |
|---|---|---|
| Exposure | How much an occupation's tasks overlap with what AI can technically do. Says nothing about whether anyone deploys it, or what happens to the worker. | The ILO's global index, the IMF's exposure measure, and the OpenAI and Pennsylvania study rate exposure.15 |
| Automation potential | The subset of exposure where the tool could do the task without a person, rather than helping a person do it. | The ILO's highest exposure gradient is defined this way: high exposure and low task variability.1 |
| Augmentation | AI making a person faster or better at a task they still perform and own. | Anthropic's usage data classifies conversations as augmentation or automation patterns.15 The IMF estimates roughly half of exposed jobs in advanced economies could benefit. |
| Productivity effect | A measured change in output per hour when workers use AI, in an experiment or a real deployment. | The customer support study: 14 percent more issues resolved per hour.13 |
| Displacement | People losing jobs, or not being hired, because of AI. The only one of these six that is about employment. | The Stanford payroll study finds reduced hiring of 22 to 25 year olds in exposed occupations.6 The Yale tracker finds no economy-wide effect.7 |
| Employment forecast | A projection of future headcount, built on assumptions about technology and the economy. | The United States Bureau of Labor Statistics projections, and the World Economic Forum's employer survey.910 |
The International Labour Organization put the central point directly in a 2026 briefing: exposure indicators "reveal technological susceptibility, not labour market outcomes", and they do not assess whether firms find it profitable to automate, how workflows will change, or how employment and wages will adjust.3 Microsoft's researchers, publishing the most detailed usage-based measure to date, wrote that concluding highly applicable occupations will be automated and lose jobs "would be a mistake, as downstream consequences of new technologies are very hard to predict and often counterintuitive".4
What the exposure studies agree on
Three very different methods, task ratings by experts and models, real usage data, and worker surveys, produce the same picture at the top and bottom of the distribution.
Most exposed: clerical and information-processing work
The ILO's 2023 analysis found that clerical work was the only broad occupation group with high exposure: 24 percent of clerical tasks were rated highly exposed and a further 58 percent moderately exposed, while in every other occupational group the highly exposed share was between 1 and 4 percent.2 Its refined 2025 index placed thirteen occupations in the highest exposure gradient, the majority of them clerical: data entry clerks, typists and word processing operators, accounting and bookkeeping clerks, statistical, finance and insurance clerks, payroll clerks, general clerical support workers, contact centre salespersons, securities and finance dealers, and financial analysts.1 Only 3.3 percent of global employment sits in that gradient, but it is 4.7 percent of women's employment against 2.4 percent of men's, and in high-income countries the gap is 9.6 percent against 3.5 percent.1
Microsoft's analysis of 200,000 real conversations with its Copilot assistant, mapped to United States occupations, found the highest applicability in interpreters and translators, historians, writers and authors, sales representatives of services, customer service representatives, telemarketers, technical writers, proofreaders, editors, public relations specialists, data scientists, web developers and market research analysts.4 The OpenAI and University of Pennsylvania study rated mathematicians, tax preparers, financial quantitative analysts, writers and authors, and web and digital interface designers as fully exposed, and interpreters and translators as the single most exposed occupation on human ratings.5 The two lists overlap heavily, and they describe information work: creating, processing and communicating text and numbers.
Least exposed: physical and relational work
The OpenAI study identified 34 occupations with no exposed tasks at all, almost all physical: agricultural equipment operators, athletes, automotive glass installers, mechanics, cement masons, short order cooks, dishwashers, dredge operators, power-line installers, floor layers, stonemasons, roofers' helpers, meat packers, tire repairers.5 Microsoft's bottom-ranked occupations included phlebotomists, nursing assistants, hazardous materials removal workers, embalmers, oral surgeons, ship engineers, dishwashers, cement masons, massage therapists, surgical assistants, roofers, and housekeeping cleaners, several with applicability scores of zero.4 The OECD's 2023 analysis named food preparation assistants, agricultural labourers, and cleaners and helpers as the least exposed to AI.8
What has actually happened to employment
Exposure tells you where to look. Employment data tells you what has happened. So far the two disagree in an instructive way.
- No economy-wide disruption. The Yale Budget Lab's tracker, updated with data to July 2026, concludes that United States data "does not provide clear evidence of labor market disruption associated with AI". Churn across occupations, AI exposure among the unemployed, and usage measures "all remain flat, lie within historical ranges, or continue along pre-AI trends".7
- A clear entry-level effect. Stanford researchers using payroll data on millions of United States workers found no widespread displacement, but employment of 22 to 25 year olds in the most exposed occupations 19 percent below where it would be had it kept pace with less exposed peers, driven by reduced hiring rather than dismissals, and concentrated in occupations where AI substitutes for tasks rather than complementing workers. Experienced workers show no gap. The authors call the findings descriptive indicators, not causal estimates.6
- Hiring tilting toward seniority. Indeed's analysis of its own postings found that from 2022 to 2026 the more exposed an occupation was, the more its postings declined, and that entry-level postings fell while senior postings rose. Software development postings rose about 15 percent from February 2025 to mid-2026 against a falling market, with most of the increase in senior roles.14
- Modest official projections. In its 2024 to 2034 analysis, the United States Bureau of Labor Statistics attributed projected declines partly to AI in a specific set of roles: customer service representatives, legal secretaries, procurement clerks and claims adjusters. Its current 2025 to 2035 projections show customer service representatives declining about 5 percent over the decade and bookkeeping clerks about 6 percent, the latter attributed to accounting software rather than AI, while software developers grow around 10 percent, accountants 5 percent, registered nurses 6 percent and electricians 9 percent. The bureau states it "assumes that the overall pace of technological change will be consistent with past experience" and that "because the capabilities of AI systems continue to evolve, these assessments can change".1011
Two occupations make good test cases. Interpreters and translators are the most exposed occupation on both the usage-based and the ratings-based indices,45 yet the Bureau of Labor Statistics still projects slight growth for them, noting that many of these jobs cannot be entirely automated.12 Radiologists were famously predicted in 2016 to become obsolete within years; a decade on, reporting based on industry and physician survey data describes rising demand and record pay.16 Neither case proves AI will never displace anyone. Both show that a whole-profession prediction from a task-level measure is the wrong inference.
Sixteen occupations, task by task
The table below is a Stay Useful synthesis of the evidence above. "AI exposure" reports what the indices say. "Likely near-term change" is our reading of the observed data and official projections, and it is a judgment. "What a worker can do" is advice, not a guarantee of anything.
| Role | Typical tasks | AI exposure | Tasks increasingly automatable | Tasks still strongly human | Likely near-term change | What a worker can do |
|---|---|---|---|---|---|---|
| Administrator or office clerk | Correspondence, scheduling, data entry, document preparation, records, coordination. | Highest. Clerical work dominates the ILO's top gradient.12 Administrative assistants and data entry clerks head the WEF's declining list.9 | Routine correspondence, data entry, formatting, first-pass summaries and minutes. | Exceptions, judgment about priorities, relationships across the organisation, knowing who to ask. | Fewer roles built only on document processing; remaining roles carry more coordination. United States projections show office and administrative support declining about 4 percent over a decade.10 | Become the person who runs the AI-assisted workflows for the team, and document the time and errors saved. |
| Customer service representative | Answering queries, resolving issues, updating records, escalation. | High. Seventh on Microsoft's applicability list.4 Named as an AI-affected decline by the Bureau of Labor Statistics.10 | Suggested replies, case summaries, routine queries with clear answers. | Difficult conversations, cases outside the script, de-escalation, deciding what the organisation will do. | Productivity rose 14 percent with AI assistance in one large deployment, most for novices.13 Projected decline of about 5 percent over the decade, with weaker entry-level hiring.610 | Build skill in the hard cases the tool hands over, and in supervising and improving the tool's suggestions. |
| Accountant or bookkeeper | Recording transactions, reconciliation, compliance, reporting, advice. | High for clerks: bookkeeping and payroll clerks are in the ILO's top gradient.1 Accountants rated fully exposed by the OpenAI study.5 | Data entry, reconciliation, first drafts of reports and client notes, explaining standards. | Sign-off, advice, judgment about a client's situation, accountability to regulators. | Bookkeeping clerks projected to decline about 6 percent; accountants projected to grow 5 percent as advisory work becomes more prominent.10 The ratio of processing to advising shifts. | Move toward the advisory and review end; own the checks on automated processing. |
| Software developer | Writing code, testing, design, review, deciding what to build. | High usage: computer and mathematical tasks are around a third of Anthropic's consumer usage and more of its business usage.15 | Boilerplate, tests, documentation, routine features, explaining unfamiliar code. | System design, trade-offs, reviewing AI output, accountability for what ships. | The most contradictory picture: early-career hiring down sharply,6 postings recovering toward senior roles,14 official projections of 10 percent growth.10 One randomised study found experienced developers slower with early-2025 tools while believing they were faster.17 | Develop judgment about architecture and review; the writing of code is becoming the smaller part of the job. |
| Writer, editor or marketing content producer | Drafting, editing, adapting, researching, publishing. | Very high. Writers and authors third on Microsoft's list; rated fully exposed by the OpenAI study; technical writers, editors and proofreaders also high.45 | First drafts, adaptations, summaries, routine copy, proofreading. | Judgment about what to say, voice, original reporting, accountability for accuracy. | United States projections put writers at zero growth, explicitly citing AI dampening demand.12 Volume roles shrink; roles defined by judgment persist. | Specialise in the parts AI is worst at: original material, editorial judgment, and verifying what the tools produce. |
| Interpreter or translator | Translating documents, interpreting speech, localisation. | The highest on both major indices.45 | Routine document translation, first drafts, high-volume low-stakes content. | Live interpretation in high-stakes settings, nuance, legal and medical accuracy, cultural judgment. | Still projected to grow slightly in the United States, with the caveat that many jobs cannot be entirely automated.12 Work shifts toward review and high-stakes settings. | Position as the reviewer and the specialist; build evidence of the errors machine translation makes in your domain. |
| Paralegal or legal support | Research, document review, drafting, discovery, filing. | High. Legal secretaries rated fully exposed by the OpenAI study; legal officials and legal secretaries on the WEF's declining list.59 | Contract review, discovery, research summaries, first drafts. | Judgment, client contact, verification of everything the tool produces, professional accountability. | Paralegals projected flat in the United States, with AI explicitly cited as limiting growth.10 Courts have sanctioned lawyers for unverified AI citations. | Become the verification layer; understand the tools well enough to catch their fabrications. |
| Teacher | Planning, teaching, assessment, pastoral care, administration. | Moderate. Education is a significant share of AI usage,15 but the core work is relational. | Differentiating materials, drafting communications, first-pass marking, administration. | Teaching, relationships, judgment about individual students, classroom management. | The WEF lists secondary and university teachers among the largest absolute growth roles.9 Administrative load may fall; the role does not. | Use the tools to reclaim time for the relational work, and keep student data out of them. |
| Nurse | Assessment, care, medication, documentation, coordination, communication. | Low. Nursing assistants are near the bottom of Microsoft's list; registered nurses sit in the augmentation-heavy group in the Stanford analysis.46 | Documentation, handover notes, information retrieval, scheduling. | Almost everything else. | Projected growth of around 6 percent in the United States; among the WEF's largest growth roles.910 Documentation burden may ease. | Shape how documentation tools are used in your setting; the clinical judgment is the job. |
| Electrician or tradesperson | Installation, repair, compliance, quoting, scheduling, customer contact. | Near zero for the trade itself on every index.45 The OECD notes construction and extraction face higher risk from automation broadly, once robotics is included.8 | Quoting, compliance paperwork, looking up regulations, scheduling. | The work. | Electricians projected to grow around 9 percent in the United States, with data centre electricity demand cited as a driver.10 The office side of the business changes. | Use AI to remove the administrative load from a small business; the trade is not where the exposure is. |
| Truck or delivery driver | Driving, loading, route management, customer interaction, compliance. | Near zero for language-model exposure.45 The relevant technology is autonomous vehicles, which the exposure indices do not measure. | Route planning, paperwork, communications. | Driving in complex environments, handling, customers, judgment on the road. | Projected growth in the United States with no autonomy adjustment; delivery drivers among the WEF's largest absolute growth roles.910 Autonomous trucking is in early commercial deployment on limited routes; its timeline is genuinely uncertain. | Watch the autonomy evidence rather than the AI headlines; build skills in the parts of logistics that need a person on the ground. |
| Salesperson | Prospecting, meetings, proposals, negotiation, account management. | High for information-heavy selling: sales representatives of services fourth on Microsoft's list, telemarketers and advertising sales also high.4 | Research, proposals, follow-ups, routine outreach, call preparation. | Relationships, negotiation, judgment about what the customer actually needs, trust. | Shop salespersons among the WEF's largest growth roles; telemarketers and door-to-door sales declining.9 Sales representatives sit in the augmentation-heavy group in the Stanford analysis.6 | Spend the reclaimed preparation time on the relationship, and track the conversion difference. |
| HR or recruiter | Sourcing, screening, interviewing, onboarding, policy, employee relations. | Moderate. Business and administrative work is broadly exposed; no index singles out HR specifically. | Job advertisements, scheduling, first drafts of policy, summarising feedback. | Judgment about people, difficult conversations, decisions about hiring and discipline, legal accountability. | United States projections show growth of around 6 percent with no AI adjustment.10 Screening decisions attract regulatory attention; human review is expected. | Own the governance of how AI is used on people decisions; that is a role in itself. |
| Financial analyst | Modelling, research, reporting, recommendations. | Very high on task ratings: financial analysts in the ILO's top gradient; quantitative analysts rated fully exposed.15 | Data gathering, model building, first drafts of reports, routine analysis. | Judgment about what the numbers mean, recommendations, accountability for them, client trust. | Exposure and projections point opposite ways: United States projections show around 7 percent growth.10 Expect the analysis to be assisted and the judgment to be the job. | Move up the chain from producing analysis to interpreting and defending it. |
| Graphic designer | Concepts, layouts, production, client communication, brand stewardship. | High for production work. Rated the fastest-declining creative role by employers in the WEF survey.9 | Variations, resizing, routine production, first concepts. | Brand judgment, original concepts, client relationships, taste. | Projected slight decline in the United States, with automated design tools cited.12 Production-heavy roles shrink; concept and brand roles persist. | Position as director of the tools rather than operator of them, and build a portfolio of judgment. |
| Manager | Deciding, allocating, coaching, reporting, coordinating, being accountable. | High exposure on task ratings, because much of the work is information work; the OECD finds managers among the most exposed to AI yet least at risk of automation.8 | Reports, summaries, plans, status updates, meeting preparation. | Judgment, allocation of people, accountability, handling what no process anticipated. | Chief executives sit in the augmentation-heavy group in the Stanford analysis.6 Less time on documents about the work; the work of managing remains. | Decide how AI is deployed on your team: to help people do more, or to do the same with fewer. That decision is the exposure. |
Which jobs are safest from AI, and what "safe" leaves out
The honest list of the least exposed occupations is a list of physical and relational work: trades, care, cleaning, hospitality, agriculture, driving, and most jobs where presence matters.458 Three qualifications matter.
- Safe from language models is not safe from automation. The exposure indices measure generative AI working on information. The OECD's broader 2023 analysis, which included other technologies, found low- and middle-skilled jobs in construction, extraction, farming and production at the highest risk of automation, and management and community and social service roles at the lowest.8 Robotics and autonomous vehicles follow a different timeline from chat assistants.
- Safe does not mean unchanged. The paperwork, communication and scheduling around physical jobs are as exposed as anyone's. A tradesperson's trade is not changing; a tradesperson's admin is.
- Growth is not the same as safety either. Employers expect nurses, teachers, delivery drivers and farmworkers to be among the largest sources of new jobs to 2030.9 That reflects demographics and demand as much as AI.
What employers expect, and why to discount it a little
The World Economic Forum's 2025 survey of over 1,000 large employers is the most cited forecast in this field. It expects 170 million jobs created and 92 million displaced by 2030, a churn of 22 percent, and 40 percent of the surveyed employers anticipate reducing their workforce where AI can automate tasks.9 That is what employers expect to do, not a count of jobs that will disappear. Its fastest-declining roles are postal clerks, bank tellers, data entry clerks, cashiers, administrative assistants, printing workers, accounting and payroll clerks, stock clerks, transport attendants, door-to-door sales, graphic designers and claims adjusters. Its fastest-growing are big data specialists, fintech engineers, AI and machine learning specialists, software developers, security specialists, and, in absolute terms, farmworkers, delivery drivers, nurses, teachers and care aides.9
Read it as what it is: what employers said they expected, extrapolated. Employer expectations have run ahead of employer behaviour. In a 2026 survey of nearly 6,000 executives in four countries, nine in ten reported no AI impact on their own firm's employment or productivity over the previous three years, while forecasting a 0.7 percent employment reduction over the next three.18 Expectations are worth knowing because they drive hiring decisions today. They are not observations.
New Zealand
New Zealand has little occupation-level evidence of its own. PwC's 2026 analysis of New Zealand job advertisements found that postings asking for AI skills more than doubled from about 3,900 in 2024 to about 9,600 in 2025, around 2.7 percent of all postings. Postings in the occupations least exposed to AI have grown far faster since 2012 than postings in the most exposed occupations, roughly 22 times against roughly 8 times, although the most exposed quartile still had the largest share of postings, around 128,000 in 2025.19 That is company research based on advertisements, not employment, but the direction is consistent with the international picture: exposed work is not disappearing, and it is growing more slowly than the rest.
The Reserve Bank of New Zealand published an analytical note in February 2026 estimating AI and robotics exposure across the New Zealand labour market using large language models, finding professional, managerial and administrative occupations substantially more exposed to AI than labourers and machinery operators, and robotics exposure concentrated in the latter.20 Treasury's 2024 analysis observed that a high-skill economy like New Zealand may be more exposed than average and that the net employment effect "remains uncertain".21
How to read your own role
A table describes an occupation. It cannot describe you, because two people with the same title have different task mixes, different employers and different judgment. Three questions translate the table into your situation.
- What share of your week is in the "increasingly automatable" column? Write down your actual tasks and sort them. That share is your practical exposure, whatever your title says.
- Is your employer deploying AI to help people do more, or to do the same with fewer people? This is the single largest factor the data identify, and it is knowable: read the policy, ask your manager, look at hiring.6
- Could you demonstrate that you make the tools produce better results than they do alone? The people insulated so far are experienced workers whose judgment the tool cannot replace.6 That is a position you can build toward at any career stage.
Our guide on whether AI will replace your job provides a full framework for scoring your own tasks, and our guide on how to stay valuable at work as AI gets better turns the third question into a plan.
Common misunderstandings
- "These jobs will disappear." No occupation in the current data is disappearing. Exposed occupations are changing shape and hiring fewer juniors.67
- "Exposure means replacement." The researchers who build the indices explicitly reject that reading.34
- "Physical jobs are safe." Safe from language models. The broader automation picture, including robotics, places some physical occupations at high risk over a longer horizon.8
- "The 47 percent figure." A 2013 estimate that rated whole occupations and predates generative AI; a task-based re-analysis a few years later found about 9 percent. It is not a current measure of anything. Our guide on whether AI will replace your job explains what each of the headline numbers measures.
- "Employers know." Employer forecasts are intentions, and so far intentions have outrun actions.918
Related questions
What jobs will AI replace by 2030?
Nobody can answer that honestly with a list. The best available forecast, from the World Economic Forum's employer survey, expects the largest declines in clerical roles such as data entry, bank tellers, postal clerks, cashiers and administrative assistants, and it is an expectation survey with a record of running ahead of reality.9 The observed pattern is roles being reshaped and entry-level hiring slowing, not occupations vanishing.67
Is accounting safe from AI?
The processing side of accounting is among the most exposed work there is; bookkeeping and payroll clerks sit in the ILO's highest gradient and United States projections show bookkeeping clerks declining.110 The professional side is projected to grow, with advisory work becoming more prominent.10 The profession is not disappearing; the ratio of processing to judgment inside it is shifting.
Will AI replace nurses or teachers?
Not on any current evidence. Both sit low on exposure for their core work, both appear among the largest sources of projected job growth, and both are relational and, in nursing, physical.4910 Documentation and administration will change. The jobs will not.
Will AI replace lawyers?
Legal support work is highly exposed: research, document review and drafting. Paralegals are projected flat in the United States with AI cited as the reason.10 Lawyers carry professional accountability that regulators and courts enforce, and courts have sanctioned lawyers who filed fabricated AI citations. Expect the profession to need fewer people producing documents and more people verifying them.
Will AI replace truck drivers?
Language models barely touch driving, so the exposure indices rate it near zero.45 The relevant technology is autonomous vehicles, which is in early commercial deployment on limited routes and on a genuinely uncertain timeline. United States projections still show growth for drivers.10 Watch the autonomy evidence, not the AI headlines.
Will AI replace programmers?
The evidence is contradictory in an instructive way: the highest usage of any occupation, sharply reduced hiring of the youngest developers, postings recovering toward senior roles, and official projections of solid growth.6141510 The job is being reshaped from the bottom, with code-writing a smaller share and judgment, review and accountability a larger one.
What percentage of jobs will AI replace?
There is no defensible number. Every percentage in circulation measures exposure or expectation, not replacement.139 The only observed employment shift so far is a 19 percent shortfall in the employment of 22 to 25 year olds in exposed occupations in the United States, relative to less exposed peers, driven by hiring, and its authors describe it as descriptive rather than causal.6 Anyone giving you a single percentage of jobs "replaced" is quoting an exposure figure.