The short answer
AI is more likely to change the tasks inside your job than to remove the job itself. Your actual risk depends on how much of your work can be done reliably from digital information, how cheaply the result can be checked, how much human judgment and accountability the role carries, and whether your employer deploys AI to replace people or to help them do more.
The large numbers you have seen, such as "40 percent of jobs" or "one in four workers", measure how much a job overlaps with what AI can do. They do not measure jobs lost. Three years after ChatGPT, economy-wide employment data in the United States and Denmark show no broad displacement.78 The one clear change so far is fewer entry-level hires in the most exposed occupations.9
The useful question is not "will AI replace my job" but "which of my tasks will change, how soon, and what will I be doing instead". This guide gives you a way to answer that for your own work.
Why "will AI replace jobs?" is the wrong level of analysis
A job is a bundle of tasks held together by a title, a contract and a lot of context. AI does not arrive at the level of the bundle. It arrives at the level of individual tasks: drafting a summary, classifying an email, reconciling a spreadsheet, answering a routine question. Some of those tasks it can do well. Others it cannot do at all, because they involve a physical object, a relationship, a decision someone has to be accountable for, or information that is not written down anywhere.
This distinction is not a rhetorical comfort. It changes the numbers. The most quoted estimate in the field, Frey and Osborne's 2013 finding that about 47 percent of United States employment was "at risk" of computerisation, was built by rating whole occupations.5 When OECD researchers re-ran the exercise in 2016 using data on the actual tasks people in those occupations perform, the share of jobs whose tasks were mostly automatable fell to about 9 percent across 21 countries.6 Same technology, same economy, a fivefold difference, because occupations labelled "high risk" still contained a large share of tasks that were hard to automate.
The International Labour Organization reached the same conclusion about generative AI in 2025. After rating nearly 30,000 tasks, its researchers found that one in four workers worldwide is in an occupation with some exposure, but only 3.3 percent of global employment sits in the highest exposure band, and they concluded that "transformation of jobs is the most likely impact" because most occupations consist of tasks that still require human input.1
What the big numbers actually measure
Almost every headline about AI and jobs comes from one of a handful of studies. They are good studies. They are also routinely misread, because "exposed" gets translated to "replaced" somewhere between the report and the headline. Here is what each one measured.
| Source | Headline figure | What it measures | What it does not say |
|---|---|---|---|
| ILO and NASK, 20251 | One in four workers in an occupation with some exposure; 3.3 percent of global employment in the highest exposure band | How much of an occupation's tasks overlap with what generative AI can do, rated task by task and validated by experts. Global, by country income group. | It does not predict job loss. The authors say transformation is the most likely outcome. |
| IMF staff, 20242 | Almost 40 percent of global employment exposed; about 60 percent in advanced economies | Exposure to AI, combined with a judgment about whether AI is likely to complement or substitute for the worker. | Of the exposed jobs in advanced economies, the authors estimate about half could benefit from AI rather than be harmed by it. |
| Eloundou and colleagues (OpenAI and University of Pennsylvania), 20234 | About 80 percent of US workers could have at least 10 percent of their tasks affected; about 19 percent could have at least half affected | Whether a large language model could cut the time a task takes by half at the same quality. Rated by people and by GPT-4 against a US task database. | The authors explicitly make no prediction about adoption or timing. |
| Frey and Osborne, 20135 | About 47 percent of US employment at risk of computerisation | The probability that a whole occupation could be automated, based on expert labels for 70 occupations extended to 702. | It rated occupations, not tasks. The task-based re-analysis found about 9 percent.6 |
| World Economic Forum, 20253 | 170 million jobs created and 92 million displaced by 2030, a churn of 22 percent | What over 1,000 large employers said they expected, extrapolated to the global workforce. Covers all drivers of change, not only AI. | It is an expectation survey, not an observation, and it is a forecast about 2030. |
| OECD, 202311 | Occupations at highest risk of automation account for about 27 percent of employment | Occupational automation risk across OECD countries, taking AI into account. | The same report found "little evidence of significant negative employment effects due to AI" so far. |
Notice what these have in common. They are estimates of what is technically possible, or what employers expect. None of them is a count of people who have lost a job to AI. That count exists, in a limited form, and it looks very different.
What has actually happened so far
The strongest evidence about what AI is doing to employment comes from studies that look at real payroll, tax and survey data after the tools became widely available, rather than at what the tools could do in principle. Three findings stand out.
Across whole economies, very little has moved
Researchers at the Yale Budget Lab and Brookings tracked the mix of occupations in the United States labour market against what happened after earlier technology shifts. Their October 2025 assessment was that the broader labour market "has not experienced a discernible disruption since ChatGPT's release", and their measures of exposure, automation and augmentation showed "no sign of being related to changes in employment or unemployment".7 Their January 2026 update found the same metrics flat or within historical ranges.7
In Denmark, where researchers could link two large adoption surveys covering 25,000 workers in 11 exposed occupations to administrative records on pay and hours, the result was what the authors call "precise zeros": no significant effect on earnings or recorded hours in any occupation, with confidence intervals ruling out effects larger than about 1 to 2 percent two years after ChatGPT launched. Users reported saving an average of 2.8 percent of their working time.8
Employers report less impact than they forecast
A 2026 survey of nearly 6,000 senior executives in the United States, United Kingdom, Germany and Australia found that 69 percent of firms actively use AI, yet nine in ten reported no impact on their own employment or productivity over the previous three years. Looking forward, they expected AI to raise productivity by an average of 1.4 percent and cut employment by 0.7 percent over three years. Their employees, asked separately, expected employment to rise slightly.13
The one place the data have moved: entry-level hiring
The exception is young workers in the most exposed occupations. Using payroll data covering millions of United States workers through June 2026, researchers at Stanford's Digital Economy Lab found that employment of 22 to 25 year olds in AI-exposed occupations now stands 19 percent below where it would be had it kept pace with less exposed peers. Experienced workers in the same occupations show no comparable gap. The change is operating mainly through reduced hiring rather than through people being let go, and it is concentrated in occupations where AI is used to substitute for tasks rather than to complement workers.9
Automation or augmentation: the distinction that matters most
The economist David Autor made the point in 2015 that automation both substitutes for labour and complements it: machines take over routine, codifiable tasks while amplifying the value of the problem-solving, adaptability and judgment that surround them, and higher output raises demand for the people who supply those.12 Whether a given deployment of AI is a substitute or a complement is not a property of the technology. It is a decision made by whoever deploys it.
The clearest evidence that this decision matters comes from the same Stanford payroll study: declines in young workers' employment concentrate in occupations where AI usage mostly substitutes for human tasks, while in occupations where AI mostly complements workers, employment is flat or rising.9 The IMF's analysis reaches a related conclusion from a different direction, estimating that around half of exposed jobs in advanced economies could benefit from AI rather than lose out to it.2
How AI is actually being used gives a rough sense of which way organisations are leaning. Anthropic's analysis of usage on its own platform, which is company research about one vendor's customers rather than a picture of the whole economy, found that about 77 percent of business use through its programming interface followed automation patterns, where the tool is given a task to complete, compared with roughly half of individual use, where people more often iterate, learn and check.10 The same source reported the individual split shifting further toward augmentation by early 2026.10
For you, the practical reading is this. Your own use of AI is mostly a matter of augmentation, and it makes you faster. The risk to your job comes from how your employer, or your employer's competitors, decide to use it. That is a reason to be the person in the room who understands the work well enough to say where AI helps and where it does not.
Why capability does not automatically equal deployment
Even where AI can do a task, several things stand between "can" and "does". Understanding them is what separates a realistic risk assessment from a frightening one.
- Cost of being wrong. A tool that is right 95 percent of the time is a gift for drafting and a liability for anything where the remaining 5 percent is expensive, dangerous or irreversible. The higher the cost of an error, the more human checking stays in the loop, and checking is work.
- Accountability. Someone has to sign the accounts, certify the design, approve the loan, or stand in front of the client when things go wrong. Regulators, insurers, courts and customers still expect that someone to be a person.
- Integration. Most real work runs on systems, permissions and half-documented processes that a general-purpose model cannot see. Getting AI into a workflow reliably is an engineering and change project, which is why executives who report using AI still report little impact.13
- Inputs that are not digital. A great deal of work depends on what was said in a meeting, what a machine sounds like, what a patient looks like, or what the site actually contains. If the inputs are not in a form the model can read, the task is not exposed in practice, whatever the index says.
- Economics. The macro-level effect is smaller than the rhetoric. One widely cited estimate by the MIT economist Daron Acemoglu puts the productivity gain from AI at no more than about 0.66 percent over ten years once you account for which tasks can actually be automated at reasonable cost.14 Goldman Sachs, whose 2023 figure of 300 million exposed jobs is often quoted, expects in its own base case that 6 to 7 percent of workers are displaced over a transition lasting around a decade, with most of them finding new roles.15
- Diffusion takes time. New Zealand's Treasury noted in 2024 that the country's historically slow uptake of new technology and low investment in intangible capital could be a barrier to realising AI's benefits.16 The same slowness cuts both ways: it delays the gains, and it delays the disruption.
None of this means the constraints are permanent. Costs fall, accuracy improves, and organisations learn. It does mean that the timeline for your job is set by your industry's tolerance for error, its regulation, its systems, and its management, not by the release date of the next model.
The Stay Useful Job Change Framework: eight things that determine your exposure
The research above points to a consistent set of factors. We have organised them into a framework you can apply to your own work. This is a Stay Useful interpretation of the evidence, built to help you think, not a research finding. Score each dimension for the task you are assessing, or for the two or three tasks that make up most of your week.
Stay Useful framework
The Stay Useful Job Change Framework
Score each dimension for a task, then read the pattern rather than the total. High repeatability, digital inputs, high error tolerance and low integration needs describe a task that is likely to change soon. High judgment, high accountability, a physical component and a high cost of mistakes describe a task that is more likely to be reshaped than removed.
Task repeatability
Does the task follow a recognisable pattern each time, with similar inputs and similar outputs? Repeatable tasks are where AI is deployed first, because the effort of setting it up pays back many times.
Digital availability of inputs
Is everything the task needs already in a document, a system or a message? If the inputs live in conversations, in someone's head, or in the physical world, the task is less exposed than its description suggests.
Error tolerance
How bad is a confident mistake? Drafting and brainstorming tolerate errors because a person reviews the result. Filing, certifying and advising do not.
Need for human judgment
Does the task require weighing competing interests, reading a situation, or deciding what the goal should be? Models are good at producing options and weak at owning a choice.
Need for trust and accountability
Does someone need to be answerable for the result, professionally, legally or personally? Where accountability is required, a person stays in the loop even when the drafting is automated.
Physical-world component
Does the task involve hands, presence, or a physical environment? Generative AI works on information. The least exposed occupations in the OECD's analysis were food preparation assistants, agricultural labourers, and cleaners and helpers.11
Cost of mistakes
Separate from how often errors happen: what does one cost? A wrong word in an internal draft costs a minute. A wrong figure in a filing, a dose, or a structural calculation costs far more, and organisations price that in.
Degree of workflow integration required
Can the task be done by pasting text into a chat window, or does it need the tool wired into systems, permissions and approvals? The more integration required, the slower deployment tends to be.
How to use it
- Write down the five to eight tasks that take most of your working week. Be specific: "reconcile supplier invoices against purchase orders" rather than "finance admin".
- For each task, mark it high, medium or low on the eight dimensions. Ten minutes per task is enough.
- Sort the tasks into three groups: likely to change soon (repeatable, digital, tolerant of error, easy to integrate), likely to be reshaped (some AI assistance, human judgment and accountability remain), and unlikely to change much (physical, relational, or dependent on information that is not written down).
- Look at what share of your week sits in the first group. That share, not your job title, is your practical exposure.
- For the first group, decide whether you want to be the person who uses AI to do those tasks faster and better, or the person who is still doing them by hand when your employer notices the difference.
Warning signs, and signs of resilience
Applied honestly, the framework tends to surface some patterns. These are generalisations from the evidence, not verdicts about any particular role.
Signs a role is highly exposed
- Most of the week produces standard documents, summaries or classifications from digital inputs.
- Output is checked cheaply, or not at all, before it is used.
- The role is defined by processing volume rather than by relationships or decisions.
- Clerical and administrative support work sits here. The ILO found it to be the most exposed occupational group.1
- Your employer is hiring fewer people into junior versions of the role than it used to.9
- Colleagues are already using AI for the same tasks and producing acceptable results.
Signs a role is comparatively resilient
- Much of the work happens in rooms, on sites, or with people, and the important information is not written down anywhere.
- You are accountable for outcomes in a way a regulator, client or court recognises.
- Errors are expensive, and the organisation knows it.
- The role involves deciding what the problem is, not just producing the answer to a stated one.
- You have accumulated tacit knowledge that lets you tell when an output is wrong. Experienced workers in exposed occupations have so far been unaffected.9
- You are the person who works out where AI helps and where it does not, rather than the person it is done to.
How this plays out across different occupations
The following table applies the task view to a handful of roles. It is illustrative. Our guide to which jobs are most exposed to AI covers many more occupations in detail, with the evidence for each.
| Role | Tasks changing first | Tasks that remain human | What the evidence says |
|---|---|---|---|
| Administrator or office coordinator | Drafting routine correspondence, scheduling, preparing standard documents, first-pass summaries of meetings and threads. | Handling exceptions, knowing who to ask, judging what matters to whom, the relationships that make an office work. | Clerical support is the most exposed occupational group in the ILO index.1 Volume of routine document work is the exposure; the coordination is not. |
| Accountant or bookkeeper | Data entry, reconciliation, first drafts of client notes, explaining a standard in plain language. | Signing off, advising, judging what a client's situation actually calls for, being accountable to a regulator. | High exposure to task change, strong accountability constraint. The OECD found high-skill occupations most exposed to AI yet "still some of the least at risk of automation".11 |
| Customer service representative | Suggested replies to common queries, summarising a case, drafting follow-ups. | Difficult conversations, cases outside the script, de-escalation, decisions about what the organisation will do for the customer. | A study of 5,000 support agents found AI assistance raised issues resolved per hour by around 14 percent, with the largest gains for the least experienced.17 Entry-level hiring in this occupation has weakened in United States payroll data.9 |
| Software developer | Boilerplate code, tests, documentation, explaining unfamiliar code, first drafts of routine features. | Deciding what to build, system design, judging trade-offs, reviewing what the tool produced, accountability for what ships. | Employment of 22 to 25 year old developers in the United States fell by nearly a fifth from 2024, while experienced developers were stable.18 The role is being reshaped from the bottom up rather than removed. |
| Nurse or care worker | Documentation, drafting handover notes, retrieving information, administrative reporting. | Assessment, physical care, communication with patients and families, clinical judgment, presence. | Low exposure on the task indices because most of the work is physical and relational. Documentation load may fall, which is a change in the job, not a threat to it. |
| Electrician or tradesperson | Quoting, compliance paperwork, looking up regulations, scheduling, customer communication. | Almost all of the actual work. | Physical occupations sit at the bottom of every exposure index.411 AI changes the office side of the business, not the trade. |
If you are early in your career, or hiring people who are
The clearest signal in the data so far is about the bottom rung. Employers appear to be hiring fewer juniors into exposed occupations, not letting experienced people go.9 The likely mechanism is straightforward: the tasks that used to be given to a new starter to learn on, the first drafts, the routine analysis, the standard responses, are the tasks AI does most readily. Stanford's 2026 AI Index reports that a third of organisations expect AI to reduce their workforce in the coming year even though large-scale losses have not yet appeared in overall employment data.18
If you are in your first years of work, this changes the strategy. The apprenticeship model, where you did the routine work until you were trusted with the judgment work, is being compressed. That makes it more important, not less, to build the judgment early: to understand why the work is done, to check the tool's output against reality, and to be able to show what you have improved. Our guide on how to stay valuable at work as AI gets better is about exactly that.
If you manage a team, the same evidence is a warning about your pipeline. A team that stops hiring juniors because AI does the junior work will have no seniors in ten years.
A note on New Zealand
The New Zealand evidence is thinner than the international evidence and mostly comes from industry surveys. Treasury's 2024 analysis concluded that the net effect of AI on employment "remains uncertain", that New Zealand's high share of skilled work may make it more exposed than average, and that a focus on skills will be important.16 The AI Forum's 2025 productivity survey of New Zealand organisations found 14 percent attributing some job losses directly to AI, up from 7 percent in earlier rounds, while 55 percent reported AI creating new career opportunities and nearly half of adopters reported reduced hiring needs.19 Those figures come from an industry body's survey of adopting organisations, not from official statistics, and should be read as a temperature check rather than a measurement.
The practical implication is the same as everywhere else, with one addition: because adoption here tends to lag, a New Zealand worker who understands the tools well is scarcer than the same worker in a market where everyone has been using them for years.
What to do in the next 30 days
Worry is not a plan. This is.
- Week 1: audit your tasks. List the tasks that fill your week and run them through the framework above. You are looking for the share of your time that sits in "likely to change soon".
- Week 2: try AI on one exposed task, properly. Pick the most repeatable task on your list. Use an approved tool, give it the full context, and check the output against your own version. Note honestly where it helped and where it was wrong. Our beginner's guide to using AI at work walks through how.
- Week 3: find out how your employer is thinking. Read the AI policy if there is one. Ask your manager what the organisation is planning. Whether AI is being deployed to help people do more or to do the same with fewer people is the single biggest factor in your risk, and it is knowable.
- Week 4: start a record. Write down one thing you improved with AI this month, with a before and after. Time saved, errors caught, output produced. Proof that you can use these tools to produce better work is the asset that travels with you.
Common misunderstandings
- "X percent of jobs will be replaced." Every large percentage in circulation is an exposure or expectation figure. None is a count of jobs lost.1234
- "If AI can do it, it will be done by AI." Capability is one input. Cost of error, accountability, integration effort and economics decide whether and when.1314
- "High-skill jobs are safe." High-skill, high-wage occupations are among the most exposed to AI on the task indices. They are also among the least likely to be fully automated, because of judgment and accountability. Both things are true.211
- "Nothing has happened, so nothing will." Economy-wide data are flat so far, but entry-level hiring in exposed occupations has weakened measurably, and adoption is still rising.79
- "Using AI at work makes me replaceable." The evidence points the other way. Occupations where AI complements workers have stable or rising employment; the risk sits where it substitutes for them.9
Related questions
Which jobs will AI replace first?
No occupation is being "replaced" outright in the current data. The tasks changing first are repeatable, text-based and checked cheaply: routine correspondence, standard documents, first-pass summaries, basic customer replies, boilerplate code. The occupations built mostly from those tasks, clerical and administrative support above all, are the most exposed.1 See our task-by-task guide to which jobs are most at risk from AI.
What jobs are safest from AI?
Work that is physical, relational, or depends on information that is not written down: trades, care, hospitality, agriculture, and most roles where presence matters. The OECD's least exposed occupations were food preparation assistants, agricultural labourers, and cleaners and helpers.11 "Safe" still does not mean "unchanged": the paperwork and communication around those jobs will change.
Will AI replace office workers?
Office work is where AI has the most reach, because the inputs are digital. The realistic expectation is that routine document work shrinks, that fewer people are hired to do only that, and that the remaining roles carry more judgment, coordination and checking. The ILO found clerical work the most exposed group but concluded that transformation, not replacement, is the most likely outcome.1
Will AI replace accountants?
The processing parts of accounting are highly exposed. The advisory, sign-off and accountability parts are not, and regulators require a responsible person. Expect the ratio of processing to advising inside the profession to shift, and expect entry-level processing roles to be the ones affected first.
Will AI replace programmers?
Not on current evidence, but the profession is changing from the bottom. Hiring of the youngest developers in the United States has fallen sharply while experienced developers are stable.918 Writing code is becoming a smaller share of the job; deciding what to build, reviewing what the tool produced, and owning what ships are becoming larger shares.
Will AI replace managers?
Managers are exposed on the task indices because much of their work is information work: reports, summaries, plans.11 The core of the role, deciding what matters, allocating people, being accountable for results, and handling the situations no process anticipated, is among the least automatable work there is. What changes is how much of a manager's week goes on producing documents about the work rather than doing the work of managing.
How quickly will AI affect jobs?
Slower at the level of the whole economy than the headlines suggest, and faster at the level of specific tasks than many people expect. Economy-wide employment data are flat three years in.78 Individual tasks, such as producing a first draft, changed almost immediately. Your own timeline is set by your industry's error tolerance, regulation, systems and management decisions.
What should I learn so AI does not replace me?
Three things, in order. How to use the tools well enough that your output is better and faster, and checked. How to redesign a piece of your own work around them rather than just chatting. And how to show the result, with a before and after. Our guide on staying valuable at work as AI gets better sets this out as specific behaviours rather than slogans.