Will AI take
your job?

Exposure is not a layoff forecast.

AI can affect a task without taking over the entire job. A role is a bundle: repeatable work sits beside exceptions, coordination, tacit knowledge, judgement, and responsibility. An AI system may draft or classify something while a person still decides whether it is correct, handles the difficult cases, and answers for the result.

That distinction does not make displacement imaginary. Employers can reorganise work, reduce hiring, or cut roles. But a capability score alone cannot tell you whether this will happen in your team. That depends on reliability, cost, workflow redesign, demand, regulation, bargaining power, and the choices an organisation makes about the gains.

Count task exposure. Do not rename it job loss.

The International Labour Organization’s 2025 global index combines task-level data, worker input, expert assessment, and model predictions. It estimates that roughly one in four workers are in an occupation with some exposure to generative AI; 3.3% of global employment is in the index’s highest exposure category. The authors say transformation of jobs is more likely than full replacement because most occupations include tasks that still require human input.

The 2026 ILO research brief makes the limit explicit: exposure measures describe technological susceptibility. They do not predict displacement, productivity gains, or the need for retraining. They cannot fully model adoption costs, institutional barriers, changing workflows, or what happens when lower costs create more demand.

These figures are global estimates across occupational categories. They do not tell an individual worker what will happen in a particular company, country, or role.

Ask what the study actually measured.

The OECD’s 2026 synthesis describes several effects at once: automation of existing tasks, new tasks and occupations, and productivity changes. It reports evidence of AI complementing human work, alongside displacement risks in routine and repetitive roles. It also cautions that high exposure does not automatically mean high automation risk; specialised cognitive work can be highly exposed while still depending on complex judgement and social skills.

A national or global index helps locate questions worth investigating. It cannot replace evidence from actual workflows and employment decisions.

Make the job visible task by task.

Write down the recurring work in a normal week. Describe outcomes, not vague labels like “communication” or “analysis.” Then assess each task with the same questions:

1. How often does it happen? Note approximate time, volume, and the amount of variation between cases.
2. What information does it need? Identify missing context, tacit knowledge, sensitive records, and whether the system is permitted to access them.
3. What counts as correct? Specify the quality bar, consequences of a mistake, and whether a result can be independently checked.
4. Where does a person decide? Mark exceptions, negotiation, care, physical action, accountability, and moments where the process needs a human to say no.
5. What happens after the task changes? Ask whether demand rises, review moves elsewhere, the task disappears, or people take on different work.

Use three provisional buckets: automate with checks for bounded, repeatable, low-consequence steps; assist and review when a system can prepare useful work but a person needs to verify it; and keep human-led when responsibility, uncertain evidence, or the consequences of error call for direct human judgement. These are workflow choices to test, not permanent labels for occupations.

Watch the organisation, not the prediction market.

For a specific role, look for changes that can be observed and compared over time:

One announcement or a short-term headcount change rarely explains the cause on its own. Compare the timing with demand, budgets, reorganisation, outsourcing, and other tools. Ask workers doing the work; they often see hidden review and exception handling before dashboards do.

Where an organisation is piloting AI, establish a baseline first: task completion, errors, time spent reviewing, customer or service quality, and the distribution of workload. Recheck after the workflow changes. Adoption counts and prompt volume do not establish that a role became more productive or less necessary.

Build judgement around a real task.

Choose one recurring task, try an approved tool on a low-risk example, and keep a record of what it gets right, what it misses, and how long verification takes. Learn the boundaries of the system you actually use. Bring the results to the people who own the workflow and ask who benefits if the work gets faster.

That approach is more useful than trying to acquire a vague collection of “future-proof” skills. It builds transferable habits: understand the work, inspect evidence, recognise failure, and make decisions about systems rather than trusting a demo.

For a role-by-role learning plan, see AI literacy at work ↗. To compare AI-assisted tasks fairly, read how to measure the complete workflow ↗.