Will AI replace my job? Start with a task audit, not a prediction

What the current evidence can tell you

The ILO's June 2026 review separates emerging productivity evidence from employment outcomes. It describes uneven gains and limited large-scale displacement in the evidence reviewed, alongside concerns about younger workers and job quality. That is a dated snapshot of a changing situation, not proof that a particular job is safe or doomed.

Exposure means a technology could affect a task. Adoption means somebody uses it. Useful automation means the whole process improves after checking, coordination, and exceptions. Those are different stages. A demonstration of a good first draft does not establish that a business can remove the person who checks the draft, understands the customer, and owns the consequences. Equally, needing a human check does not mean the task's economics cannot change.

Make a five-column task audit

List the recurring tasks you completed last week. For each one, record time spent, the output somebody needed, the cost of an error, how the result is checked, and the next person or system it depends on. Keep the units mundane: hours, handoffs, corrections, and waiting. A task called 'marketing' is too broad; 'turn interview notes into a draft case study' is something you can test.

Mark which tasks are easy to reverse. Rewriting an internal draft is different from sending a customer a contractual commitment. Start with a reversible task where you already know how to recognise a good result. Otherwise an impressive output may simply conceal mistakes you cannot evaluate. The audit is not a score of your personal worth. It is a map of where tools, judgment, and coordination meet.

Find the bottleneck before buying speed

Suppose drafting is quick but approval takes days. Faster drafting can create a larger queue without improving delivery. Try Amdahl's law: make one part of the process dramatically faster while leaving the rest unchanged. Notice how the total improvement remains limited by work that the tool does not accelerate.

Now look back at your audit. Ask whether the limiting step is execution, unclear requirements, access to data, review capacity, or permission to act. A useful AI skill may be less about producing more text and more about preparing better inputs or detecting errors sooner. Measure an end-to-end outcome, such as an accepted deliverable, rather than counting drafts generated.

A worked example: a customer-support team

Imagine a support agent spending an illustrative ten hours each week drafting replies and another ten investigating account histories, checking policies, and resolving unusual cases. These numbers are invented for teaching; they are not measurements of a real team. Halving drafting time would save five hours before verification and setup costs. It would not halve the entire workload.

1. Choose one low-stakes reply type. Define an acceptable answer using the team's existing policy and have the normal reviewer check it.
2. Compare several similar cases with and without the tool. Record total time, corrections, and whether the customer actually got a resolution.
3. Review the results before expanding. If checking consumes the savings, improve the workflow or choose another task instead of hiding the cost.

Choose a skill bet with a small downside

The opportunity cost experiment makes an overlooked point concrete: time spent learning one tool is time unavailable for something else. Compare a narrow tool tutorial with a transferable skill such as writing clear requirements, evaluating evidence, or explaining tradeoffs. A transferable skill can remain useful even when a particular product changes.

Use optionality to think about the shape of the decision. A short project that produces a portfolio example, a reusable checklist, and feedback from a real user creates several future paths. A costly commitment justified only by a single forecast leaves fewer ways to adapt. This is a planning principle, not a claim that every small experiment succeeds.

Run a two-week learning experiment

Set one question: can I produce a better checked result for this task? Choose a fixed time budget and a stopping date. Save examples of failures as carefully as successes. Write down what you expected before the experiment, so you cannot quietly redefine success after seeing the output.

Explore the multi-armed bandit to see why endlessly trying new options and prematurely committing to one option can both waste information. Your career is not that simulation: opportunities change, feedback is noisy, and relationships matter. The useful lesson is to make room for exploration while still finishing work that teaches you something.

• Pick one reversible task, not an entire occupation.
• Measure checked output and total effort.
• Keep a record of failures and exceptions.
• Decide when to continue, change, or stop.

What this approach does not settle

An individual audit cannot predict hiring decisions, bargaining power, or a company's strategy. A tool that helps you today may also change entry-level training opportunities tomorrow. Discuss changes with colleagues and managers rather than treating all workplace uncertainty as a personal optimisation problem.

The World Economic Forum's 2025 skills outlook is useful context about employer expectations, but expectations are not realised outcomes. Combine broad reports with direct evidence from your own work. The practical question is not 'Can I predict the future perfectly?' It is 'What can I learn this month that leaves me better able to respond?'