The short version

  • Choose one small task instead of trying to predict your whole job.
  • Write down the information, steps and checks it needs, and who approves the result.
  • Learn the part you cannot yet explain or evaluate.

Start with a smaller question

‘Can AI do my job?’ is a difficult place to start. Try: ‘Can it help with this one task, and how would I know if the result was right?’ A task might be answering an inquiry. The full process also includes finding the right information, checking the answer and resolving the person’s problem.

The ILO’s exposure index assesses task capabilities and finds clerical work particularly exposed. Its April 2026 methods review warns that exposure alone cannot predict displacement, productivity or reskilling needs. Economic and organizational conditions affect what happens next.[1][3]

Write down five things

  • The purpose: who needs this result, and what would help them?
  • The information: which records or sources does the task depend on?
  • The steps: what happens from the first request to the finished result?
  • The check: how would you catch an error or an unusual case?
  • The decision: who can approve the result, and what happens if it is wrong?

Use the gap to choose what to learn

If you cannot check the answer, learn more about the subject or ask a qualified person. If you understand the task but struggle with a repetitive step, try a small example using public or fictional information. At work, first check which tools and information are permitted.

Checking and coordination may become easier to automate too. The aim is not to find a permanently human-only step. It is to learn something useful now, keep evidence of your work and notice when the requirements change.

Draw the task. Name the check. Assign the decision. Follow the outcome.

A benchmark is not a working day

METR’s time-horizon benchmark estimates the human-expert task duration at which a model reaches a specified success rate, such as 50% or 80%. It primarily covers well-specified technical tasks. The duration is not how long the agent runs autonomously, and long-task estimates are uncertain.[2]

Improving benchmarks are a reason to retest assumptions, not a conversion table from model scores to redundancies. A task taking an expert four hours does not mean a model with a four-hour horizon can reliably replace half a working day.

An example, not a reported outcome

A customer refund request

Before
Treating ‘write a response’ as the complete task.
Small trial
Separate finding the order, checking policy, drafting the response and authorizing a refund.
What to check
The authorized person verifies the records and exception rules. Afterward, track whether the customer’s problem was actually resolved.
Adapt this to your situation

One thing to do now

Write the steps of one familiar task on paper. Circle the step you find hardest to explain or check. Use it to choose your next lesson or question to a colleague.

Make it fit my situation
Sources & limits

Read the evidence.

The practical steps are our editorial recommendations. Numbered citations link research claims to their sources. Expand a source for its findings, methods and limits.

  1. Generative AI and Jobs: A Refined Global Index of Occupational ExposureInternational Labour Organization · Working Paper 140 · May 20, 2025

    Finding. About 25% of employment is in occupations with some exposure; 3.3% is in the highest band.

    Limit. Exposure describes potential task change, not observed adoption or job losses. The highest band is included in the 25%.

    Type
    Modeled exposure
    Period covered
    2025 task-capability assessment
    Scope & method
    Global employment; occupational tasks assessed using worker input, experts and AI, then linked to employment data.
    Where to look
    Executive summary; global employment by exposure gradient. DOI: 10.54394/HETP0387.
    Read the original source ↗ (new tab)
  2. Task-Completion Time Horizons of Frontier AI ModelsMETR · Updated May 8, 2026

    Finding. Reports human-expert task duration at specified model success rates, including 50% and 80%.

    Limit. Not autonomous runtime or whole-job competence. Long-duration estimates, especially beyond 16 hours, are uncertain.

    Type
    Capability benchmark
    Period covered
    TH1.1 estimates; page snapshot reviewed September 2026
    Scope & method
    Primarily software, machine-learning and cybersecurity tasks with measurable completion criteria.
    Where to look
    Definitions, reliability selector and limitations below the chart.
    Read the original source ↗ (new tab)
  3. Workers’ exposure to AI: What indicators tell us—and what they don’tInternational Labour Organization · April 17, 2026

    Finding. Capability indicators alone cannot establish displacement, productivity or reskilling needs.

    Limit. Methodological guidance, not a new estimate of employment change.

    Type
    Methods review
    Period covered
    2026 review of exposure indicators
    Scope & method
    Review of what task-based AI exposure measures can establish.
    Where to look
    Discussion of indicators, adoption constraints and labor-market outcomes.
    Read the original source ↗ (new tab)

Reviewed September 5, 2026. Publication dates and periods studied can differ. Exposure is not job loss; projections are not outcomes. Reading time covers the main briefing; learning resources, source notes and linked research are additional.

Browse the research library