Evaluate a specific task with real review costs. An occupation’s exposure score cannot tell you whether an AI workflow will work in your workplace.
To understand how AI might change your work, list your recurring tasks and examine one workflow at a time. Look at its inputs, required quality, permissions and consequences of mistakes. Test a small permitted use against your normal process, counting review and repair time as well as generation time. This is a planning exercise, not a prediction of job security.
The ILO’s 2025 occupational exposure research examines tasks and potential technological overlap. Exposure is not the same thing as a recorded redundancy or an individual forecast. Our task audit translates that distinction into questions a worker and manager can discuss together.
Map the work that actually fills your week
Use a representative week rather than your job title. Break “customer support” into activities such as reading a request, finding the applicable policy, drafting a reply, approving an exception and recording the outcome. They have different inputs and different kinds of responsibility.
For each activity, record how often it happens, roughly how long it takes and how you recognise a good result. Include coordination and follow-up. If a draft is fast but creates extra clarification work for colleagues, measuring only the drafting step will give an incomplete picture.
Set the boundaries before choosing a tool
Ask which information may be used, which tools are approved and who can authorise any external action. Suggesting a response and sending it to a customer are separate permissions. The same distinction applies to recommending a payment and executing one.
NIST’s generative AI profile offers voluntary risk-management guidance. The practical checklist below is our synthesis, not a NIST assessment or certification. Start with a low-consequence task whose result you already know how to evaluate.
Use a task worksheet
| Task | Possible assistance | Required check |
|---|---|---|
| Draft a routine reply | Propose wording from approved material | Check facts, tone and applicable policy before sending |
| Summarise meeting notes | Group recorded points | Check attribution and avoid inventing decisions |
| Handle an unusual complaint | List questions to investigate | Keep the responsible person’s judgment and approval |
These are hypothetical examples. A workplace may prohibit a use that appears simple. Add a row for data restrictions and a named reviewer before trying anything. If the proposed reviewer cannot recognise a wrong result, choosing a different trial is more useful than adding a checkbox labelled “human review.”
Compare the whole workflow
Choose comparable examples and define an acceptable result before starting. Record setup, prompting, generation, checking and correction separately. In an illustrative trial, a task taking 30 minutes manually might take 8 minutes to draft with AI plus 17 minutes to check and repair. The saving is 5 minutes, not 22. These numbers illustrate accounting, not a measured productivity result.
Record quality problems as well as time. Did the output omit a necessary condition? Did it add a promise no one approved? A serious error may matter more than a small average saving. Keep failed examples in the record so the discussion is based on the full trial rather than its best demonstration.
Decide what to change and who benefits
Use the findings to choose whether to continue, redesign or stop the workflow. If it helps, agree who maintains the source material and what triggers rechecking. Changes to tools, policies or task complexity can invalidate an earlier result.
Discuss the effect on people too. Does the trial free time for difficult cases, add review pressure, or remove practice that a junior worker needs? These are questions for the people doing and managing the work, not answers a model’s benchmark can supply.
Continue with tasks versus job titles and who benefits from AI productivity. Use the audit as a concrete starting point for that conversation, with your assumptions and unresolved questions visible.