AI may remove parts of your job, change the order of the work or raise the standard expected of you. That is not the same as proving that your whole job will disappear.
The useful question is not “Will AI steal my job?” It is “Which tasks in my job are exposed, and what remains valuable when those tasks get cheaper?”
What the evidence actually says
The ILO and NASK global index, published in May 2025, assessed task exposure across occupations. It found that one in four jobs worldwide had some potential exposure to generative AI, with a higher share in high-income countries.
That figure describes exposure, not confirmed redundancies. The ILO says transformation is more likely than full replacement because most jobs combine tasks that vary in technical feasibility, cost, responsibility and need for human involvement.
That distinction matters. An assistant might draft a report, but it does not automatically own the decision, persuade a sceptical stakeholder, notice an unsafe assumption or accept liability for the outcome.
Audit your job as a collection of tasks
Write down the recurring tasks from a normal fortnight. Do not start with your job title. Titles hide the real work.
For each task, record:
- Input: what information, request or event starts it?
- Output: what must exist when it is finished?
- Rules: how much of the decision can be written down?
- Context: what private, local or relationship knowledge changes the answer?
- Consequence: what happens if the output is wrong?
- Ownership: who reviews, approves and carries the risk?
Then place it in one of four buckets.
| Task type | Sensible treatment |
|---|---|
| Repetitive and low consequence | Candidate for automation |
| Draftable but judgement-heavy | AI drafts, human reviews |
| Sensitive or high consequence | Tight controls and named approval |
| Relationship, negotiation or accountability work | Human-led, perhaps AI-assisted |
This is an operating model, not a prediction. A real employer may automate less because the data is poor, integration is expensive or the risk is unacceptable. It may automate more if the work is already structured and measured.
Build evidence instead of collecting prompts
“Prompt engineering” is not a durable career moat on its own. Product interfaces and models keep changing.
A stronger approach is to show that you can improve a real task safely:
- establish a baseline using real examples;
- define what a good result looks like;
- test the tool on ordinary, awkward and failure cases;
- record the time spent checking and correcting its output;
- protect confidential or personal data;
- keep a human owner for consequential decisions.
The result may be faster. It may also reveal that the tool moves effort from drafting to checking. Both findings are useful.
What to strengthen
The safest bet is not to become “irreplaceable”. No one can promise that.
Strengthen the work that sits around the model:
- choosing the right problem;
- obtaining clean, authorised context;
- spotting contradictions and missing evidence;
- handling exceptions;
- explaining decisions;
- earning trust;
- accepting responsibility for the final outcome.
These are not mystical human advantages. They are observable parts of work that employers and clients can assess.
Boundaries checked on 24 August 2026
This article does not predict whether a particular role, employer or sector will cut jobs. Exposure studies estimate technical potential, while actual outcomes also depend on demand, investment, regulation, organisational choices and worker voice.
Do not put employer, client, health or other sensitive data into an AI service until its contract, account type, retention settings and approved-use policy have been checked. Product licences and privacy terms vary and change.
For a practical way to inspect how AI changes your own work, join the Microsoft Copilot Adopters Space.