Most "AI-proof jobs" lists are written by people who have never had to defend a hiring decision. They rank job titles, when automation has never really worked at the title level. It comes for tasks, and every job is just a bundle of them.
Why the job-title framing fails
A radiologist and a paralegal both have work a model can produce a decent first pass on. Neither profession has vanished. What changed inside them is the mix: fewer hours spent making a first draft, more hours spent checking, deciding, and being answerable for what goes out the door.
If you plan a career around a title someone put on a list, you are betting on a category. If you plan around tasks, you are betting on what you actually do all day, which happens to be the thing you can change.
The tasks that are genuinely exposed
There is a rough pattern to what gets automated first, and it has little to do with difficulty. Work is exposed when it is:
- repetitive, with the same shape every time
- text in, text out, or data in, data out
- well specified, so that "done correctly" can be defined in advance
- reviewable by someone else quickly and cheaply
Summarising a fifty-page report is exposed. Drafting the standard clauses of a contract is exposed. Producing the monthly variance commentary from a spreadsheet is exposed. First-line ticket triage is exposed. None of that means the person doing it disappears. It means those hours shrink, and the rest of the job has to carry the salary.
The tasks that resist
The stubborn parts are less flattering than the "creativity and empathy" answer everyone gives.
Physical presence in messy environments. Site work, care work, field repair, on-the-ground logistics. The world there is neither tidy nor instrumented, and the situation keeps changing while you are standing in it.
Accountability and sign-off. Somebody has to be liable. A model can draft an audit conclusion, but it cannot be the name at the bottom of the page and it cannot be struck off a register. Wherever a signature carries legal or financial weight, the signing is the job.
Negotiation and trust built over time. Renewing a contract with a client who is annoyed at you is not an information problem. It is a relationship with a history and a balance of power.
Judgment under ambiguity. Deciding what to do when the data is incomplete, the stakeholders disagree, and there is no correct answer, only a defensible one.
Three of those four are about consequences rather than intelligence. That distinction is what your CV has to make visible.
Audit your own role in an afternoon
Take your last two weeks. Not your job description, your calendar and your sent folder.
- List every recurring task in plain language, with a rough share of your week.
- Score each one against the four exposure criteria above.
- Mark the tasks where a mistake of yours lands on you.
- Add up the share of your week sitting in the exposed column.
The number tends to be higher than people expect and less alarming than it sounds, because the exposed hours are usually the ones nobody enjoys anyway.
Rewrite the CV around the half that survives
Most CVs are built back to front. They describe output volume, which is exactly the part that just got cheap. Lead each bullet with the decision, the ownership or the outcome, and let the volume sit at the end as supporting evidence.
Before: Produced monthly financial reports for four business units.
After: Owned the monthly close for four business units; caught a pricing error that had gone unnoticed for two quarters and recovered 40k in under-billed revenue.
Before: Wrote product documentation for the API.
After: Decided what went into the API docs after support tickets showed users stalling on authentication, cutting auth-related tickets by a third.
Before: Managed supplier relationships.
After: Renegotiated three supplier contracts during a price increase, holding two at the previous rate and moving the third to a shorter term to keep leverage for next year.
The pattern is the same each time. A verb of judgment first (owned, decided, renegotiated), then the friction that made it hard, then a number attached to the result instead of the activity. "Wrote 60 documents" measures activity. "Cut auth tickets by a third" measures an effect.
Show that you use AI, without the empty line
In plenty of roles this is now an expected signal rather than a differentiator, which is precisely why the generic version is worthless. "Proficient in AI tools" tells a recruiter nothing at all. What lands is the specific workflow you changed and what it bought you.
Something like: "Built a review workflow using an LLM to pre-screen 200 monthly supplier invoices for mismatches, freeing around six hours a week for exception handling." That describes someone who moved their own hours toward the non-automatable half of the job, which is the trait the hiring manager is actually hunting for.
Keep it honest. If you reuse a set of prompts for competitive research, say so and say what came of it. If you tried a tool and dropped it, that is a solid interview answer, not a confession.
One thing to do this week
Open your CV and count the bullets starting with a verb of production (wrote, produced, managed, supported, prepared) against those starting with a verb of judgment (owned, decided, prioritised, negotiated, recovered). If production is winning, you have your afternoon's work. Rewrite the top three bullets of your most recent role so the first six words carry a decision you made, and push the number to the end of the line.
If you are rebuilding from your LinkedIn profile rather than a blank page, Postulit carries the structure across so your time goes into the wording instead of the formatting.