How to Rewrite a Job Description for AI: Tasks First, Not the Title
You’ve decided a role on your team needs rethinking. Good — almost nobody does. Now you’re staring at the existing job description, which is two years old and describes work that has partly stopped existing, and you have no idea where to start.
Here’s the mistake almost everyone makes at this point: trying to redesign the job. Don’t. Redesign the tasks, and let the job description fall out of that.
Start from the number that makes this manageable
Google’s ATLAS study mapped nearly 15 million real AI interactions onto 800-plus occupations and 4,000 tasks. AI showed up somewhere in 68% of those occupations — but in the median occupation it touched just 21% of the tasks.
That second number is the useful one. You are not rewriting a job. You’re rewriting about a fifth of one. Everything else in the role is unchanged, and pretending otherwise is what turns a two-hour exercise into a project nobody finishes.
Step 1: Inventory the tasks, not the responsibilities
Get the real list. Not the job description’s list — the actual one.
The fastest way is to ask the person doing the job to log a normal week: every discrete thing they did, with rough time against each. A week is enough. You’ll get somewhere between fifteen and forty lines, and the A flat list of what a job actually consists of, at the level of “reconcile the vendor spreadsheet” rather than “manage vendor relationships.” Job descriptions are written at the second level, which is exactly why they’re useless for this — you can’t ask whether AI does “manage relationships.” is the whole basis for what follows.
Two rules. Log what they did, not what’s supposed to happen. And don’t let anyone tidy it — the eleven minutes a day spent reformatting someone else’s spreadsheet is precisely the kind of line that matters here.
Step 2: Sort every line into three piles
Go through the list once, marking each task:
| Pile | Status | Definition |
|---|---|---|
| 1 | AI drafts it | The first pass is now a machine’s job — summaries, formulas, first versions. |
| 2 | AI can’t touch it | Judgment, relationships, physical presence, context. The spine of the new role. |
| 3 | AI made it pointless | Tasks that only existed to feed a step that is now automated. Delete outright. |
Pile 3 is the rarest and the most valuable — the hand-formatting, the re-keying, the status update a dashboard now answers. Pile 1 is where the actual work of this exercise is, and where step 3 comes in.
Step 3: Add the checking time back
This is the step that gets skipped, and skipping it is why redesigned roles don’t survive contact with reality.
When a task moves to pile 1, the time doesn’t vanish — it shrinks and changes shape. Workday surveyed 3,200 leaders and employees and found that 37% of the time saved by AI is lost again to Correcting, clarifying or rewriting weak AI output. Workday’s framing is a “tax on productivity” — roughly four hours going back out for every ten gained. . Nearly four hours back out of every ten saved.
So if you subtract eight hours of drafting from a role, you do not get eight hours back. You get about five, and you have created three hours of a genuinely different task: checking. That task belongs on the new description by name. A role that absorbs AI output without anyone accountable for verifying it isn’t a redesigned role — it’s the same role with a new failure mode.
Step 4: Rewrite the description around what’s left
Now, and only now, open the document. The rewrite is nearly mechanical:
- Delete everything from pile 3.
- Demote pile 1 — these become “produces X” one-liners, not the body of the job.
- Promote pile 2 to the top. This is what the role is now for.
- Add the verification work from step 3, explicitly.
Concretely, for a Junior Frontend Developer:
Before
- Translate Figma mockups into boilerplate HTML/CSS.
- Write initial React components for new features.
- Draft standard unit tests.
- Fix minor UI layout bugs logged by QA.
After
- Own the structural integrity of new UI features: review and correct AI-generated React components before they hit production, and be the person who catches the hallucinated logic or accessibility failure.
- Decide which layout bugs require fundamental rewrites, and take them to Design with a recommendation.
- Maintain the strict codebase standards that determine whether the generated code is actually maintainable at scale.
Same person, same title, roughly the same hours. But the second version can be hired against, reviewed against, and argued with — and it describes a job that will still exist in two years.
Step 5: Set a date to check
Write down what you expect to be true in ninety days: which tasks should have disappeared, roughly how much time should have moved, and what “the checking is happening properly” would look like. Then actually look, on that date.
Without this, you can’t tell a redesign that worked from one that quietly reverted — and reverting is the normal outcome. The description goes back in the drawer and the old shape of the job reasserts itself within a quarter.
The takeaway
Role redesign sounds like an HR programme. It’s a two-hour exercise with a list, three piles, and a document — done for one role at a time, by the person who manages it.
The two things that make the difference: work from the real task list rather than the job description, and put the checking time back in. Almost every redesign that fails, fails at one of those two, and the second one is the one nobody sees coming.
Sources: Google’s AI & Economy ATLAS v1.0 for the 68% of occupations / 21% of tasks figures (14,653,926 de-identified interactions sampled 6–19 April 2026, mapped to 800+ occupations and 4,000 tasks). The 37% rework figure is from Workday’s Beyond Productivity: AI Value report with Hanover Research (3,200 respondents, half leaders and half employees, fielded November 2025), reported by CFO.com — worth noting Workday sells AI-featured HR software, so it has an interest in how organisations think about this. The five-step procedure, the three piles, and the worked example are mine, not theirs.
Related: AI Role Redesign: Every Job Is Changing, Almost No One Is Rewriting Them is why this is worth doing at all — and the research on how few organisations bother. How Much of Your Job Can AI Actually Do? About a Fifth is where the 21% comes from and what it means from the employee’s side of the desk.