Where AI earns its place at work, and the jobs to keep it away from

Background

Most advice about using AI at work is a list of things it can do. It can do nearly all of them. The better question is which jobs are worth handing over, because the same tool that turns a messy email thread into a tidy action list will also invent a VAT total with a straight face.

If a job smells automatable, AI is one option. Often the better one is a formula, a template or fixing the process so the job doesn't exist. I'd start there, and use AI for the jobs where the value is in the words and the reading, with a person checking the result.

This guide is the test I'd run before handing anything over. For prompts by job, there's a map of the tips and guides. For which tool to open, there's a first taste of ChatGPT, Claude and Copilot.

Setup

Five questions before you hand a job over

1. Is the output a draft that a person will read before it goes anywhere? If the answer is no, stop. Nothing here should reach a customer, a patient or a payroll run without a person looking first.
2. Can you check the answer in a minute? Reading a summary against the original is quick. Re-doing a reconciliation to see whether it was right isn't, and then you haven't saved anything.
3. Can you paste the input without breaking a rule? If it's personal, clinical or confidential, use a tool your organisation has approved for that data, or take the names out first.
4. Is the job mostly words and reading, not arithmetic? Summaries, rewrites, sorting and spotting what's missing suit a model. Totals, counts and exact matching belong in the spreadsheet.
5. Does the job repeat? A task you do every week earns a saved prompt. A one-off might be quicker to do yourself.

Four yes answers is a good candidate. Fewer than that, think again.

Jobs that tend to pass

Each of these is a draft you can check against something real:

• Meeting notes in a fixed shape. You were in the meeting, so you'll spot what's wrong. See the shape worth copying.
• A shared inbox turned into a worklist. The triage prompt asks for an owner and a deadline per thread, which is more use than a summary.
• A runbook from what one person knows. Interview them, then let the model structure it: a runbook from tribal knowledge.
• Renewal dates out of a pile of contracts. It quotes the clause, you check the PDF: contract renewal dates.
• Ticket triage. Agree the categories first, then let it sort: categories first.
• A handover note before leave. The questions it asks you matter more than the draft.

Jobs to keep it away from, or hold at arm's length

• Anything where it originates a number. Do the sums in the spreadsheet and let the model write around them. Never let it invent a number.
• Decisions about a person. A review, a reference, a disciplinary note. It can help structure notes after the names are out. A person makes the call.
• A process that's broken. If nobody can say how the job works, a model will describe it fluently and wrongly. Fix the process, or write down how it really runs first. A small project without the ceremony is often enough.
• Anything you couldn't tell was wrong. If you'd have no way to spot a mistake, the answer can't be used.

A first week

Pick one job that passed the test. Do it the usual way and with the tool, side by side, for a week. Ask it to list its assumptions each time. At the end of the week, keep the prompt if it saved time and you trusted the result. If it didn't, you've lost a week's side-by-side.