ChatGPT, Claude, Copilot or Gemini · Practical walkthrough

Know when not to use AI for the task in front of you

Once a tool is useful for eight things, it starts getting pointed at the ninth without anyone deciding to. The ninth is usually the one with a named person in it, a number that has to be exactly right, or a reader who assumed you wrote it.

Task by Task15 min read + practice

What you'll make

a one-line go, part or no-go decision for a task you’re about to start, and a short list of the tasks in your own work where the answer is a standing no.

What you'll need

a task you were about to start, five minutes, and a willingness to arrive at “do this by hand”. A screen you never fail isn’t a screen.

Example result.

TASK: Q3 stock variance (£4,100 between system and count)
DECISION: AI for part of it
WHY: must be exactly right. Reconcile in the spreadsheet; AI helps with the lookup formula and drafts the board note once the figure is confirmed.
DATE: Monday 12 October

These tools are good at drafting from material you give them, reshaping text you already have, summarising things you can check, and being an undemanding audience while you work out what you think. They’re a poor choice for exact arithmetic, binding wording, decisions about individuals, anything you can’t verify, and messages whose whole value is that a person wrote them. The difficulty is that the ninth task doesn’t announce itself.

1. Check you’re allowed to use it at all

Before usefulness, permission. Is the account you’d use approved for work, and is this kind of information allowed in it? Being able to sign in isn’t permission to use company information. If you don’t know, don’t paste the material: Work out what you are allowed to put into an AI tool at work shows how to find out.

Why this works: a task that fails on permission doesn’t need the rest of the screen.

What you should see: a named account and a yes for this material, or a question sent to the person who decides.

2. Ask the four questions

Thirty seconds, before you open anything.

Could you tell if the answer were wrong? Could you actually check it, with a source, in reasonable time? If not, and it will be acted on, stop. An unverifiable answer you pass on is a claim you’ve made. “Ask the assistant again” isn’t a check.
Is the value in the document or in the thinking? If a colleague asked what you think about the supplier, a balanced two-page assessment answers a question nobody asked.
Does it have to be exactly right, or roughly right? A first-pass summary or a list of things to consider can be roughly right. A reconciliation can’t.
Would the reader feel misled if they knew? Not whether it’s against policy: whether they assumed something about how it was made, and would mind being wrong.

Why this works: each question catches a different way the ninth task goes wrong: unverifiable, beside the point, imprecise, or quietly dishonest.

What you should see: a clear pass, or the one question the task fails. If everything on your list fails, the problem probably isn’t the list.

3. Check it against the standing no’s

These are worth treating as no’s at full scope:

Decisions about a named person. Shortlisting, performance scoring, redundancy selection, credit decisions. The ICO’s guidance says human involvement has to be active, not a token gesture. Glancing at a ranking an assistant produced is the token gesture. That guidance is under review after the Data (Use and Access) Act, so ask your data protection lead before relying on it.
Binding wording. Contract clauses, terms, privacy notices, safety instructions. A plausible draft that’s subtly wrong gets reviewed as a document rather than questioned.
Regulated advice passed on as yours. Medical, legal, tax, financial. You become the person who gave the advice.
Exact arithmetic in a chat window. Reconciliations, VAT, payroll totals, stock counts. Use the spreadsheet or the system that owns the number.
Facts you can’t verify. Regulations, statistics, case references, quotations. Confident answers can contain wrong facts and invented references.
Messages whose whole value is that a person wrote them. Condolence, an apology for your own mistake, a thank-you, a reference. A fluent condolence note is worse than a clumsy one.
The decision you’re avoiding. “Should we restructure the team?” produces a tidy answer with considerations on both sides. Reading it feels like progress. It’s forty minutes not making the decision.
Work whose purpose is your own learning. If you’re meant to form a view, having it assembled for you removes the only output that mattered.

Why this works: a concrete list catches the task before habit does.

What you should see: your task either clear of the list, or matched to one line.

4. Narrow it, then write the one line

Several no’s become reasonable at a narrower scope. The move is the same each time: take the assistant off the judgement and put it on the material around it.

“Rank these applications” is a no. “List the questions I should ask HR before this process starts” is fine.
“Work out whether the variance is real” is a no. “Here’s the confirmed figure from the system; help me write the explanation for the board” is a yes.

I want help with part of this task but must keep [the judgement] myself. Task: [description, without sensitive details]. Suggest two or three narrower things you could do with material I already have, such as organising, extracting or drafting from confirmed facts. For each, say what I'd need to check. Don't make or recommend [the judgement].

[the judgement] is the decision that stays with you; [description, without sensitive details] describes the task in general terms.

Then record the decision in one line: task, decision (AI / AI for part of it / by hand), why, date. You’ll face the same task next month and won’t remember the reasoning.

Why this works: the narrowed task is checkable, and the judgement stays visibly with a person.

What you should see: a narrower task with a check named, or a firm “by hand”.

Check the result

The decision names which question or standing no it turned on.
Any “part of it” says exactly which part the assistant does.
The judgement you kept is written down as yours.
Permission was settled before anything was pasted.

The usual failure: the narrowed task creeps back into the judgement. You asked for wording around a confirmed fact, and the draft adds a conclusion you didn’t reach: a cause, a recommendation, a verdict. It reads naturally, which is why it gets through. Read the output for any claim that wasn’t in what you gave it.

Your draft includes [the added conclusion], which I haven't established. Remove it. Use only the facts I gave you, and where the reason or decision is still open, say that it's still being investigated.

[the added conclusion] is the line that went beyond your facts.

Worked example: four tasks on a Monday

Fictional example. An operations lead opens their list on Monday 12 October.

1. Monthly supplier renewals summary, from contracts in the company workspace
2. 23 applications for a warehouse supervisor role
3. Q3 stock variance: £4,100 between the system and the count
4. A note to a colleague whose father has died

Renewals: AI. Checkable against the contracts, roughly right is fine for a watchlist, and the dates are verified before anyone acts.
Applications: by hand. A decision about named people, and a glance at a ranking isn’t active human involvement. The lead uses AI only for interview questions drawn from the job description.
Variance: AI for part of it. The reconciliation happens in the spreadsheet. Once the figure was confirmed, the lead asked for a board explanation. The draft said: “The variance is most likely due to damaged stock written off informally.” Nobody had established that. The repair prompt with “the cause, damaged stock written off informally” gave: “The count is £4,100 below the system figure. The cause is still being investigated.”
Condolence note: by hand. Would the reader feel misled? Yes, and that’s the end of the analysis. The effort is the message.

Two yeses, one partial, one firm no, in a couple of minutes. Without the screen, the second task would have cost an afternoon of unwinding.

Task kit for this job: Download (ZIP) · Reusable instructions

Go deeper

Optional training: The uses people refuse that they shouldn't

Caution has its own failure. Reshaping text you wrote is low-risk and useful: notes into an update, a long document into a one-page brief. The facts came from you and you can see whether they survived. Summarising something in front of you is fine, because you can check it. Talking a problem through with something that asks obvious questions is an old technique that works. Formatting, first-draft structures and writing up a process you described out loud are all sensible. The common thread: you supplied the material and you can verify the output.

Optional training: Write the team version

A personal screen scales badly; a team’s list of standing no’s scales well. Keep it to one page, and make every entry a real task from your work. “We don’t use AI to assess supplier credit risk” is usable. “Exercise appropriate caution with sensitive data” has never changed anyone’s behaviour. Next to each no, put the narrowed version that’s allowed and who owns the list. Revisit it when the tool, the data or the rules change.

Optional training: Reassess after the first attempt

A task that passed the screen can still turn out unhelpful. If the output needs heavy repair, ask why: missing information, an awkward input format, or a job that’s quicker done directly. For a recurring task, run a two-week pilot and count the checking. Stopping an unsuitable workflow is a useful result.