Using AI assistants at work, properly

AI assistants — ChatGPT, Claude, Gemini, Copilot — are now normal office tools, like spreadsheets. The people who use them well are noticeably faster. The people who use them badly get caught.

Here's how to be in the first group.

What they actually are

That last part is the whole lesson.

The four-part prompt

Vague question in, vague answer out. Good prompts have four parts.

❌ Weak: "Write an email about the project."

✅ Strong:

Role: You are an experienced project manager.

Task: Write an email to a client explaining that delivery has slipped from 12 March to 26 March because their data arrived two weeks late.

Context: The client is important, generally reasonable, and dislikes long emails. We want to keep the relationship warm.

Format: Under 120 words. Professional but friendly. Offer one option to receive a partial delivery on 12 March.

Same tool. Utterly different output.

Five things it's genuinely great at

1. 1First drafts. A blank page is expensive. A bad draft you improve is cheap.
2. 2Explaining. "Explain what an amortisation schedule is like I'm 12."
3. 3Reformatting. Turning messy notes into a clean table, or a long document into bullets.
4. 4Brainstorming. "Give me 20 ideas, including some bad ones."
5. 5Rehearsal. "You're the interviewer for a junior accounts role. Ask me five questions, one at a time."

Four things it's genuinely bad at

1. 1Facts, numbers and dates. It will produce a completely fabricated statistic, with a completely fabricated source, in a confident tone.
2. 2Recent events. Its knowledge stops at a training cutoff date.
3. 3Your company's specifics. It does not know your policies, your customers or your systems.
4. 4Maths that matters. It's improved, but never trust an unverified calculation in a financial document.

The rules that keep you safe at work

1. Never paste confidential data. Customer records, salaries, unreleased financials, source code — assume anything you type may be stored and reviewed. Many companies have an approved internal tool for exactly this reason. Ask before you paste.

2. You own the output. If you send an AI-written email with a wrong number in it, that's your mistake, not the tool's. Nobody will accept "the AI said so".

3. Verify anything checkable. Every name, number, date, law and quote. If it cites a source, open the source.

4. Declare it when it matters. Using AI to polish your writing is fine and normal. Using it to write an assessment that's meant to measure you is not. When in doubt, ask.

5. Don't let it replace your thinking. Use it to go faster on things you understand. The moment you're pasting outputs you can't evaluate, you've stopped being the person in charge.

A workflow that works

1. 1You decide what's needed and why.
2. 2AI produces a draft, fast.
3. 3You check every fact and rewrite anything that doesn't sound like you.
4. 4You send it, and you own it.

Notice you're in three of the four steps. That's the correct ratio.