How to use AI at work: a practical starting point.

Most advice about using AI at work is either breathless or useless: "10x your productivity" on one side, a list of 100 prompts you will never open again on the other. What actually works is smaller and duller. Pick one task you already do, learn to brief an assistant properly, and check what comes back.

This article is the starting point. It shows you how to choose the first task, how to write a brief that gets a usable result, what a real before-and-after looks like, and the checks that stop you forwarding something embarrassing. Everything here works with the free tier of ChatGPT, Copilot, Gemini or Claude - the tools will change, the method will not.

Where should you start? One task, not a transformation.

The people who get real value from AI at work did not "adopt AI". They found one recurring task where a decent first draft saves them time, got good at that, and only then added another. The right first task has four properties:

• It produces a first draft. Writing, summarising, tidying, reformatting - work where "80% done in two minutes" is genuinely useful.
• It recurs. Weekly status updates, meeting recaps, routine emails. A task you do once is not worth learning on.
• You can judge the output. You know what a good version looks like, so you can spot a bad one in seconds.
• It contains nothing confidential. No client names, salaries, unreleased figures or personal data. More on this below.

Good first candidates: turning meeting notes into minutes, drafting a reply to a long email thread, converting bullet points into a status update, writing a first-pass job advert, or explaining a spreadsheet formula someone else built. Bad first candidates: anything legal, anything with numbers you cannot verify, anything going straight to a customer.

How do you brief it? Like a colleague, not a search engine.

The word "prompt" makes this sound like a technical skill. It is not. A prompt is just a brief - the same thing you would give a capable temp on their first morning. Weak results almost always trace back to a weak brief, not a weak tool. A working brief has four parts:

The fourth part is the one most people skip: what good looks like. One sentence of standards changes everything - "keep it under 150 words", "do not promise dates", "match the tone of the example below". If you have a past example of the output done well, paste it in and say "match this style". That single move outperforms almost any clever prompt trick, and it is the core of the briefing patterns in our library of work prompts that hold up.

A worked example: notes into a status update.

Here is the whole method on one real task. You have just come out of a project catch-up with notes like this:

supplier delay on packaging - now w/c 11 Aug. Priya chasing. website copy signed off. budget - £2k over on photography, offset by print saving. Need decision from Dan on launch date by Fri.

The Friday status update usually takes 25 minutes of staring at a blank page. Instead, you give the assistant this brief:

What comes back is a clean, correctly structured update: packaging delay under Risks with the new date, sign-off and the budget position under Progress, the launch-date decision flagged to Dan with the Friday deadline. You will still edit it - perhaps the budget line reads too cheerful, and you want "cost neutral overall" rather than "great news". That edit takes two minutes. The task that took 25 minutes might become eight, and the structure is better than the one you wrote by hand, because you never bothered with a Decisions section before.

That is the honest scale of the win: not ten times faster, but, in our estimate, 15 to 20 minutes back on one recurring task, every week, with a result you are happier to send. Multiply that across the handful of tasks in where AI genuinely saves time in a working week and it becomes hours.

Check it before it goes anywhere.

AI assistants produce fluent, confident text that is sometimes wrong. The failure mode at work is not bad writing - it is plausible writing containing an error you did not catch. So every output gets the same three checks before another human sees it:

• Facts: verify every name, number, date and claim you did not supply yourself. If you cannot verify it, delete it.
• Voice: read it aloud. If a sentence is not something you would say, rewrite it - colleagues notice when your emails stop sounding like you.
• Gaps: ask what is missing. Assistants summarise what you gave them; they do not know about the thing you forgot to mention.

A useful rule: the assistant drafts, you decide. Anything that commits you to a date, a number or a promise is yours, not the tool's.

Know what never goes in the chat box.

Before the habit forms, set the boundary. Public AI tools are not a safe place for client data, personal information about colleagues, unreleased financials, passwords or anything covered by a confidentiality agreement. Most of the tasks above work fine with names swapped out and numbers rounded. The full rules - what is safe, what to anonymise, what to keep out entirely - are in using AI safely at work. Read that before you paste anything you would not put on a noticeboard.

How do you build the habit in one week?

Skill with AI at work is a habit, not knowledge. The one-week version: pick your task on Monday and write the brief once, properly. Use it every day it applies, keeping the brief in a notes file so you refine rather than retype. By Friday you will know your saved time per use, and the brief will be twice as good as Monday's. Then, and only then, pick a second task - email and meeting notes are the usual next candidates. The 30-day version of this plan, with a task menu per job family, is the backbone of our book AI at Work.

Frequently asked questions