How to Use AI at Work: A Practical Starting Guide

Most advice on using AI at work is either too vague ("boost your productivity!") or too technical. This guide is neither. It's for managers and team members who want AI to take real hours off their week, starting this month.

It covers which tasks to start with, how to ask for output you can actually use, how to check it, and the few rules that keep you out of trouble.

Start with the tedious work, not the important work

The most common mistake is starting with the big strategic problem. Start with the boring stuff instead: the weekly report, the meeting notes, the email you send thirty times a month.

Good first tasks share three traits:

They happen often. Daily or weekly, so the time saved adds up.
You can check them quickly. You know what good looks like and can spot errors in a minute.
Mistakes are cheap. A clumsy internal summary is fine; a wrong figure in a customer invoice isn't.

Tasks that usually work well

Writing first drafts: emails, proposals, job ads, policy updates, replies to reviews.
Summarising: long email threads, meeting transcripts, reports, contracts (for a first read, not for the final legal view).
Restructuring: turning rough notes into a structured document, a table or a checklist.
Explaining: a spreadsheet formula, a clause, a technical term, in plain language.
Comparing: two supplier quotes, or three versions of a document, side by side.
Preparing: questions for an interview, an agenda for a meeting, objections a client might raise.
Translating and adjusting tone: a formal email made friendlier, or English text turned into clear Portuguese.

Tasks to hold back on

Anything final that goes to a customer, regulator or court without a person reading it.
Calculations you will rely on without checking.
Facts, figures and citations you haven't verified. AI tools can invent sources that look real.
Decisions about people, such as hiring, performance or pay.

How to ask for output you can use

The quality of what you get depends almost entirely on how you ask. "Write a follow-up email" gets you something generic. A good request reads like a brief to a capable new colleague on their first day.

The five parts of a good request

Context. Who you are, what the situation is, what has happened so far.
Task. Exactly what you want produced.
Audience. Who will read it and what they care about.
Format. Length, structure, tone. "Three short paragraphs, no bullet points, friendly but direct."
Example. Paste something you've written before that you liked.

Then add one line most people skip: "Before you start, ask me any questions you need answered." It often improves the result more than anything else, because the tool stops guessing.

A before-and-after example

Weak:

Write an email about the delayed project.

Strong:

I run operations at a 60-person logistics company. A client's warehouse system rollout is delayed by two weeks because their data export arrived late. Write an email to their operations director explaining the delay and the new date (14 November). Tone: transparent and confident, not apologetic. Under 150 words. End with one clear next step. Ask me questions first if anything is unclear.

Iterate instead of starting over

Treat the first answer as a draft. Say what to change: "shorter", "less formal", "move the new date to the first line", "you assumed X, but actually Y." Two or three rounds usually beat a single perfect prompt.

Save what works

When a request gives you good output, save it. A shared document of tested prompts per role is one of the most useful things a team can build. It turns one person's trick into everyone's habit.

How to check AI output

AI writes fluently and with confidence whether it's right or wrong. Checking is your job, and it's a skill.

Facts and numbers: check every one against the source. If the tool cites something, open it.
Omissions: summaries can drop the one detail that mattered. For important documents, ask: "What did you leave out?" Then read the original section yourself.
Tone and promises: make sure a draft doesn't commit you to something you didn't intend, like a discount, a deadline or a guarantee.
Fit: does it sound like you and your company? If not, give it an example and ask again.

Count the time you spend checking when you judge whether AI saves time. If a task takes ten minutes to generate and forty to fix, it's not a good AI task yet.

Five rules that keep you safe

Agree these rules before people start, ideally on one page:

Use company-approved tools and company accounts. Business plans generally come with stronger data-handling terms than personal free accounts. Check the current terms of the tool you choose.
Never paste sensitive data into unapproved tools. That means customer personal data, employee records, bank details, passwords, unreleased financials and anything under NDA.
A person reviews anything that leaves the building.
Be honest about AI use when a client or colleague asks.
Name one person to ask when someone isn't sure.

If you're in the EU, note that the AI Act requires organisations that deploy AI systems to take measures to ensure their staff have sufficient AI literacy (Article 4). A clear policy plus practical training is the sensible baseline.

A 30-day plan for a team

Week 1: agree the rules and pick the tasks

Publish the one-page rules. Ask each person to pick one tedious weekly task to try.

Week 2: try it on real work

Everyone uses AI on their chosen task at least three times. Note time taken, including checking.

Week 3: share and improve

Run a 45-minute session where people show what worked and what didn't. Collect the good prompts in a shared document. Show a few examples of AI output that was wrong, so everyone learns to spot it.

Week 4: make it the default

For each role, write down where AI is now the first step, for example "the weekly client update starts with an AI draft from my notes." Decide on the next task to tackle.

This kind of practical, task-based approach matches what the evidence suggests works. A UK government review of AI upskilling found that effective training is practical, tied to real tasks and built into everyday work, rather than generic courses. (GOV.UK, 2026)

When to go beyond the chat window

Once a task is reliably done with AI assistance and the steps are always the same, it may be a candidate for an AI agent. An agent does the work end to end, with a person approving where it matters. Answering routine calls and booking appointments, or chasing overdue invoices, are typical examples. See what our AI agents do.

Get the basics right first, though. An agent built on a process nobody understands will only make mistakes faster.

FAQ

What's the best way to start using AI at work?

Pick one repetitive, low-risk task you do every week, such as meeting notes or a standard email, and use an approved AI tool on it for a week. Check every output and track the time saved.

How do I write a good AI prompt?

Give context, the task, the audience, the format and an example, then ask the tool to ask you questions before it answers. Treat the first result as a draft and refine it.

Is it safe to put company data into AI tools?

Only into tools your company has approved, on company accounts, and never sensitive personal, financial or confidential data unless the tool and your policy allow it explicitly.

Will AI replace my job?

AI takes on tasks, not whole jobs, and the people who learn to use it well become more valuable. The practical move is to learn which parts of your work it can take on and which need your judgement.

Which AI tool should we use?

Whichever business-grade assistant fits the systems you already use, for example the one built into your office suite. The habits matter more than the brand.

See where AI fits in your week

Want to know which of your tasks suit AI, and how many hours it could give back? Our free AI Discovery assessment takes about ten minutes and gives you a snapshot of exactly that.

Take the free assessment →