How to manage a team that uses AI

If that sounds familiar, you are in a common position. Some of your team may already be using AI tools at work. Some may not have tried them at all. A few may be using AI without realising it, because it has quietly appeared inside software they already use. You can't manage what you can't see, so the first job is to find out, openly and without blame.

This guide is for line managers, team leaders and small-business owners who are not AI specialists. It covers the decisions you need to make, a conversation to have with your team, and a plan for your first 30 days. Your organisation's policy and approved tools still take priority over anything here.

Start from what is actually happening

Don't assume nobody uses AI, and don't assume everybody does. Ask.

The way you ask matters more than the question. If the first thing people hear is "who's been using ChatGPT?", the honest ones will feel accused and the rest will go quiet. If the first thing they hear is "I want us to use these tools well, so I need to know what's useful and what worries you", you will get far more truth.

Priya's approach was simple. At the weekly operations meeting she said: "AI tools are here and some of us will be using them. I'm not looking for anyone to blame. I want to know what's working, so we can agree some sensible rules together." Then she went first, admitting she had used Copilot Chat to tidy a supplier letter.

The eight things a manager has to settle

You don't need a long policy. You do need a clear answer to each of these, written down somewhere your team can find it. The simple team AI rules guide has a one-page template that covers all eight.

1. Which tools are approved

Find out which AI tools your organisation has approved, and on which accounts. The same brand name can cover very different products. A personal ChatGPT account and a company ChatGPT Business workspace are treated differently by OpenAI. Microsoft 365 Copilot Chat signed in with a work account runs under the organisation's Microsoft 365 terms, with enterprise data protection. A personal Gemini account and Gemini under a Google Workspace work account may be under different terms.

If nothing is approved yet, say so plainly and tell the team who is deciding and roughly when. Ask Ravi, or whoever runs IT in your organisation, what is already available. Some teams already have access to a work AI tool without realising it.

2. Where the data boundaries sit

This is the decision with the most risk attached. Your team needs to know which information can go into which tool. A useful starting split:

• Public or general information (published material, generic wording help): usually fine in any approved tool.
• Internal business information (plans, procedures, figures): approved work tools only.
• Customer information and personal data about staff, customers or the public: only where policy allows, in approved work tools, with the minimum needed.
• Sensitive personal data such as health, ethnicity, religion, trade union membership, criminal records or safeguarding information: treat as off limits unless your written policy explicitly says otherwise.
• Passwords, access codes, bank or card details: never, in any AI tool.

Removing a name does not always make information anonymous. A complaint about "the only account manager in the Midlands" still identifies Alex. The Can I put this in ChatGPT? tool walks through these decisions one piece of information at a time, and it is a good thing to share with your team.

3. Who is responsible for the result

The person who sends, submits or acts on AI-assisted work is responsible for it, exactly as if they had written it themselves. "The AI wrote it" is never an explanation for a wrong figure in a customer quote. As a manager, you stay responsible for the decisions you make, including any that were informed by AI output.

4. How work gets checked

AI tools can state wrong information confidently. The industry term is hallucination. Agree what checking looks like for your team, scaled to the risk. A draft internal note needs a read-through. A customer quote needs every figure checked against the price list. A summary of a contract needs every date and notice period checked against the clause itself. The guide on reviewing AI-assisted work gives you a checklist.

5. When AI use should be disclosed or recorded

Decide when people should say that AI was used. A sensible default: tell the reader when a document was substantially drafted with AI and the reader would reasonably expect to know, and always flag it when handing work to someone else to check. Some organisations also want a short note on file for certain work, such as which tool was used for a report that informs a decision. Whether and how you record this is a matter for your organisation's policy.

The point of disclosure is to help the checker. A reviewer who knows a summary came from AI will check it differently.

6. What to do when something goes wrong or is unclear

People need a named person to ask when they're unsure, and a clear instruction for when something has gone wrong: if information went into a tool it shouldn't have, tell the named person straight away. Speed matters more than blame, because your organisation may need to assess it quickly. The shadow AI guide covers what to do in that situation.

7. Training

People can't follow rules for a tool they don't understand. Give the team time to learn: a short session, a shared set of good prompts, or time to work through a free course. The free course on this site includes a lesson on where AI answers go wrong, and The AI-Safe Desk covers what can and can't go into a tool.

8. A feedback loop

Rules can go out of date within months. Tools change, licences change, and your team will find uses you didn't expect. Put a short standing item on your team meeting: what worked, what went wrong, what we should change. Keep a simple list of good prompts and near misses. Review the rules at a set interval.

A team conversation guide

This is the conversation Priya used. It takes 30 to 45 minutes. Run it as a discussion, not a presentation, and don't ask people to name who did what. Take notes without names.

Opening (5 minutes)

Say why you are having the conversation and what it is not. For example: "This isn't an audit and nobody is in trouble. I want us to agree how we use AI, so it helps and doesn't cause problems." If you have used AI yourself, say so.

Current use (10 to 15 minutes)

• Which AI tools have you come across at work, including features inside other software?
• What have you tried them for? What was useful?
• What did you try that didn't work, or took longer than doing it yourself?
• Has anyone been unsure whether something was OK to paste in?

Worries and hopes (10 minutes)

• What worries you about AI in our team? (Jobs, being monitored, mistakes, customers noticing, extra work.)
• What would you most like help with?
• Is there anything you'd want to stay fully human, whatever the tools can do?

Boundaries (10 minutes)

• What kinds of information do we handle that should never go into an AI tool?
• Which of our work goes to customers or informs decisions, and so needs careful checking?
• Who should people ask when they're not sure?

Close (5 minutes)

Tell the team what happens next and when. For example: "I'll write up interim rules from this by Friday, share them for comment, and we'll look again in a month." Then do it.

If you want to turn anonymised meeting notes into a first draft of findings, an approved work AI tool can help. Keep names and customer details out of the notes before you paste them.

Below are my notes from a team discussion about how we use AI tools at work. The notes contain no names. Using only these notes, group what was said under four headings: current uses, what worked, worries, and requests. Under current uses, mark any that involve a personal account or an unknown tool. Then list up to five questions I should get answered before writing team rules. Do not add anything that is not in the notes. If something is unclear, say so instead of guessing.

[paste your anonymised notes]

Your first 30 days

This plan assumes you are a manager without an AI specialist in your team, in an organisation that may or may not have a policy. Adjust the timings to fit.

Mistakes managers commonly make

Pretending it isn't happening. Silence doesn't stop people using AI. It stops them asking you about it.

Leading with a ban and no alternative. If people have found a tool that saves them an hour a week and the only answer is "no", some will carry on quietly. Give an approved route, even a limited one.

Treating every use as equally risky. Asking AI to suggest a clearer subject line is different from pasting in a customer's complaint history. Rules that treat them the same get ignored.

Over-trusting the output. It is easy to relax once a tool has produced a few good drafts. The error that matters is usually the one that looks right: a date one day out, a figure from the wrong column, a policy clause that doesn't exist.

Making the checking invisible. If you only praise speed, people will skip checks to look fast. Thank people for catching errors.

Letting AI make decisions about people. Hiring, discipline, performance ratings and similar decisions need a human decision-maker who can explain the reasons. The task suitability guide explains why.

Where the law comes in

UK data protection law applies to personal data whatever tool is used. The ICO publishes guidance on AI and data protection for organisations. For government teams, the GOV.UK AI Playbook sets out principles for safe use. This guide is not legal advice. Your organisation's policy, data protection officer or legal adviser has the final say on what is allowed.