How to write an AI usage policy your team will actually follow

Someone in your company has probably already pasted a client contract into a free chatbot to get a quick summary. Carelessness is rarely the reason. They had a deadline, the tool was one tab away, and nobody had told them where the line was.

Companies usually react in one of two ways. Some ban AI tools, and people keep using them on their phones. Others say nothing, and every employee draws their own line. Neither protects your clients’ data, and neither helps people use AI well.

A short AI usage policy fixes this. We don’t mean a 30-page document written for lawyers. We mean two pages that answer the questions people actually ask: which tools can I use, what can I put into them, do I need to check the result, and who do I ask when I’m not sure. This post covers what to put in it, gives you an outline to copy, and shows how to roll it out so it doesn’t end up in a folder nobody opens.

Why a ban and silence both fail

A ban feels safe, but it only moves AI use out of sight. Someone who has found that a chatbot saves them an hour a day won’t stop. They’ll switch to a personal account on a personal device, where you have no contract with the provider and no idea what went in.

Silence has the same effect with less guilt. The careful employee avoids AI completely. The enthusiastic one pastes in payroll data. Both believe they’re doing the right thing.

There is also a legal reason to write things down. Since February 2025, the EU AI Act has expected companies that use AI to take steps so their staff understand it well enough to use it responsibly. This is called the AI literacy duty. GDPR applies as soon as personal data goes into a tool. A clear policy plus a short training session is the simplest way to show you took both seriously. We cover the legal side in AI, GDPR and the EU AI Act.

What the policy should cover

Eight sections are enough for most companies. Each one should fit in a few sentences or a short table.

The eight sections of a usable AI policy

• Approved tools Which tools, which plan, company login only
• What data can go in Green, amber and red, with examples
• Checking outputs Who is responsible for what AI produces
• Telling customers When you say that AI was involved
• Confidentiality and IP Client contracts, secrets, other people’s work
• Who to ask One named person, not “IT”
• New tools and uses How to propose them and get an answer
• Incidents What to report, to whom, how fast

1. Approved tools

List the tools people may use and, just as important, which plan. The same assistant can be fine on a business plan and a problem on a free personal account. Business and API plans from the large providers usually come with a data processing agreement and terms saying your data isn’t used to train their models. Free and personal plans often have weaker terms, and some use conversations to improve the model unless you switch that off.

Name each tool precisely: the assistant on your company’s business plan, the AI features built into your helpdesk or office suite, and any private model you run for sensitive data. If you only allow one tool, say so. A list of three approved tools is more useful than a list of twenty banned ones.

2. What data can go in

People will read this section more than any other, so make it concrete. A three-colour system works well. Green data can go into any approved tool. Amber data only goes into approved business tools with a signed data processing agreement. Red data never goes into an external AI tool.

Green, amber, red: what can go into an AI tool

	Green	Amber	Red
Examples	Public website text, published price lists, your own drafts without names	Internal emails, client correspondence, non-public figures, supplier contracts	Health data, national ID numbers, bank details, payroll, passwords, data a client contract excludes
Approved business tools	Yes	Yes	No
Private model you control	Yes	Yes	Partly
Who decides	You, no need to ask	You, within the rules	The policy owner, case by case

Make the examples match your company. In an accounting firm, a client’s draft accounts are amber, and anything containing national ID numbers or bank account details is red. In a clinic, patient data is red for any external tool. It belongs only in a private setup that your data protection officer has reviewed, such as a model running on your own servers (see self-hosted LLMs and private AI). In a recruitment agency, CVs are personal data. They can be amber in an approved tool for parsing and summarising, but the decision about a candidate stays with a person.

Give people a way to turn amber into green. Removing names, replacing a client with “Client A” and cutting the parts the task doesn’t need often makes a document safe to use.

3. Checking what AI produces

One rule covers most of it: the person who uses the output owns it. AI drafts, a person sends. If a wrong figure reaches a client, “the AI wrote it” is not an answer.

Be specific about what must be checked. Names, numbers, dates, quotes from documents and anything that sounds like a legal or tax statement must be checked against the source. Anything that goes to a client or gets published is read in full. AI does not make final decisions about people: hiring, credit, discipline. For processes where AI runs at volume, you need a designed review step instead of a general rule. We cover that in human in the loop.

4. Telling customers and partners

Decide when you tell people that AI was involved. Some of this is set by law. From August 2026, the EU AI Act requires that people are told when they are talking to an AI system, such as a chatbot on your website, unless it’s obvious. Beyond the legal minimum, pick a position and write it down. For example: “We don’t label routine emails drafted with AI help, as long as a person has checked and sent them. Chatbots and automated messages always say they are automated.”

Also check your client contracts. Some forbid using AI tools on the client’s data, or require you to say when you do.

5. Confidentiality and intellectual property

Treat an AI tool like any other outside supplier. If a client contract or NDA says their documents can’t be shared with third parties, that includes AI tools unless the contract says otherwise. Passwords, API keys and access tokens never go into a prompt.

Two points on intellectual property. Don’t ask AI to rewrite someone else’s article, image or code and then publish it as your own. And the rules on who owns purely AI-generated content are not settled and differ between countries. For work you deliver to clients as yours, make sure a person shapes and edits it, and check what your tool’s terms say about outputs.

6. Who to ask

Name one person and a backup. “Ask IT” or “ask legal” sends people on a hunt, and most will guess instead. Promise an answer within a few working days.

7. Proposing a new tool or a new use

People will find new tools every month. Give them a short, predictable route, otherwise they’ll just start using them.

How a new AI use gets approved

1. 01 Someone proposes it What for, what data, which tool
2. 02 Owner checks Data colour, terms, data processing agreement
3. 03 Small trial Two or three people, real but safe data
4. 04 Added to the list Policy and prompt library updated

8. Incidents

Say what counts as an incident: red data entered into an unapproved tool, an AI mistake that reached a client, or a tool showing data it shouldn’t. Ask people to report within 24 hours to the policy owner. If personal data is involved, the owner decides whether it is a data breach under GDPR, which may need to be reported to the data protection authority within 72 hours.

A template you can copy

Fill in the brackets, cut what doesn’t apply, and keep it to two pages.

[Company name] AI usage policy
Version 1.0 · Owner: [name] · Backup: [name] · Next review: [date]

Purpose: we want everyone to use AI where it helps, without putting
client, employee or company data at risk. This applies to all employees
and contractors.

1. Approved tools
   - [Assistant], [business plan], company login only
   - AI features in [helpdesk / CRM / office suite]
   - [Private model on our servers]: for amber and approved red data
   Personal accounts are not used for company data.

2. What data can go in
   Green (any approved tool): public information, own drafts without names
   Amber (approved business tools only): internal documents, client
   correspondence, non-public figures
   Red (never in external tools): health data, national ID numbers,
   bank and payroll data, passwords and keys, anything a client
   contract excludes
   Not sure? Treat it as one colour higher and ask [owner].

3. Checking outputs
   You are responsible for anything you send, publish or decide with AI.
   Check names, numbers, dates, quotes and legal statements against the
   source. AI does not make final decisions about people.

4. Telling others
   Chatbots and automated messages say they are automated.
   [Our position on AI-assisted emails and documents.]
   We follow client contracts that restrict AI use.

5. Confidentiality and IP
   Client confidentiality applies to AI tools like any other supplier.
   No passwords, keys or tokens in prompts.
   We do not publish AI output that copies someone else’s work.

6. Who to ask
   [Owner], [channel or email]. Answer within [3] working days.

7. New tools and new uses
   Send [owner]: what for, what data, which tool.
   Safe uses start with a short trial by two or three people.

8. Incidents
   Report within 24 hours to [owner]: red data in an unapproved tool,
   an AI error that reached a client, a tool showing data it should not.
   Nobody is penalised for reporting.

Training: [date of session]. Prompt library: [link].

How to roll it out

A policy nobody has heard of protects nobody. Four weeks is enough for a company of 20–200 people.

Rolling out the policy in four weeks

1. 1

   Week 1

   Draft it with the people who use AI

   The owner, three or four heavy users and someone from finance or HR. Start from the template.

2. 2

   Week 2

   Test it on real cases

   Take ten things people actually did with AI last month. Does the policy give a clear answer for each?

3. 3

   Week 3

   Set up the approved tools

   Business accounts, a signed data processing agreement, training on your data switched off.

4. 4

   Week 4

   Thirty minutes per team

   Walk through green, amber and red using examples from their own work.

5. 5

   Every 6 months

   Review it

   New tools, incidents, and the questions people asked the owner.

The order matters. If you announce the policy before the approved tools are ready, you’ve banned the old way without offering a new one.

Mistakes we see

• Writing it for lawyers. If a sales rep can’t tell in ten seconds whether they may paste a customer email, the policy failed.
• Banning without an alternative. Every “no” should come with “use this instead”.
• Rules nobody can apply. “Use AI responsibly” is a wish, not a rule. “Never put national ID numbers into an external tool” is a rule.
• No owner. Questions go unanswered, so people stop asking.
• Never updated. Tools and their terms change every few months. A policy from two years ago is probably wrong about at least one of them.

What to do next

1. Make a list of the AI tools people in your company already use. Ask, don’t guess.
2. Pick the approved tools and set up business accounts.
3. Fill in the template above with your own green, amber and red examples.
4. Test it on ten real cases, then run a 30-minute session per team.

A policy tells people where the lines are. It doesn’t teach them to use AI well. For that, see getting your team to actually use AI. If your rules mean some data can’t leave your own environment, we can talk through a private setup.