On this page
Your team has very probably started using AI tools already. Without asking you, each with their own account and their own approach. That isn’t bad news; the bad news is that there’s no written rule for it.
An AI usage policy is a short internal document that sets out which tasks a team may use these tools for, with which data and under which checks. Its purpose isn’t to stop people using them, but to make that use safe and consistent.
What happens in a team without a policy?
Without a written AI usage policy, everyone applies their own interpretation, and those interpretations can be far apart. When team rules for AI stay verbal, the rule changes from person to person.
One person pastes a client contract in full and asks for a summary. Another never opens the tool because they assume it “might not be allowed”. A third drops a figure the model produced into a report without checking it. None of them means any harm; all three are guessing the answer to a question nobody has answered for them.
The real risk here isn’t a technical vulnerability; it’s inconsistency. Having overly cautious and overly relaxed behaviour side by side in the same team means both lost productivity and risk.
The policy shouldn’t run past one page
Most AI usage policies I’ve seen don’t get read, because they’re fifteen pages of legal language. A policy nobody reads is the same as no policy.
The goal should be this: a new starter can read it in five minutes and know what to do. That takes three things: concrete examples, clear boundaries and short sentences. Being thorough and being long are not the same thing.
Seven sections every policy needs
A useful AI usage policy answers the following seven questions. The order can change, but none should be missing.
1. Which tools are approved?
List the tools the team may use by name. Also state who to ask if someone wants to use a tool that isn’t on the list. Without this, shadow use is inevitable.
2. Which data can be entered, and which can’t?
This is the heart of the policy. Rather than abstract wording, name the types of data. The table in the next section was prepared for exactly this. If you’re considering a setup that works with your own data, the data preparation steps in this article rely on the same classification.
3. Which tasks can it be used for?
Drafting, summarising, generating ideas and reformatting are usually fair game. Performance reviews, screening decisions in hiring and final text that goes straight to a client need separate care.
4. How is output checked?
State in writing which output must always be checked by a person. Any output containing figures, dates, names, quotations or legal wording falls into this group. These are the four areas where language models are most prone to being confidently wrong.
5. Transparency with clients and the outside world
Decide how AI-assisted output is presented to clients. In some sectors disclosure is an obligation; in others it’s a choice. Either is fine, but it shouldn’t be left unclear.
6. Who is responsible?
A one-line clause, and the most important one: responsibility for the output lies with the person who uses it. “The model wrote it” is not a justification. When this sentence is explicitly in the policy, the habit of checking takes root by itself.
7. Who do you ask?
Name the person to go to when in doubt. Without a name nobody asks, everyone guesses, and the policy’s purpose is lost.
Data classification table
On the AI data security side, what helps most isn’t long explanations but a short classification. The three-category structure below is a starting point most businesses can use as it is. You’ll need to adapt the examples to your sector, but the category logic usually holds.
| Category | Example data | Rule |
|---|---|---|
| Allowed | General industry information, public content, draft text you wrote yourself | Can be used freely |
| Careful | Internal process descriptions, project details without names, anonymised examples | Can be used once identifying details are removed |
| Prohibited | Client personal data, contract text, financial records, access credentials, health data | Never entered into any tool outside an approved business environment |
The phrase “approved business environment” in the third row is deliberate. The business versions of some tools commit not to use submitted data for model training and come with a separate contract. Which version of which tool you use directly affects this table, so the policy should name the version.
A one-page policy template
The template below is the seven sections above on a single page. Keep the headings and fill in the content for your situation.
- Purpose: This document helps our team use AI tools safely and consistently. It aims to clarify use, not restrict it.
- Approved tools: [tool names and versions]. Before using a tool that isn’t listed, speak to [name].
- Data rule: The allowed, careful and prohibited categories are defined in the table below. If in doubt, treat the data as prohibited.
- Use cases: Drafts, summaries, ideas and reformatting are allowed. For [listed tasks], get approval first.
- Checking rule: Any output containing figures, dates, names, quotations or legal wording is verified by a person before it’s published.
- Transparency: [Situations in which clients are informed].
- Responsibility: Responsibility for the output lies with the person who uses it.
- Questions: Whenever you’re unsure, speak to [name]. Asking is always the right answer.
- Review date: [date]
This template gets read because every line states an action. Sentences that describe principles are easy to write but don’t get applied; sentences that say “in this situation, do this” do.
Three mistakes to avoid
- Prohibitive language. A policy that opens with “use is forbidden” doesn’t stop use; it just makes it invisible. Shadow use is unaudited use, and far riskier.
- Writing without tool names. Generic phrases like “AI tools” mean everyone understands something different. Write the names, and update them when the list changes.
- Publishing it once and forgetting it. Announcing the policy and shelving it makes it obsolete within three months. Tools and capabilities change fast.
Keeping the policy alive
An AI usage policy takes effect not on the day it’s published but on the day it starts being applied. The written document is half the job; the other half is the team actually being able to follow it.
The routine I’ve seen work in practice is this: after publishing, the policy is walked through in a short session using examples, the team tries two examples from their own work, and a month later the questions that come up are gathered and added to the policy. The third step is skipped in most places, but it creates the most value, because real ambiguities only appear through use.
Set a review date every six months and put it at the top of the document. An internal document without a date never tells anyone it’s out of date.
Frequently asked questions
We’re a small team. Do we really need a policy?
Even a five-person team needs an AI usage policy, because the issue isn’t team size but how sensitive the data is. If you work with client data, one person getting it wrong once is enough. In a small team the policy is shorter, but it should exist.
Who should write the policy?
The person who knows the work should write the draft, and a decision-maker should approve it. Policies written only by legal tend to be too strict to follow; those written only by the technical side tend to underestimate the risk. Having both read it gives the best result.
Can employees use their personal accounts?
This is the most common question about employee use of AI, and the answer depends on your sector. If you allow personal accounts, you need to spell out in much more detail which data can’t go into them. If you don’t, banning them without offering an approved alternative doesn’t work.
Is this policy a legal requirement?
It depends on your sector and the regulations you’re subject to. If you have data protection obligations, controlling which data employees transfer where is already your responsibility, and this policy is the practical tool for that control. For a definitive answer, ask your legal adviser; this article isn’t legal advice.
Setting it up with your team
Let’s draft the policy with your team
In corporate AI training programmes we draft the AI usage policy together with the team, using your own real examples. You’ll find how the programme works on the service page.
Updated:


