How to Write an Internal AI Usage Policy (With a Template)
The short answer
A good internal AI policy fits on two pages and answers four questions clearly: (1) which data never goes into any AI tool — personal data, customer identity and financial details, signed contracts, whole source code, unannounced commercial information; (2) which tasks may be delegated to AI; (3) who is responsible for the output — always the human who approved it; (4) which tools are approved, and who approves them. Policies that describe the permitted space get followed; policies that ban everything only make usage invisible.
Why you need a policy — and why it's urgent
Let's start with a plain fact: you didn't start AI usage in your company. Your team already uses it — from their phones, on personal accounts, without telling you. When you don't write a policy, usage doesn't stop; it goes underground.
Invisible usage has two costs. First, data: a customer list containing personal data pasted into a free account can become a GDPR matter. Second, quality: when nobody knows how something was produced, there's no chain of responsibility once a flawed output reaches a client.
On our side this isn't theoretical. At Gurizon we build our clients' CRM systems, and with healthcare brands like ICEFUE and OFM Klinik there's patient data involved. In work like that there's no "let's just try it and see" — which data may go where has to be in writing before the project starts. Working in healthcare taught us that discipline early, and we carried it across to every client.
The four rules of a policy that works
Never exceed two pages
Nobody reads a ten-page policy, and an unread policy is no policy. Write it like a note to the team, not a legal document. Short sentences, concrete examples.
Describe the permitted space, not a list of bans
Policies that open with "the following is forbidden" make people assume everything not listed is fine — and push them into hiding. Saying instead "these three data types never leave, everything else is fine in approved tools" is both enforceable and builds an honest culture.
Attach responsibility to a name
This is the policy's most important sentence: responsibility for an output lies with the person who approved it. "The AI wrote it" is not a defence — not to the client, not legally. Once that's in writing, output quality rises on its own.
Keep an approved-tools list and update it
Which tools are approved, who approves them and how someone proposes a new one — all in writing. The approval criterion can be a single sentence: tools on a business plan that commit to not training on our data are approved.
A template you can copy
Adapt the text below with your own company name. I wrote it as internal communication, not as a legal document; if your sector is regulated (health, finance, legal), have counsel review it.
- PURPOSE — This document defines where and within what limits [Company] employees may use AI tools. Its aim is not to restrict use but to make it safe.
- DATA THAT NEVER GOES IN — Customer identity data (names together with ID/passport numbers, addresses, phone or email lists), health and financial data, the full text of signed contracts, entire source code, unannounced commercial information (pricing strategy, acquisition talks, new product plans) and employee records are never entered into any AI tool.
- PERMITTED USES — Drafting, editing and summarising text; research using publicly available information; ideation; coding assistance (within the limits above); translation; drafting decks and content; producing meeting notes (provided participants are informed).
- RESPONSIBILITY — Responsibility for any AI-assisted output lies with the employee who uses and approves it. Nothing going outside (to a client, supplier or the public) may be sent without an employee reading and approving it.
- APPROVED TOOLS — [tool list], used with company accounts, are approved. Personal or free accounts may not be used with company data. New tool proposals go to [owner]; the criterion is that the tool is on a business plan and commits to not training on our data.
- TRANSPARENCY — If a client asks whether AI was used in producing the work, we answer honestly. Where an output is entirely AI-generated and this affects the nature of what's delivered, the client is told.
- INCIDENTS — If data is entered by mistake, it is reported to [owner] the same day. No sanction applies to the person reporting; failing to report does.
- REVIEW — This policy is reviewed every six months. Questions and suggestions: [email].
Pay attention to item seven: stating that whoever reports a mistake won't be punished is the clause that decides whether the policy works at all. Someone who expects punishment hides the mistake — and a hidden data leak costs many times a reported one.
How should you roll it out?
- Don't just email it. Walk through it in a 30-minute meeting; the point is explaining what's restricted and why.
- Open with permissions, not restrictions. A presentation that starts with "here's what you can freely use" lands far better than one that starts with bans.
- Show a real example. Explaining why pasting a specific customer list is a problem beats three paragraphs of text.
- Assign the policy to a person. A policy without an owner goes stale within six months and nobody updates it.
When I have this conversation with my team I use one sentence: "My goal isn't to stop you using it, it's to protect you." That's true, because most of the risk sits on the employee: the person who pastes the wrong data is the one left most exposed. Present the policy as a safe area with clear edges rather than a restriction, and people actually follow it.
Key takeaways
- Your team uses AI whether or not you write a policy; the only difference is whether you can see it.
- Keep it under two pages — an unread policy is no policy.
- Describe the permitted space instead of listing bans; ban lists push people into hiding.
- The critical sentence: responsibility lies with the human who approved the output.
- The no-punishment-for-reporting clause decides whether the policy functions at all.
Frequently asked
I have a small team — do I really need a policy?
Even at three people, yes — but one page will do. The three that matter: which data never goes in, who owns the output, which accounts are used. In a small team a policy isn't bureaucracy; it's the single document that puts everyone on the same boundary.
How do I stop employees using personal accounts?
Don't block it — make it unnecessary by providing company accounts. People use personal accounts mostly because the company offered no alternative. At $20–30 per person per month, that's the cheapest way to remove the risk.
Do we have to tell clients we use AI?
There's no blanket legal requirement for most businesses, but check your confidentiality and subcontractor clauses. Commercially my advice is: be honest when asked, and volunteer it where it affects the nature of what you deliver. Anything discovered later costs more than anything disclosed up front.
Which use carries the most data-protection risk?
Pasting lists containing personal data (customers, candidates, patients) into a tool for analysis. It's usually done in good faith and "just this once". Second is processing meeting recordings without telling participants — informing them is a simple fix that most companies skip.