Answer

How do you write an AI policy for your company?

Justas Butkus is a fractional AI officer based in Vilnius, Lithuania, working with mid-market companies and scale-ups across the UK, EU and US at two to four days a month. He builds and operates the production AI systems he advises on, and takes no commission, referral fee or revenue share from any vendor.

Why most AI policies fail

Picture the week after a fifteen page AI policy goes out. It was copied from a template written for a large regulated enterprise, it was sent round by email with a polite request to read and acknowledge it, and nobody did, because it opens with a definition of artificial intelligence. Meanwhile the team carries on using exactly the tools it used before. The only thing that has changed is that they no longer mention it.

That is the part worth sitting with, because the policy has made things strictly worse. The risk is the same as it was, and your visibility of it has gone. People who feel the rules are not meant for them stop telling you what they use, and a document that was supposed to bring AI into the open has pushed it quietly underground.

So the test of a policy is not whether it is complete. It is whether a new employee can read it in five minutes and know what they may do. If you cannot pass that test on a Monday morning with a coffee in your hand, the policy is not finished, however thorough it looks in the shared folder.

The six questions your policy must answer

A policy that works is really six answers written in plain words, and the useful ones are the ones that name something. Read them as the questions an employee would otherwise be asking you in the corridor.

The employee, question one

Which tools am I allowed to use?

A named list: "approved tools are X, Y and Z". A line such as "tools must be appropriate to the task" answers nothing, so skip it.

The employee, question two

What must I never paste into one?

Your own list, specific to your business: customer personal data, unreleased financials, credentials, anything under a client confidentiality clause. Vague categories get interpreted generously.

The employee, question three

When does a person have to check what it wrote?

Named situations: anything sent to a customer, anything affecting pay or employment, anything going into a legal or financial document.

The employee, question four

How do I get a new tool approved?

A named person and a realistic timeframe. If approval is slow or unclear, staff route around it and you are back to invisible risk.

The employee, question five

What if something goes wrong?

Who to tell and how fast, plus an explicit line that reporting a mistake made with an approved tool is not a disciplinary matter. Without it you will not hear about failures.

The employee, question six

Whose document is this, and is it still current?

A name and a date. Unowned policies rot within months in a field moving this fast.

Notice that none of those answers needs a lawyer to read it. That is deliberate, and it is most of the craft.

What to leave out

The hard part of a short policy is not what you write but what you resist writing, because every missing paragraph feels like a gap somebody will point at in a meeting. Four things are worth cutting anyway.

  • Definitions of artificial intelligenceNobody needs them and they date immediately.
  • Ethical principles as standalone statements"We will use AI responsibly" constrains nothing. Put the constraint in the rules instead.
  • Technology-specific instructionsName tool categories and approval routes, not model versions.
  • Anything you will not enforceAn unenforced rule teaches staff that the whole document is decorative.

How this relates to the EU AI Act

There is a quiet industry built on the idea that the EU AI Act obliges you to have an AI policy, and it is worth being precise here because the confusion is being sold commercially. The Act asks for something narrower and, in some ways, more sensible.

  • Article 4 requires measures supporting AI literacy among staff and contractors operating AI systems on your behalf, as replaced by the Digital Omnibus with effect from 27 July 2026. It is a duty of effort. There is no mandated curriculum, examination or certificate.
  • The Act does not require an AI policy document, and it does not require appointing an AI officer. Unlike GDPR with its Data Protection Officer, no role or governance structure is mandated.
  • A written policy is still the cheapest evidence that you took proportionate measures, which is what you will be asked to demonstrate in a procurement questionnaire or an insurer's renewal form long before a regulator asks.

A two-page structure that works

Put the six answers in this order and you have the whole document. It is short enough that people will read it, and each heading is a question someone will actually ask.

  1. 01

    Purpose

    Three sentences. Why this exists and who it applies to, including contractors.

  2. 02

    Approved tools

    A named list, with the approval route for anything not on it.

  3. 03

    Never put this into an AI tool

    A specific list drawn from your own business.

  4. 04

    Human review required

    The named situations where a person checks the output first.

  5. 05

    If something goes wrong

    Who to tell, and the explicit no-blame statement.

  6. 06

    Owner and review date

    A name and a date.

Anything beyond that belongs in a separate governance document, read by a different audience. Keep it away from the two pages your staff will actually open.

So go back to that Monday morning. If a new colleague started tomorrow, could they read your policy before their first coffee goes cold and know what they may do? If not, what is the one line you would cut first?

Short answer

A usable AI policy is short, specific and answers six questions: what staff may use, what they may never put into a tool, when a human must check output, who approves new tools, what to do when something goes wrong, and who owns the policy. Anything longer gets ignored, and an ignored policy is worse than none.

If you need this written rather than explained

It is a bounded piece of work: a couple of pages, specific to your tools and your data, with a named owner.

Frequently asked questions

What should an AI policy for employees include?

Six things: which tools are approved, what must never be entered into any tool, when a human must review output, how new tools get approved, what to do when something goes wrong, and who owns the policy. Roughly two pages.

Does the EU AI Act require an AI policy?

No. Article 4 requires measures supporting AI literacy among people operating AI systems on your behalf, and the Act mandates no policy document and no AI officer role. A written policy is still the cheapest way to evidence proportionate measures.

How long should an AI policy be?

About two pages. The test is whether a new employee can read it in five minutes and know what they may do. Longer policies get ignored, and an ignored policy leaves the risk unchanged while removing your visibility of it.

Should the policy name specific AI tools?

Yes, for the approved list, with a clear route to get something added. Principles without a named list get interpreted generously, and staff route around approval processes that are slow or unclear.

Does an AI policy apply to contractors?

It should. The EU AI Act obligations extend to other persons operating AI systems on your behalf, which includes contractors and service providers, so your policy should say so explicitly.

Who should own the AI policy?

A named person with a review date, not a function or a committee. In a field moving this fast, an unowned policy is out of date within months.