An AI policy is a short agreement on how AI may be used in the company: which tools are approved, which information is never entered, who owns the questions, and when a person reads the result before it is used. It fits on one page. Start with the template here, adapt it to your business and have a lawyer read it before you share it.
Why you need a policy now
Employees already use AI tools, often on their own initiative. Without a shared agreement nobody knows what is allowed, which information may be entered or whom to ask when something is unclear.
A policy gives security to the people who want to use AI and protection to the company. Keep it short. A ten page policy is read by nobody, and a one page policy is read by everyone.
What an AI policy should contain
- Purpose. What you want to achieve with AI and who the policy applies to.
- Approved tools. Which services may be used for work, and how to ask for a new one.
- Information that is never entered. For example personal data on customers and employees and trade secrets, unless the tool is approved for that specific task.
- Review. Who reads the result before it is sent, published or changes anything.
- Responsibility. Who owns the policy and whom to ask.
- Mistakes and incidents. How to say that something has gone wrong, and that doing so is welcome.
- Knowledge. How new colleagues learn the rules. The EU AI Act sets requirements for AI literacy among staff who work with AI. Ask a lawyer what that means for you.
A one-page template
Copy the template, replace what is in square brackets and delete what does not fit. Read it together with someone who knows the law before you share it.
AI policy template
AI policy for [Company name] Version [date]. Owner: [name and role]. 1. Purpose We use AI to [purpose, for example save time on administration]. This policy applies to everyone who works for [Company name]. 2. Approved tools [List of tools]. A new tool is approved by [role] before it is used for work. 3. Information that is never entered Personal data on customers and employees, [category] and [category], unless the tool is approved for that specific task. 4. Review Everything AI has written or suggested is read by a person before it is sent to a customer, published or changes anything in our systems. 5. Agents and permissions An AI agent gets access to what the task requires and nothing more. It makes no changes without approval from [role]. 6. Questions and mistakes Questions go to [role]. If something has gone wrong, say so straight away, so that we can correct it and learn from it. 7. Review of the policy The policy is read again every [interval] by [role].
A checklist before the first tool
- Task. Which recurring task should AI help with?
- Owner. Who in your company owns the result?
- Information. Which information does the tool need to see, and does the policy allow it?
- Agreement. Does the tool process personal data, and if so, is there a data processing agreement?
- Stop point. Where does a person read the result before anything happens?
- Follow-up. When will you look at whether it works and what needs to change?
How to keep the policy alive
Read the policy again after a couple of months and collect the questions colleagues have asked in the meantime. The questions show where the text is unclear. Add what is missing and delete what nobody has needed.
Let one person own the policy and make it easy to find. A policy in a folder nobody opens works as if it did not exist.
Common mistakes
Start simple and improve. A short policy that is used is better than a complete one that stays in a folder.
- Too long. A policy nobody has the energy to read is not followed. Delete everything that does not decide anything.
- Only bans. A policy that says no to everything creates hidden workarounds. Say what is fine and how to ask for more.
- No owner. Without a person who answers questions the text goes stale.
- Written by a few. Ask those who use AI every day what is unclear. They know where the rules are needed.
- Never read again. Tools change. Set a date for the next review.
Three rules that tend to work
- Read everything before it goes out. What AI has written is read by a person before a customer sees it.
- Ask before you connect. A tool that will be connected to your systems is first approved by the owner of the policy.
- Say so when something goes wrong. Whoever finds a mistake says so straight away, so that everyone can learn from it.
How we work with this
With us, every assignment starts with a mapping of how the work flows. The price is fixed per step, with a cap, and is set before the step begins. We sign a data processing agreement before we touch any information.
A policy helps you decide the limits before an agent starts working. The text here is general information and does not replace legal advice. Ask a lawyer or your data protection officer to assess what applies in your business.
Common questions
What is an AI policy?
A short written agreement on how AI may be used in the company: which tools, which information is never entered, who owns the questions and where a person reads the result.
Does a small company need an AI policy?
Yes, and it can be shorter. As soon as anyone in the company uses AI for work, it helps that everyone knows which information may be entered and whom to ask.
Do we have to ban certain tools?
You do not need to ban anything in advance. Decide which tools are approved and which information is never entered, and let the policy say how to ask for a new tool.
Who should own the policy?
A person with the mandate to answer questions and update the text, often someone in management or the person responsible for IT or data protection.
How long should an AI policy be?
One page is enough for most companies. A longer policy is hard to read, and then it is rarely followed.
Is a template from the web enough?
A template is a good start. Adapt it to your tools and tasks and have a lawyer read it before you share it.