1. Why a written policy comes before any tool
Many companies discover that employees are using generative AI tools when something goes wrong: a text containing a client's data pasted into an external service, an invented answer ending up in a quote, or a confidential document uploaded to be summarised. The problem is not that the person wants to work better, but that nobody has clearly explained where it is permitted and where it is not. A written policy turns an implicit rule into a shared one.
The policy also has a demonstrative function. The European AI regulation requires organisations that use artificial intelligence systems to ensure an adequate level of AI literacy among their staff, and an internal document, kept up to date and signed for acknowledgement, is one of the tools for doing so. Check with an adviser which obligations apply to your case before considering the policy complete.
2. What to prepare before writing
Before writing a single line you need four pieces of information. The first is a list of the uses already under way, including unofficial ones: ask heads of department which tasks they carry out with AI tools and what data they enter. The second is a map of company data, classified as public, internal, confidential and personal. The third is a list of approved tools, or tools to be assessed, together with their contractual terms. The fourth is the name of a person responsible for the policy.
The owner does not necessarily have to be a technician. Usually it is someone who knows the processes and has the authority to enforce the rules, supported by the IT manager for technical aspects and by the data protection contact for legal ones. If these roles do not yet exist, appoint them before writing the document: a policy without an owner is not updated and gradually loses its value.
3. The steps for writing the policy
The sequence matters. First decide the purpose, then classify the data, then set out the approved tools and permitted uses, and finally fix verification and reporting. Skipping a step leads to rules that do not connect with each other and that staff struggle to follow in concrete cases. The steps do not have to happen in a single meeting, but each must conclude with a text agreed by the owner before moving to the next.
Write in plain language and use examples. A sentence such as "do not use sensitive data" says nothing to someone who does not know what that means. A sentence such as "do not paste the text of a contract or a list of customers into tools that have not been approved" can be applied the same day. Avoid references to laws in the body of the text: it is better to refer to an annex checked by an adviser, so the policy remains readable and easy to update without being rewritten each time.
- Define the purpose: one or two sentences saying why the policy exists and to whom it applies.
- Classify the data: list the categories that may never enter an unapproved AI tool, such as personal data, customer data, contract documents and non-public information.
- List the permitted tools: state which are approved for which uses and which are not, without necessarily naming commercial products in the text.
- Set the permitted use cases, with concrete examples for each department.
- Set the verification rules: who checks an output before it leaves the company and in which cases a review is required.
- Define how to report an incident, for example data entered by mistake, and whom to contact.
- Have the text reviewed by the legal adviser and the IT manager, then formally approved by management.
4. What it must contain, area by area
The table summarises the minimum areas. Not all of them require a long chapter, but each must have an explicit answer. The rules shown are generic examples: they must be adapted to the real activities, the data processed and the tools actually in use. Each row of the table corresponds to a question an employee may ask in a moment of doubt, so it is best read as a list of questions to be clarified.
One area often overlooked is exceptions. Anyone with a special case must know whom to ask for authorisation and how the decision is recorded. Without an exception procedure, people find shortcuts, and shortcuts become the rule. Also state who can grant an exception and for how long. Write down the duration of the exception and how it is reviewed, otherwise it remains valid without anyone noticing.
| Area | Question it answers | Example of a generic rule |
|---|---|---|
| Scope | Who it applies to and which tools it covers | Applies to all staff and to contractors with access to company systems |
| Data | Which information may never be entered | Personal data, contract documents and confidential information do not go into unapproved tools |
| Permitted uses | What AI may be used for | Drafts of generic texts, summaries of public documents, translations of non-confidential materials |
| Verification | Who checks the output and when | Any content intended for customers or for operational decisions is reviewed by a competent person |
| Reporting | What to do in case of error | Data entered by mistake must be reported to the owner immediately, without trying to delete it alone |
5. Common mistakes when drafting
The first mistake is writing a policy that prohibits everything. A blanket ban is quietly circumvented, and the work moves to personal tools, where nobody can see anything. It is better to define a simple authorised route, with approved tools and clear use cases, and to prohibit explicitly only what is genuinely dangerous for data and customers. A targeted prohibition, by contrast, says exactly what not to do and why.
The second mistake is treating the policy as a document to be filed away. It must be read, used and updated when the tools, the data processed or the applicable rules change. The third is failing to explain it: a text sent through the internal mail without a moment for discussion is forgotten within days. The fourth is failing to provide for who decides on new cases, those nobody has yet imagined, which are often the riskiest.
6. How to check that the policy works
A policy is effective when people know what to do in concrete cases. A simple check is to ask a sample of employees from different departments a few questions: which data they may not enter, whom they ask for authorisation, and what they do if an output seems wrong. The answers show where the text is unclear or has not really been read.
Also watch the indirect signals. If the IT manager detects use of unapproved tools, or if recurring exception requests arrive, the policy should be revised or an authorised alternative offered that meets the need. Plan a periodic review of the document and a review each time one of its premises changes, recording the date, version and approval of every amendment. Spontaneous reports, when they come in, are also a useful signal.
7. The guides to read next, in the right order
This policy calls for a closer look at some topics, which have dedicated guides. It is best to read them in this order: first unauthorised use, then data, then the choice of use cases. The topics, posed as questions, are these. Each one answers a question the policy alone cannot close, because it requires data, examples or a decision by the company. Starting from the phenomenon rather than the rule helps make the rules concrete.
You do not need to read everything before writing the policy. But the sections on data and on unauthorised use should be clarified first, because they determine the most important rules. In practice it is enough to be clear about who uses what, which data is involved and who decides on exceptions; the rest can follow without holding up the team's work.
- What shadow AI is, meaning the ungoverned use of AI tools, and how to identify it.
- How to prepare documents and data before using them with an AI system.
- Which first use cases to choose, using a method that assesses value and feasibility.
- What typical risks language models carry, such as prompt injection.
- How to reduce model hallucinations and check their answers.
8. How to start
This week, ask three heads of department which tasks they carry out with AI tools and what information they enter. A formal questionnaire is not needed: two questions in a meeting are enough to bring out most real uses. Note the answers without judging them, because they are the basis of the text. Also note who has already tried new tools, as they are often the most useful people for building the rule together with the company.
The next step is to prepare a short draft, following the steps in section three, and have it reviewed by the IT manager and the data protection contact. Only after that review should it be presented to staff, in person or in a meeting, leaving room for questions before the rule comes into force. Remember that acknowledgement must be recorded, with a date and a signature or equivalent confirmation.
Frequently Asked Questions
Is a policy legally required?
A written policy is not always an explicit obligation, but the European AI regulation requires those who use artificial intelligence systems to ensure an adequate level of AI literacy among staff, and an internal document is a natural way to demonstrate this. The specific obligations depend on the type of systems used and the sector, and should be checked with an adviser.
Who should sign the policy?
All staff who use digital tools to work, including contractors with access to company systems. The signature for acknowledgement serves to show that the rule has been communicated. The document should be approved by management, because rules without an authority behind them struggle in difficult cases.
Can I ban generative AI altogether?
You can, but in practice a total ban often leads to hidden use on personal tools, which are harder to see and control. It is usually more effective to name a few approved tools, the data that must not be used and the cases in which a review is required. A targeted ban on specific data and activities is more sustainable over time.
How long should the policy be?
Long enough to answer the most frequent questions, and no longer. For many organisations a few pages are sufficient, with separate annexes listing the approved tools and the use cases. A long text that nobody reads is less useful than a short one that everybody knows. Review the length once the policy has been in use for some time.
How often should it be updated?
There is no single schedule. It is best reviewed when the approved tools change, when the types of data processed change, when the applicable rules change, or when incidents or recurring exception requests emerge. Record every change with its date, version and approval, so that anyone can tell which rule was in force at any given time.
Is the policy enough to prevent incidents?
No. A policy reduces risky behaviour, but it does not replace technical controls such as access permissions, activity logging and human review of outputs. The policy says what to do; the controls make it possible to do it and show whether someone does not.
Who is responsible if an employee makes a mistake?
It depends on the case and the employment contract, and should be checked with a legal adviser. In general the policy serves to clarify responsibilities and reporting routes, not to shift blame onto one person. A system that punishes whoever reports a mistake encourages incidents to be hidden, which is the opposite of what is needed.