AI policy template: rules for staff and for AI agents

A company AI policy sets out which AI tools staff may use and for what, which data never goes into them, who checks the output and who is accountable for what. An AI policy template needs eight sections for that, and a ninth with separate rules as soon as AI agents act in your systems.

A chat tool produces a draft that a person reads before using it. An AI agent, by contrast, reads a complaint from the service inbox, pulls order data from the ERP and creates a draft credit note there, so what happens depends on permissions and checkpoints rather than on one person's diligence.

Keep the policy short enough for staff to go through it in full during induction training. Anything that changes often, such as the list of approved tools, belongs in a register maintained by IT that the policy refers to.

No law requires a document called an AI policy. Several duties are hard to meet demonstrably without written rules, though, and the policy is where they come together. This overview is not legal advice; the assessment for your company belongs with data protection, compliance or counsel.

Since 2 February 2025, Article 4 of the EU AI Act has required providers and deployers to take measures to support the AI literacy of their staff and of others using AI systems on their behalf, which covers any company using AI tools in its business. Since the amendment by Regulation (EU) 2026/1744 they need not guarantee any specific level of literacy, and neither Article 99 of the Regulation nor Germany's implementing act, the KI-MIG, sets a separate fine for it.

The European Commission's Q&A on AI literacy states that no certificate is needed and that organisations can keep an internal record of training. Germany's market surveillance authority, the Bundesnetzagentur, adds that a lack of AI literacy can be regarded as a breach of the duty of care, particularly where damage occurs.

Since 2 August 2026 the transparency duties of Article 50 have applied too. Deployers who use AI for deep fakes, or for text published to inform the public on matters of public interest, must disclose that the content is artificially generated or manipulated, unless such text has undergone human review or editorial control and a natural or legal person holds editorial responsibility. Four further sets of rules make a written policy sensible even without an explicit duty.

The template has eight sections that you fill in with your own roles, tools and deadlines; replace the names in square brackets with the responsible function in your company.

Word the third section on data classes with particular care. A tool can be replaced, but an entered design drawing or price list cannot be retrieved if the provider stores inputs or uses them for training, which is why IT checks these two points before approving a tool for confidential data.

AI agents need a section of their own because they do not just generate text but act with write access in the ERP, CRM, ticket system or DMS. Rules about what staff type into a chat window do not reach them, because the agent works without anyone seeing each step.

Two risks explain rules 9.2 and 9.5. An agent reading emails or PDFs can follow instructions hidden in them (prompt injection), and draft-only write access contains the damage. A model change by the provider can shift behaviour across thousands of cases without any change in the workflow, which a test set run before every release exposes.

Section 9 does not set details per workflow; it sets the conditions under which an agent may go live at all. The details sit in the control set of each workflow, described in governance inside the workflow, and the roles behind them in the operating model for AI in operations. The glossary of AI agent terms defines control set, shadow mode and confidence threshold.

The policy governs what people may and must do, and the control set governs what one particular agent may do in one particular workflow. The policy requires a control set for every agent and names who approves it; the workflow defines its content.

The policy might say, for example, that an agent writes in draft status only until shadow mode meets the acceptance criteria. The control set for claims handling then says that the agent may create draft credit notes in the ERP, that cases above the threshold go to quality management and that refusals never go out without a claims handler.

The two stay separate because a policy is revised once a year and agreed with employee representatives, while a control set changes with every new threshold and model version. If the details sat in the policy, every recalibration would have to go through the same approval round.

The Workflow Assessment determines, among other things, the risk class and owners for the top candidates and sets the first workflow's checkpoints in the build scope. The AI Act check gives a first classification under the AI Act beforehand.

Where there is a works council or other employee representation, roll the policy out together with it. In Germany, the works council co-determines the introduction and use of technical devices designed to monitor employees' behaviour or performance under section 87(1) no. 6 BetrVG, and the Federal Labour Court treats a device as designed for this when it is objectively suitable (order of 13 December 2016, 1 ABR 7/15). An AI tool that links inputs to user accounts collects that kind of information.

Where the works council has to assess the introduction or use of AI, section 80(3) BetrVG deems calling in an expert necessary. Allow time for that and present the tool register, data classes and agent rules together before the first go-live, so that tools are not negotiated one by one; whether such a framework suffices in a given case is for your employment lawyers. Outside Germany, check the local rules on informing and consulting employee representatives.

For training under Article 4, the Bundesnetzagentur suggests three building blocks, namely identifying needs, designing measures, and documenting, evaluating and refreshing them. Its guidance lists type, content and duration of measures and the participants as what to record. Review the policy itself at least once a year and after every change in law or tools.