In brief
An AI policy template should pair every rule with a system setting and the record it leaves: approved tools behind SSO, data classes per tool, DLP, audit logs, training records. The AI Act itself does not require an AI policy; ISO/IEC 42001 (clause 5.2) does for an AI management system.
Key takeaways
- The AI Act does not, as such, require an internal AI policy (as of 3 November 2026); it does require AI literacy measures under Article 4, applicable since 2 February 2025.
- Regulation (EU) 2026/1744 rewrote Article 4: companies must take measures to support AI literacy, but need not guarantee any specific level for each person.
- ISO/IEC 42001:2023 requires top management to set an AI policy (clause 5.2) for an AI management system.
- Every rule in the policy needs a control and evidence: an app catalog status, a DLP rule, an audit log, a training record.
- Marking an AI app as unsanctioned in Microsoft Defender for Cloud Apps does not block it by itself; blocking runs through Defender for Endpoint or a supported proxy.
A company AI policy is the internal rule set that names the approved AI tools, the data allowed in each, how use is logged and how new tools get approved. An AI policy template is worth using only if every rule points to a setting someone can check. A policy that cannot be shown in a system is not a control. It is a document about a control. The EU AI Act does not require one; ISO/IEC 42001 does. Here are the sections, the control and evidence behind each rule, and where AI governance and security consulting helps.
Does the AI Act require an AI policy?
No, not as such (as of 3 November 2026). The AI Act assigns duties by role and risk class, and no article asks a company that uses ChatGPT, Copilot or Gemini to adopt an internal AI policy. What applies to almost every company is Article 4 on AI literacy, which has applied since 2 February 2025.
The Digital Omnibus on AI, Regulation (EU) 2026/1744, replaced Article 4. Providers and deployers "shall take measures to support the development of AI literacy of their staff", and the duty "does not require providers or deployers to guarantee any specific level of AI literacy of any individual".
A policy rule on training, plus a record of who completed it, is the simplest way to show those measures.
Stricter duties come with high-risk systems. Deployers must use them according to the instructions for use and assign human oversight to competent people (Article 26). For Annex III systems this applies from 2 December 2027. If your staff only use general-purpose assistants, the policy is good practice and useful evidence, not a legal duty.
National rules add authorities, not a policy duty. In Poland the Act on Artificial Intelligence Systems (Journal of Laws 2026, item 1003) set up a new commission, KRiBSI, as the authority, and its provisions on proceedings and fines apply from 28 October 2026. In Germany the KI-MIG, the national act implementing the AI Act, makes the Bundesnetzagentur the default market surveillance authority.
What should an AI policy template contain?
Keep it short enough that people read it. A company AI policy that works has eight sections, each with a named owner:
- Scope and owner. Who it covers (staff, contractors, interns), who owns it and when it is reviewed.
- Approved tools. A register of AI tools, the account type allowed (company tenant, never a private account) and the purpose of each.
- Data classes per tool. Which classes of information may go into prompts and uploads of each tool.
- Approval path. How someone requests a new tool or use case, who decides and how fast they answer.
- Human review. Which outputs a person checks before they reach a customer, the codebase or a decision about a person.
- Logging and retention. Which interactions are logged, where, and for how long.
- AI literacy. What training each role gets and how completion is recorded (Article 4).
- Incidents and concerns. How to report a prompt that leaked data or an output that caused harm.
How do you verify each rule in your systems?
Write the control and the evidence next to every rule. A rule without a setting behind it is a wish. A setting that leaves no record cannot be shown to an auditor, a customer or your board. The table is the core of the template.
| Policy rule | Control in the system | Evidence |
|---|---|---|
| Only approved AI tools | SSO for approved tools; other generative AI apps marked unsanctioned and blocked on endpoints or at the proxy | tool register, app catalog status, block events |
| Company accounts only | sign-in through your identity provider, company domain only | user list per tool, onboarding and offboarding report |
| No confidential or personal data in public tools | endpoint DLP rule that blocks pasting and uploading sensitive information to AI sites | DLP alerts and overrides |
| Interactions are logged | audit log of the AI tool, exported to your log system | log entries, retention setting |
| Vendors do not train models on business data | business plan and contract terms, model training switched off in the admin console | contract, data processing agreement, dated screenshot of settings |
| Every role is trained | course assigned per role in your learning platform | completion records per person |
| New tools go through approval | request form or ticket workflow with a named approver | tickets with decision and date |
Vendor documentation (as of 3 November 2026) shows where the gaps are. Marking an app as unsanctioned in Microsoft Defender for Cloud Apps "doesn't block use" by itself; blocking runs through Defender for Endpoint or a supported proxy. Copilot interactions are recorded in Microsoft Purview Audit (Standard) as CopilotInteraction events.
Purview endpoint DLP can block pasting or uploading sensitive information to third-party AI sites in Edge, Chrome and Firefox, but only on devices onboarded to Purview. OpenAI keeps compliance logs for ChatGPT Enterprise and Edu for 30 days, so a longer retention period in your policy means exporting them.
In Google Workspace the Gemini app is on by default for licensed users over 18, and admins can turn it off per organizational unit or group.
Why is a ban not a control?
A ban leaves no evidence that anyone follows it and pushes use to private accounts and phones, where you see nothing. That is shadow AI, and our article on shadow AI as a data security risk shows what leaks that way.
An approval path works better: a short form, a named approver, a response deadline and a register entry. Each request becomes a recorded decision, which ISO/IEC 42001 expects and customer security questionnaires ask about.
The tool register also becomes your list of AI systems in use. You need that list anyway to check your role and risk class under the AI Act, as described in our guide to AI Act obligations for companies.
Which data may go into prompts?
Reuse the data classification you already have instead of inventing one for AI. The policy then maps each class to a type of tool:
| Data class | Public tool, private account | Company tenant under contract | AI on your own infrastructure |
|---|---|---|---|
| Public | allowed | allowed | allowed |
| Internal | not allowed | allowed | allowed |
| Confidential (contracts, source code, customer data) | not allowed | only with approval of the data owner | allowed |
| Personal data | not allowed | only with a legal basis, a data processing agreement (GDPR Article 28) and, where required, a DPIA (Article 35) | same rules as any other processing |
Before a tool gets personal data, run the same GDPR checks as for any other processor. On 6 August 2026 the Polish data protection authority UODO published checklists for organizations planning to use an AI tool and stressed that they do not replace a risk analysis or a DPIA.
Classification works only if someone owns the data. That groundwork is covered in data governance as the first step to AI.
What does ISO/IEC 42001 add to an AI policy?
ISO/IEC 42001:2023 specifies requirements for an AI management system and is meant for organizations that provide or use AI systems. Under clause 5.2, top management sets an AI policy. The policy must fit the organization's purpose, give a framework for AI objectives, commit to applicable requirements and continual improvement, and be documented and communicated.
Annex A adds controls for policies related to AI (A.2) and for the use of AI systems (A.9). A certificate from an accredited body confirms that the management system meets the standard. ISO/IEC 42001 is not a harmonized standard under the AI Act and gives no presumption of conformity. How to get an AI management system ready for certification is on our AI Act and ISO/IEC 42001 page.
Where to start: controls first, the policy last
- List the AI tools in use: app discovery logs, expense reports, a short staff survey.
- Choose the approved tools and move users to company accounts with SSO.
- Map your data classes to tool types, as in the table above.
- Set up the approval path and the tool register.
- Turn on audit logs and set their retention.
- Assign AI literacy training per role and record completion.
- Write the policy last, as a description of what is already set up.
The order matters: a policy written after the controls describes reality, and each rule already has its evidence. To set up the controls or check the evidence with us, get in touch.
Frequently asked questions
- Is an AI policy mandatory for companies?
Not under the AI Act itself (as of 3 November 2026). An AI management system under ISO/IEC 42001 requires one (clause 5.2), and customers and auditors often ask for it.
- Who should approve the company AI policy?
Top management. ISO/IEC 42001 (clause 5.2) makes top management responsible for setting the AI policy; day-to-day ownership usually sits with IT, security or compliance.
- Can we simply ban ChatGPT?
You can, but a ban leaves no evidence and moves use to private accounts. An approved tool with company accounts, a DLP rule and audit logs gives you both control and a record.
- Does the AI policy cover contractors and freelancers?
It should. The AI literacy duty in Article 4 covers staff and other persons dealing with the operation and use of AI systems on the company's behalf, so contractors who use your AI tools belong in scope.
- How do we show AI literacy under Article 4 of the AI Act?
The amended Article 4 requires measures, not a guaranteed level per person. Role-based training with completion records, and the policy rule that requires it, are the usual evidence.
- Does ISO/IEC 42001 certification mean AI Act compliance?
No. ISO/IEC 42001 is a management system standard, not a harmonized standard under the AI Act, so it gives no presumption of conformity.
