Enterprise AI governance framework template

Your printable governance framework template

Work through the sections in order. Sections 1 to 3 take one leadership meeting. Sections 4 to 9 need the owners of the source systems in the room. Section 10 closes the document and sets the date of the next review.

1. 1. Purpose and scope

   State why the organization uses AI, which business decisions it may inform or act on, and which systems, teams and regions this framework covers. Name what is out of scope, such as personal use of consumer tools on unmanaged devices, so that an exclusion is a decision rather than an accident.

   Owner / decision / evidence:

2. 2. AI inventory

   List every AI system in use or under evaluation: assistants, connectors and MCP servers, agents, AI features embedded in the ERP or CRM by the vendor, and internal models. For each, record the owner, the systems it reads or writes, the users who can reach it and the date it was last reviewed. An inventory nobody maintains is the first control to fail.

   Owner / decision / evidence:

3. 3. Roles and accountability

   Name the executive sponsor, the owner of each AI system, the data owners whose records it touches, the security reviewer and the incident contact. Separate the person who approves a use case from the person who builds it. Write down who can change permissions, metric definitions and model settings, and who must be told when they do.

   Owner / decision / evidence:

4. 4. Risk tiering

   Define three or four tiers, from low to high, based on the consequence of a wrong output, the sensitivity of the data involved and whether the system acts or only answers. A read-only assistant on a sanitized dataset and an agent that posts journal entries need different review depth, approval authority and monitoring. Assign every inventory item a tier and record the reasoning.

   Owner / decision / evidence:

5. 5. Data access and permissions

   Require that every assistant, connector and agent acts under the identity and permissions of the person using it, enforced by the source system rather than by a shared service account or a prompt. Record which data classes are excluded entirely, for example payroll, health or card data, and where each exclusion is enforced: in the source system, in the connector, or in the assistant's settings.

   Owner / decision / evidence:

6. 6. Business definitions and evidence

   Decide which metric definitions are certified, who owns each one and where the definition lives, so that an AI answer about revenue, margin or pipeline uses the same definition as the board report. Require that any answer used for a decision can show its evidence: the query or records it used, the definition applied and the policies in force at the time.

   Owner / decision / evidence:

7. 7. Human approval for actions

   List the actions an AI system may propose but a person must approve, such as creating a vendor, approving an invoice, changing a price, sending a customer communication or posting to the ledger. State the approval authority by tier, how the approval is recorded and which actions are never delegated to an agent regardless of its confidence.

   Owner / decision / evidence:

8. 8. Monitoring, incidents and audit

   Define what is logged for each system (who asked, what was accessed, what was returned, what was approved), how long logs are kept and who reviews them. Describe what counts as an AI incident, such as a wrong answer acted upon, a permission bypass or a suspected prompt injection, and the response path. Choose the measures you will track, for example answer acceptance rate, correction rate and access denials, before the first pilot starts.

   Owner / decision / evidence:

9. 9. Third-party and connector review

   Set the review required before any vendor model, connector, MCP server or embedded AI feature is enabled: who builds and supports it, the authentication method, the scopes requested, read versus write tools, and the retention and training-data commitments confirmed on the vendor's own page. Name the person who owns the vendor relationship. The connector approval checklist linked below covers this step in detail.

   Owner / decision / evidence:

10. 10. Review cadence

    State how often the framework, the inventory and each high-tier system are reviewed, what triggers an out-of-cycle review (a new connector, a model change, a regulation change, an incident) and who signs off. Record the version and date of this document so that an auditor or a new team member knows which rules were in force.

    Owner / decision / evidence:

How to use this template

The ten sections map to the four functions of the NIST AI Risk Management Framework: Govern, Map, Measure and Manage. If your organization already reports against the AI RMF, use the table below to show where each function is covered. If it does not, the table is a useful order for the first review: govern first, then map, then measure and manage.

Which standards does this template reference?

Three public references shape the structure: the NIST AI RMF, ISO/IEC 42001 and the EU AI Act. None of them requires this exact template, and completing it does not make an organization compliant with any of them. It gives you the owners, decisions and evidence that each one expects you to be able to show.

• NIST AI RMF 1.0, a voluntary framework released in January 2023, organized around the Govern, Map, Measure and Manage functions. NIST notes that the framework is being revised, so check the NIST page for the current version before citing it in a policy.
• ISO/IEC 42001:2023, the management-system standard for AI. It sets requirements for establishing, running and improving an AI management system, and an organization can choose to be certified against it. Sections 1, 3, 8 and 10 of this template produce the kind of documented scope, roles, monitoring and review that a management system expects.
• The EU AI Act, which sets obligations by risk category. The linked site is an independent explorer of the Act maintained by the Future of Life Institute; rely on the official text and your counsel for obligations. Section 4 (risk tiering) and section 7 (human approval for actions) are where an EU-exposed organization will spend the most time.