Guide · Governance · 7 min read · Updated October 2026

AI governance framework: who decides what your AI agents may do.

Governance used to mean an acceptable-use policy for chat tools. Once agents send emails, update records and prepare payments, it has to answer harder questions: who owns each agent, what it may do alone, and who reads what it did.

What is an AI governance framework?

An AI governance framework is the set of rules a company uses to decide what its AI agents may do: who owns each agent, what it may do alone, what always needs a person's approval, how its actions are logged and reviewed, and how changes are tested before release. Once agents act rather than only answer, it is how leadership keeps control.

Looking for the technical side? AI agent security →

Christos Papadimitriou, theagency47 · Updated October 2026

What AI governance means once agents act

A chat assistant produces text that a person reads and then decides what to do with. An agent takes the next step itself: it replies to the customer, files the invoice, updates the CRM. The person who used to be the last check is no longer in the loop by default. Governance is how you decide, in advance and in writing, where that person stays in the loop and where they do not.

The technical controls that enforce those decisions, such as scoped access, sign-in and audit trails, are covered in our AI agent security guide. This guide is about the decisions themselves.

The five parts of the framework

1. Ownership: one named owner per agent

Every agent has one person who answers for it: usually the head of the process it works in. The owner reads the logs, approves changes, and can switch it off. An agent "owned by IT" or "owned by the vendor" in practice has no owner.

2. Autonomy levels: what an agent may do alone

Set one of three levels per agent, per action: it suggests and a person acts; it acts with approval; or it acts on its own inside limits. Most agents start at the first or second level and move up only after their tests pass and the owner signs. This is bounded autonomy in practice.

3. Approval rules: what always needs a person

Some actions should need a person no matter how well an agent performs: anything that sends money, deletes data, commits the company contractually, or makes a decision about an individual. Write the list down and make it part of every agent's specification. The pattern is human-in-the-loop, with escalation rules for the cases the agent should hand over. In practice, approvals and logs like these are what clients see in OS47.

4. Logs and review: who reads them, how often

Logging everything is easy. Reading it is the control. Decide who reviews each agent's actions, how often (daily in the first weeks, then weekly or by sample), and what triggers an immediate look, such as a spike in overrides or a complaint.

5. Change control: evals before every change

A new prompt, a new model version or a new knowledge source can change behaviour. Every change runs against the agent's test suite before release, and fails back to the last good version if the score drops. No test, no change.

Who sits on the AI steering group

	Mid-size company (25–250 staff)	Large company (250+ staff)
Chair	Owner or CEO	Executive sponsor (CEO, COO or a board member)
Members	The heads of the departments with agents, the IT lead or provider, whoever handles data protection	CIO, CISO, DPO, CFO, HR, heads of departments in scope, workers' representatives where required
Cadence	Monthly, 45 minutes	Monthly steering, quarterly strategy review
Decides	Next agent, autonomy changes, budget	Portfolio, policies, risk acceptance, budget per department

Who holds which decision, from the yes to the off switch, is set out in who owns the decision to adopt AI agents: CEO vs CIO.

Staff AI policy: personal tools vs company agents

Two different things need two different rules.

Personal AI tools (a chat assistant an employee uses to draft or summarise): which tools are approved, under which company account, what data may never be pasted in, and that the employee stays responsible for the output.
Company agents (software acting on company systems): only deployed through the framework above, with an owner, an autonomy level and logs. Nobody connects an agent to a company system on their own initiative.

Keep the staff policy to one page. A long policy nobody reads governs nothing.

EU rules in plain words

GDPR applies whenever an agent processes personal data: a lawful basis, data minimisation, a processing agreement with every vendor involved, and, under Article 22, no decisions with legal or similarly significant effect on a person made solely by automated means without safeguards. High-risk processing needs a data protection impact assessment.

The EU AI Act sorts AI uses into risk classes: prohibited practices, high-risk uses (for example recruitment screening or credit scoring), uses with transparency duties (for example telling people they are dealing with an AI), and minimal risk. A company that uses an AI system is its deployer. Most back-office agents, such as email triage or invoice extraction, are generally not high-risk; agents that help decide about people can be. Deployers also have to support their staff's AI literacy, and for high-risk uses must assign human oversight to people with the competence and authority to act. If your agents run on the Grok API, add its output watermarking, on by default from 7 November 2026, to your decisions.

The dates that apply and what they mean for the person who decides are in our CEO vs CIO post. Check each use case with your legal adviser; this guide is not legal advice.

Template: the one-page agent charter

Every agent gets one page, signed by its owner before it goes live and updated when its autonomy changes.

Field	Example
Agent and purpose	Invoice agent: extracts supplier invoices and matches them to purchase orders
Owner	Head of finance operations (named)
Systems and data it may read	Accounts payable inbox, purchase orders, supplier master data
Actions it may take alone	Create draft entries, flag mismatches
Actions that need approval	Post entries, any payment, any supplier change
Never	Contact suppliers, change bank details
Escalation	Amount above threshold, unknown supplier, low confidence: to the owner's queue
Logs and review	All actions logged; daily review for 2 weeks, then weekly sample
Tests	Test suite re-run before every change; minimum pass score agreed
Off switch	Who can stop it, and how fast (tested before launch)

How we apply these rules on our own builds is on our Trust page. For the company-wide programme the charter sits in, see AI transformation services.

FAQ

Questions about AI governance.

What should an AI governance framework include?

Five parts: a named owner for every agent, an autonomy level per action, a list of actions that always need a person, a log review routine, and change control that tests every change before release. Add a staff policy for personal AI tools and a short check against GDPR and the EU AI Act.

Who should own AI governance in a company?

The chief executive owns the limits and chairs or sponsors the steering group; each agent has an owner in the business; IT owns access, logs and the off switch. See CEO vs CIO for the full split.

Is an AI policy for staff the same as AI governance?

No. A staff policy covers personal AI tools employees use. Governance also covers company agents that act on company systems: who owns them, what they may do alone and how they are reviewed.

Does the EU AI Act apply to our AI agents?

If your company uses AI systems in the EU, you are a deployer and some duties apply, such as supporting staff AI literacy. Most back-office agents are not high-risk; agents that help decide about people can be. Check each use case with a legal adviser.

How often should AI agents be reviewed?

Daily in the first weeks after launch, then weekly or by sample, with an immediate review when overrides, complaints or costs spike. The steering group looks at the whole portfolio monthly or quarterly.