How Can Enterprises Move AI Governance From Policy to Practice?




An enterprise can assess an AI system, confirm that it meets governance requirements and approve it for production. But approved is not the same as allowed: that approval does not automatically authorize every action the system can perform.
In my conversations with risk, compliance, AI governance and technology leaders across European financial institutions, this is the gap that comes up most often: the framework exists, the assessment has been done, and yet nobody can say with confidence what the system is actually allowed to do once it is running.
Most established enterprises already have AI policies, risk processes and defined responsibilities. These provide a foundation for determining acceptable use and the conditions under which systems operate. The challenge is ensuring those conditions continue to apply when an agent acts, particularly as its capabilities or operating environment changes.
What Happens After an AI System Is Approved?
Consider an AI agent supporting an enterprise service desk. It has been approved to retrieve incident information, consult internal knowledge bases and update service records.
Now suppose the agent receives a request to change a production configuration. The organization has a policy requiring additional authorization because an incorrect change could disrupt critical services.
The governance decision is clear. But does the system prevent the agent from making the change without approval?
An instruction telling the agent to follow company policy is not an enforceable restriction. The requirement has to be reflected in the controls governing the action. Instructions live inside the model's context, where they can be misread, overridden by later instructions or manipulated through the content the agent processes. A control that must hold has to sit outside the model, at the point where the action is executed: the tool, the API, the permission.
This matters particularly for agents that can select tools and sequences of actions at runtime. An assessment evaluates intended use, known capabilities and expected operating conditions. It remains relevant, but permissions, connected applications or responsibilities can change after approval. Changes that affect the approved risks or restrictions need further review.
The same pattern appears wherever agents operate. In banking, an agent approved to answer customer queries and update contact details becomes a different system the moment it can initiate a payment, adjust a credit limit or close an account. The assessment did not cover those actions; the controls must ensure the agent cannot perform them until it does.
What Changes When Governance Becomes Operational?
The response is not to require human approval for every agent action. Enterprises need to distinguish between activities according to their potential consequences. Governance that blocks everything fails as surely as governance that blocks nothing; it simply fails on the value side, and the enterprise ends up with an expensive assistant instead of an agent.
Apply controls according to risk
Retrieving authorized incident information and changing a production configuration carry different risks. The first can proceed within established permissions. The second could affect system availability and may be difficult to reverse.
The organization must decide which actions the agent can perform independently, which require authorization and which remain prohibited. A policy establishes the requirement; the relevant technical control applies it when the action is attempted. The step in between, translating a requirement written for people into a control a system can apply, is where most organisations are weakest. The policy lives in a document owned by Compliance; the permission lives in a configuration owned by Engineering; and the mapping between the two is rarely written down, let alone maintained.
Monitor activity and review exceptions
Suppose the agent repeatedly attempts a configuration change that requires approval.
The control appears to be preventing the action, but the repeated attempts raise another question: do the agent's instructions, permissions and assigned responsibilities still reflect its intended role?
Monitoring provides information about the pattern. The responsible team can then determine whether the system needs revised instructions, different permissions or further governance review.
A control can prevent an unauthorized action while still revealing a mismatch between expected and actual behavior.
Use evidence to evaluate controls
An operational log can show that a configuration change was blocked. That alone does not establish whether the correct governance requirement was applied.
The organization also needs to know which restriction applied, whether authorization was required and what outcome followed.
Reviewing operational evidence against the approved requirement helps determine whether the control worked as intended and whether its restrictions remain appropriate as the agent's operating conditions change. Evidence of this kind is most useful when it is produced as a by-product of operation rather than reconstructed afterwards for an audit. If every blocked or authorised action already carries the identity of the requirement that governed it, the review becomes a matter of reading, not forensics.
How Does This Put Responsible AI Into Practice?
Responsible AI principles such as accountability and human oversight need practical meaning after deployment. A requirement for human intervention must establish which actions need approval, when that intervention occurs and who reviews exceptions. This is no longer only a matter of principle. The EU AI Act treats risk management as a continuous process across the system's lifecycle (Article 9), requires that high-risk systems can be effectively overseen by people (Article 14) and that their operation is logged (Article 12), and expects providers to monitor systems after they are placed on the market (Article 72). ISO/IEC 42001 asks for the same discipline in management-system terms. Each of these obligations is, in practice, a question about what happens after approval.
Risk and Compliance teams own these requirements, while AI, Data and Engineering teams apply them within the systems they operate. Human judgment remains necessary for determining acceptable risk and deciding when oversight should change. In regulated industries this is the familiar three-lines-of-defence model applied to a new kind of actor: the teams that run the agent are the first line, Risk and Compliance the second, and Internal Audit needs evidence it can test rather than narratives it has to trust.
Where Should Enterprises Start?
Extending governance into operation does not require replacing existing frameworks. A useful starting point is one AI system in production and one policy requirement that materially affects its actions. Starting with one is a way to learn the mechanics, not the end state. The institutions I work with have hundreds, in some cases more than a thousand, AI use cases in production, and the chain from requirement to control to evidence has to be repeatable across all of them. What is done by hand for one system will not be done at all for the five-hundredth.
Three questions help:
- Where is the requirement applied? Does the relevant system enforce it?
- Who reviews exceptions or changes? Is responsibility clear after deployment?
- What evidence is available? Can the organization establish whether the control worked as intended?
The answers show where existing governance is effective and where further oversight is needed. For enterprises expanding AI adoption, operational governance helps determine which activities can proceed autonomously, where authorization remains necessary and when changing conditions require review.
Policies and assessments remain the foundation. Their requirements must also continue to govern what AI systems are permitted to do in operation.
This is the logic behind how we approach AI governance at GFT: prescribe the requirements for each AI system, enforce them where the action happens, and let the evidence compound into something auditors, regulators and the business can rely on. The tooling matters, but the shift is mainly one of mindset: from governance as a gate to governance as an operating layer.
That raises the next question: when a control blocks an action, can the enterprise trace it back to the risk decision that justified it? That is the subject of the next article in this series.


