AI Governance
Agentic AI governance: the controls every AI agent needs before it acts
AI agents do not just answer questions; they take actions in your systems. Kriv AI helps regulated organisations decide what each agent may do, who answers for it and how every action is recorded, before it goes live.
Agentic AI governance is the set of controls that decide what an AI agent may do on its own, what needs a person's approval, who is accountable for it and how its actions are recorded. Because agents act in systems rather than just producing text, governance has to cover identity, permissions, approvals, logging and a way to stop the agent, in addition to the policy that applies to any AI system.
why different
Why agentic AI governance is different
An assistant drafts something and a person decides what to do with it. An agent takes the next step itself: it searches a system, updates a record, sends a message or calls another tool. That changes the governance question from 'is this output acceptable?' to 'what is this software allowed to do, and how would we know if it did something it should not?'
The general structure still applies: a policy, an inventory, owners and risk tiers, as set out in our AI governance framework. Agents add a layer of operational controls on top, because their risk sits in their actions as much as in their words. Our explainer on agentic AI versus traditional automation covers the underlying difference.
framework
An AI agent governance framework in seven controls
This is the control set Kriv AI applies to every agent before it goes live. It is our own practice, not an external standard.
| Control | The question it answers | What good looks like |
|---|---|---|
| Owner | Who answers for this agent? | A named business owner, recorded in the AI inventory with the agent's purpose. |
| Identity | What does the agent act as? | Its own service identity, never a person's account, so its actions are distinguishable. |
| Permissions | What can it reach and change? | Only the systems and actions its one task needs; read-only wherever possible. |
| Approval points | What must a person approve? | Anything that leaves the organisation, moves money, changes a record of consequence or touches a customer. |
| Logging | What did it do, and why? | Every input, tool call, action and approval recorded where reviewers can find it. |
| Monitoring | How would we know it went wrong? | Agreed checks on its outputs and actions, reviewed on a set cadence, with alerts for anything outside its limits. |
| Stop | How do we halt it? | A tested way to suspend the agent immediately, and a named person allowed to use it. |
risk tiers
Matching AI agent governance to risk
Not every agent needs the same weight of control. An agent that only reads documents and prepares a summary for a person is low risk; one that can update customer records or send external messages is much higher. We tier agents by what they can change, not by how sophisticated the model is.
We start organisations with read-only agents that prepare work for a person, and widen what an agent may do only after it has run under review and its logs show it behaves as intended. Our page on AI agent security risk covers the technical side of that progression.
regulated
Agentic AI governance in regulated industries
In banking, SR 26-2, which superseded SR 11-7 in April 2026, states that generative AI and agentic AI models are not within its scope. That does not leave agents ungoverned: they remain under the bank's own risk management and governance. Our page on agentic AI in investment banking compliance shows the controls applied in that setting.
In healthcare, life sciences and insurance the same principle holds. We write each agent's controls to the policy and frameworks your organisation already answers to, such as the NIST AI RMF, ISO/IEC 42001 or the EU AI Act. Your risk and compliance functions keep the final judgement on what each agent may do.
unapproved agents
Agents nobody approved
Agent features now come built into assistants and automation tools, so teams can create an agent without going through IT or risk review. Those agents can act under a person's own account, with that person's full access, and nobody else knows they exist.
The fix is not a ban. It is a simple route to register an agent, a default rule that agents act under their own identity with limited permissions, and a regular check of the platforms your teams use for agents that are not in the inventory. Bringing an unapproved agent under the seven controls is quicker than discovering it after it has done something it should not.
platforms
Governance across agent platforms and tools
Agents run on many platforms: orchestration frameworks, assistant products and automation tools with agent features. The seven controls apply on all of them, but how you implement each one differs by platform. Our page on governance for AI agent orchestration platforms covers the platform side, and our AI governance tools page covers software for keeping the inventory and approvals.
For worked examples of these controls in practice, see our agentic AI governance case studies.
how an engagement runs
How an agentic AI governance engagement runs
1. Discovery call
We learn which agents you run or plan, on which platforms, and how your risk functions review new systems. Work is scoped and priced after this short discovery call.
2. Agent inventory
We list every agent with its owner, purpose, permissions and risk tier, including agents built by teams on their own.
3. Controls written
We write the seven controls for each agent to your policy and agree them with the owners and reviewers.
4. Configure and test
We configure identities, permissions, approvals and logging, test the stop control, and hand over the records your reviewers will need.
Related resources
Continue exploring
Straight answers
Frequently asked questions about Agentic AI governance: the controls every AI agent needs before it acts
What is agentic AI governance?
It is the set of controls for AI systems that act, not just answer: who owns each agent, what identity it acts under, what it is permitted to reach and change, what needs a person's approval, how its actions are logged and monitored, and how it can be stopped. It sits on top of the policy and inventory that apply to all AI systems.
How is AI agent governance different from AI model governance?
Model governance focuses on whether a model's outputs are accurate and fair. AI agent governance adds control over actions: permissions in real systems, approval points, logging of every step and a way to halt the agent. An accurate model can still cause harm if the agent around it can do more than its task needs.
Does SR 26-2 cover AI agents?
No. SR 26-2, which superseded SR 11-7 in April 2026, states that generative AI and agentic AI models are not within its scope, and it does not set forth enforceable standards. Agents used by a bank still fall under the bank's own risk management and governance, so they still need owners, controls and records.
What is the Model AI Governance Framework for Agentic AI?
It is a published framework title that some readers search for by name. We do not summarise third-party frameworks here; read it at its publisher's site. Whatever external framework you adopt, the practical work is the same: an inventory of agents and written controls for each one, which is what we help put in place.
Should AI agents ever act without human approval?
Low-risk, reversible actions inside a narrow task can run without approval once the agent has proved itself under review, such as filing a document in the right folder. Anything that leaves the organisation, moves money, changes an important record or affects a customer should keep a person approving it.
How do we find AI agents we did not approve?
Check the admin consoles of the assistants and automation tools your organisation licenses for agents, flows and connected apps, and ask teams directly with a simple registration route that does not punish disclosure. Then bring each one into the inventory with an owner and the seven controls, or switch it off.
Who should own an AI agent?
A named business owner who answers for what the agent does, supported by IT for identity and permissions and by risk or compliance for review. Agents built by teams on their own still need an owner, which is why the inventory is the first step in any engagement.
How much does agentic AI governance work cost?
Our work is scoped and priced after a short discovery call, once we know how many agents you run or plan, on which platforms and what review your organisation requires.
Talk to the team that would do the work
Bring your requirements to a working session with the person who'll actually deliver.
Book a Discovery Call