Boards answer for model risk under SR 11-7, and senior management runs the framework that keeps it within limits. For AI agents, that means the board sets how much autonomy it will accept and receives reports showing agents stay inside it, while senior management owns the inventory, policies, validation and independent challenge that make those reports true.
What does SR 11-7 governance require, in plain words?
SR 11-7 is the interagency model risk guidance used by US banking supervisors, including the Federal Reserve and the OCC. It has three parts: sound development and use of models, validation, and governance, policies and controls. This piece is about the third.
In substance, the guidance places ultimate responsibility for model risk with the board of directors. The board should understand the level of model risk the bank carries, make sure it stays within the bank's tolerance, and hold senior management to account for the framework. Senior management turns that into policies, assigns roles, keeps a complete inventory of models, funds validation and makes sure internal audit tests whether the framework works.
A recurring idea is effective challenge: critical review by people who are competent, independent of the model's development and have enough influence to force change. Governance is what gives challenge its teeth.
Who does it apply to?
The guidance applies to banking organisations supervised by the agencies that issued it, and it shapes expectations well beyond them. Whether a given agent falls within its definition of a model is a separate question, covered in is an AI agent a model under SR 11-7 and does SR 11-7 apply to AI agents.
Our view: any agent whose outputs inform a decision or produce one should be in the model inventory. Arguing an agent out of scope rarely survives an examination and leaves the board blind to risk it is meant to own.
What changes when the model is an agent that acts?
Traditional model governance assumes a person sits between the model's output and the action. A credit score informs a decision; a person decides. Board reporting therefore focused on model performance.
An agent removes that person for some actions. It retrieves data, decides what to do next and acts, sometimes across several systems. Three governance consequences follow.
First, the board's risk tolerance must cover authority, not only accuracy. A board should decide which categories of action agents may take without a person, and at what scale. Depth is a practical way to express it: drafting for a person to send is depth 2; moving money is depth 4.
Second, aggregate model risk grows. SR 11-7 already asks banks to consider risk from interacting models. Agents that call other agents, or share sources, can fail together, so the inventory must record dependencies.
Third, human oversight becomes a control that needs evidence. Our glossary entry on what human oversight is sets out why approvals without review do not count.
How do the board and senior management meet the guidance for agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Board sets model risk tolerance | Board-approved limits on agent authority by action type and depth | Board minutes and a written autonomy statement |
| Board is informed of model risk | Quarterly report: agents by depth, certificate status, incidents, overrides | Board packs and recorded questions |
| Senior management maintains policies | Agent policy covering approval, validation, change and retirement | Approved policy with owners and review dates |
| Complete model inventory | Every agent and capability listed with owner, depth, sources and dependencies | Inventory extract reconciled to production |
| Independent validation | Agents certified on the bank's own cases by staff independent of the builders | Validation reports with pass floors and findings |
| Effective challenge | Validators can block release or force depth reduction | Records of blocked releases and resolved findings |
| Ongoing monitoring | Every action checked against bounds and filed with its verdict | Append-only action record and drift reports |
| Internal audit review | Audit tests the agent framework, not just individual agents | Audit reports and management responses |
| Vendor models governed | Vendor agents validated on the bank's use, with vendor evidence reviewed | Vendor due diligence and in-house test results |
The autonomy statement in the first row is the control most banks lack. Without it, the board cannot say whether an agent that settles small disputes on its own is inside or outside tolerance.
What will examiners and auditors ask?
- Show us the board's statement of risk tolerance for agents. When did the board last discuss it?
- Which agents in the inventory act without a person approving each action?
- How does validation for an agent differ from validation for a scoring model?
- Who can stop an agent from going live, and has that ever happened?
- How do you know the agent in production is the version that was validated?
- How do you monitor agents that rely on a vendor's underlying model?
For a fuller list of board-level questions, see what a board should ask about AI agents.
What does SR 11-7 governance not cover?
It does not tell a bank how much autonomy is acceptable; that is for the board. It does not prescribe validation tests for agents, which are covered in our piece on SR 11-7 model risk management for AI agents. It does not address fair lending, privacy or consumer protection, which sit with other laws such as ECOA and GLBA. And it does not apply directly outside its supervised population, although other supervisors, such as the PRA with SS1/23 and OSFI with Guideline E-23, set comparable expectations. For the control model behind all of this, see agent governance as a bound.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Questions people ask
Does the board need to approve every AI agent?
No. The board sets tolerance and policy and receives reports. Senior management approves individual agents within that policy, escalating to the board anything that would exceed tolerance, such as a first agent permitted to move money.
How technical should board reporting on agents be?
Not very. Directors need to know which agents act on their own, whether they are certified, what went wrong and what changed. Technical detail belongs in validation reports that the board can request.
Who provides effective challenge for agents?
Model validation, risk management and internal audit, each independent of the team that built or bought the agent. Business users can challenge too, but they are not independent of the outcome.