Effective challenge of an AI agent means people who are independent, competent and given real authority examine its brief, sources, tools, limits and behaviour, and can force changes. Under SR 11-7 the test is whether challenge changes something. For an agent, the clearest sign is that challenge reduced or conditioned what the agent may do.
What does SR 11-7 say effective challenge is?
The interagency model risk guidance SR 11-7 makes effective challenge a guiding principle of model risk management. It describes it, in substance, as critical analysis by objective, informed parties who can identify a model's limitations and assumptions and produce appropriate changes. The guidance says effective challenge depends on three things together:
- Incentives: those challenging must have reason to do so, supported by separation from development, appropriate compensation and governance that rewards challenge.
- Competence: they need the technical knowledge and modelling skill to understand the model, and knowledge of the business it serves.
- Influence: they must have the authority and standing to make their findings matter, backed by senior management commitment.
Challenge is not confined to validation. It applies throughout development, implementation and use, and model users can be a source of it. For whether agents are within SR 11-7's scope, see is an AI agent a model under SR 11-7.
Who does it apply to?
Banking organisations supervised by the Federal Reserve and the OCC, with the FDIC applying the same approach. Many lenders and fintechs adopt it because bank partners and investors expect it. For the short answer on scope, see does SR 11-7 apply to AI agents.
What does effective challenge look like for an agent?
Agents change what each of the three conditions requires.
Incentives. Agents are often built by business teams under pressure to clear backlogs. The people who built the agent want it to act. Challenge must come from people whose success does not depend on the agent going live with full authority.
Competence. Challenging a credit scorecard needs statistical skill. Challenging an agent needs two kinds of skill at once. Technical: how language models fail, how prompt injection works, how tool permissions and limits are enforced. Business: what a correct outcome looks like in the process the agent performs, whether that is complaint handling, payment exceptions or arrears contact. A challenger with only one of these will miss half the risk. In practice, teams pair a model validator with a senior practitioner from the business process.
Influence. Challenge that cannot change the agent's authority is commentary. The challenger's findings must be able to keep a capability at drafting, set a lower limit, require more testing, or stop deployment. If findings routinely lead to an action plan while the agent keeps acting, influence is missing.
Where should challenge happen in an agent's life?
At design. Is the brief's description of the task correct against policy? Are the sources the right ones? Are the tools the minimum needed?
At validation. Do test cases reflect real use? Are the pass marks set high enough for the risk? Do the controls actually stop the agent exceeding its limits?
In use. Reviewers of held actions are a continuous source of challenge. When a reviewer overrides an agent's draft, that override is challenge in its purest form, and it should flow back to the agent's owner.
At change. When the model provider updates a model or the business adds a tool, someone independent should ask what has changed and whether authority should drop until re-tested.
Take an illustrative case. A US bank's operations team builds an agent that resolves card disputes under a set amount. Validation, staffed by a model validator and a disputes specialist from another region, finds the agent mishandles disputes involving recurring merchant charges. The validators' report keeps those disputes at drafting and asks for more test cases. The business objects; the head of model risk upholds the finding. Six weeks later, after retesting, the restriction lifts. The record shows challenge, response and outcome.
How do you meet the effective challenge principle for agents?
| Requirement | Control | Evidence it leaves |
|---|---|---|
| Incentives | Challengers independent of the agent's build team and business targets | Reporting lines; role descriptions |
| Competence | Challenge teams combine AI technical skill and business process expertise | Reviewer profiles; training records |
| Influence | Findings can restrict authority, block deployment or require retesting | Policy granting authority; examples of restrictions |
| Challenge at design | Review of brief, sources and tools before build completes | Design review notes |
| Challenge at validation | Independent validation on the bank's own cases | Validation reports with findings |
| Challenge in use | Reviewer overrides tracked and reported to the agent's owner | Override reports |
| Challenge at change | Independent review of model, brief and tool changes | Change review records |
| Escalation | Disputes between business and challengers escalated and resolved | Escalation log |
For validation specifically, see SR 11-7 model validation for agents.
What will examiners ask?
- Who challenged this agent, and how are they independent?
- What skills did the challengers bring, technical and business?
- Show us a finding that changed what an agent may do.
- How do reviewer overrides reach the agent's owner?
- What happens when the business disagrees with a finding?
What does the principle not cover?
It does not prescribe who must challenge or how often. It does not, on its own, satisfy validation or monitoring expectations, though it runs through both. It does not address conduct or data protection risks directly, which other rules govern. For the wider set of binding rules, see our guide to AI agent regulation mapped to controls.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
Where Pryme fits
Pryme makes challenge bite at the point of action. A finding can lower a capability's depth or tighten a customer limit, and the minimum chain enforces it on the next action. Held items carry the facts and sources a reviewer needs to disagree, and every override is filed in the append-only record.
Questions people ask
Can model users provide effective challenge?
Yes. SR 11-7 recognises that users can challenge models. For agents, staff who review held actions see the agent's work every day; their overrides are valuable challenge if they are tracked and acted on.
Is a vendor's own review a form of effective challenge?
It may be useful input, but the vendor is not independent of its product and does not see your use. The bank's own challenge is still needed.
How do you show effective challenge happened?
With records of findings that changed something: a restriction on authority, a fixed source, a revised brief or a blocked deployment. Reports with no resulting change suggest challenge lacked influence.