For most bank uses, yes. SR 11-7 is US interagency guidance on model risk management, and examiners apply it to any quantitative method that turns inputs into estimates used for decisions. An agent that scores, classifies, forecasts or recommends fits that description. One that only drafts routine text may not, but the bank should decide and record it.
Who does SR 11-7 apply to?
SR 11-7 is supervisory guidance issued by the US federal banking agencies, and it is used in examinations of banking organisations they supervise. It is guidance, not a statute, but examiners expect banks to meet it in proportion to their size, complexity and use of models.
In substance, it expects:
- A broad model definition, covering methods that process inputs into quantitative estimates, including those that rely partly on judgement.
- Sound development, implementation and use, with documented purpose, data, assumptions and limits.
- Validation by people with the competence and independence to provide effective challenge: conceptual soundness, ongoing monitoring and outcomes analysis.
- Governance: board and senior management oversight, policies, a model inventory and clear roles.
- Vendor models held to the same standards, even where the bank cannot see inside them.
Whether a particular agent meets the definition is examined in is an AI agent a model under SR 11-7.
This is an explanation of the rules, not legal advice; confirm how they apply to you with counsel.
What changes when the model is an agent that acts?
SR 11-7 describes a model in three parts: inputs, processing and reporting. A traditional model's reporting component hands an estimate to a person, who decides what to do. An agent often skips that hand-off. Its output is the action: an account flagged, a payment queued, a limit changed.
That shifts two things. First, the model's use and its output merge. SR 11-7 warns that model risk comes from fundamental errors and from incorrect or inappropriate use. With an agent, inappropriate use can happen inside the agent, when it applies its reasoning to a case it was not built for. Second, the vendor model problem deepens. The underlying model belongs to a provider, can change without notice, and cannot be fully inspected. SR 11-7's expectation that vendor models are validated through testing and outcomes analysis becomes the main line of defence.
What may the agent do, and what stays with a person?
- Agent may: gather data, produce estimates and recommendations, draft documentation for validators.
- Agent may, within limits: take routine actions where outcomes are monitored and reversible, such as classifying alerts for review queues.
- Stays with a person: credit decisions and adverse actions, overrides of model outputs, changes to the agent's purpose, approval of validation findings, and acceptance of the agent's use after material changes.
An illustrative example
Take a US regional bank that deploys an agent to prepare commercial loan risk ratings. The agent reads financial statements, covenant reports and relationship notes, and proposes a rating with its reasons. A credit officer accepts or changes it.
The bank's model risk team concludes the agent is a model: it turns inputs into a rating that drives pricing, reserves and approval authority. It goes into the inventory, tiered high. Validation checks conceptual soundness (does the agent follow the rating methodology?), tests it on past loans with known outcomes, and sets ongoing monitoring of how often officers change its rating. When the provider releases a new model version, the bank reruns the validation set before the agent is used again. How such an entry is recorded is covered in SR 11-7 model inventory for AI agents.
Which controls meet SR 11-7 for agents?
| SR 11-7 expectation | Agent control | Evidence |
|---|---|---|
| Inventory | Entry per agent with purpose, sources, tools and owner | Inventory record |
| Development and use | Documented brief and limits; use restricted to tested purpose | Brief, scope bounds |
| Validation and effective challenge | Independent testing on the bank's own historical cases | Validation report |
| Ongoing monitoring | Override rates, outcome drift, model version changes | Monitoring dashboards and alerts |
| Vendor models | Change notice terms; revalidation on provider updates | Contract, test reruns |
Banks with European operations face separate duties; compare whether the EU AI Act applies to AI agents. The cross-regulation map is in AI agent regulation, mapped to controls.
Questions people ask
Does SR 11-7 apply to non-bank fintechs?
Not directly. It is aimed at supervised banking organisations. A fintech partnering with a bank may find the bank applies SR 11-7 standards to the fintech's agents through its third-party oversight.
Is a general chatbot a model under SR 11-7?
A chatbot that answers questions from approved content usually is not, though it needs other controls. If it begins to estimate, score or recommend in ways that drive decisions, reassess it.
What is effective challenge for an agent?
Critical review by people who are independent of the agent's build, competent to assess it, and able to change or stop its use. Testing on the bank's own cases is central to it.