SR 26-2 and Bank Customer Service Chatbots: Scope and Evidence

What changed and what SR 26-2 now says

SR 26-2, the revised interagency supervisory guidance on model risk management, was issued on April 17, 2026 by the Federal Reserve, in coordination with the OCC and FDIC. It serves in part to rescind prior model risk management guidance, which the letter references as SR 11-7, and it reflects supervisory experience and industry feedback since SR 11-7 was issued. A scoping memo that still cites SR 11-7 as the operative text is out of date as of November 2026.

Two features matter for chatbots. First, the guidance is risk-based and tailored, commensurate with a banking organization's size, complexity and extent of model use. It is expected to be most relevant to banking organizations with over $30 billion in total assets regulated by the Federal Reserve. Second, a Federal Reserve Board speech on May 1, 2026 said the agencies amended the guidance to clarify that it does not apply to generative or agentic AI, and that supervisors had expanded the scope of the previous guidance beyond its original purpose.

The question has shifted from whether a chatbot is a model under SR 11-7 to which architecture the chatbot is and which framework governs it. For tools and systems outside scope, SR 26-2 is understood to direct that the bank's risk management and governance practices guide the appropriate governance and controls. Out of scope does not mean uncontrolled. It means the bank carries the burden of explaining and evidencing its own control choices.

Scoping position by chatbot architecture

Classify each chatbot by three attributes: how it produces outputs, which customer data, account systems and tools it can reach, and who operates it. This framing is analytical, not guidance text. The table shows where each architecture lands and what the determination depends on.

• Rules-based: Decision trees or keyword databases trigger preset, limited responses. Whether the system meets the model definition is a classification decision to make against the guidance text and record, with the provision relied on.
• Retrieval-augmented: An open analytical question. If the system returns retrieved text deterministically, it resembles a rules-based system. If a generative model composes the answer, it falls toward the generative treatment. Document which behavior applies.
• Generative (LLM-based): Treated as outside SR 26-2 according to the May 2026 Federal Reserve Board speech. Governance and controls rest on the bank's own risk management practices.
• Vendor-hosted: Scope follows the output-generation method, not the hosting arrangement. Where a vendor model is covered, Section VII validation and monitoring expectations apply. Vendor generative AI sits with third-party risk management and internal governance.

Vendor-hosted chatbots and consumer-protection exposure

Section VII addresses vendor models the guidance covers. As quoted, it describes validation of vendor products by internal or outside parties, and ongoing monitoring and outcome analysis to assess whether the models are accurate, remain fit for purpose, and continue to be reliable, even where the bank does not receive the underlying code, data or methodology. For a covered vendor chatbot, the bank should therefore keep due diligence records, validation or testing results, and monitoring outputs itself, not rely on the vendor to hold them. Where the vendor system is generative, the same evidence still supports the bank's own governance and third-party risk management.

Scope outcome does not change consumer-protection exposure. The CFPB states that chatbots must comply with all applicable federal consumer financial laws and that institutions may be liable for violations. It identifies risks including inaccurate information, failure to recognize a consumer invoking federal rights, and failure to protect privacy and data. Its research also reported that each of the ten largest commercial banks had deployed chatbots and that banks were moving toward LLMs. A chatbot that touches complaints or disclosures needs clear ownership and an escalation path for rights invocations it may fail to recognize, whether or not SR 26-2 applies.

Interaction-level evidence for in-scope and out-of-scope chatbots

Whatever the scoping result, a reviewer assessing a chatbot that handles customer data, account information, complaints or disclosures will want to see how it behaved on actual interactions. Retaining interaction-level records is a suggested practice rather than a stated SR 26-2 requirement, but it is the practical way to support either a model risk management file or an alternative governance file.

Useful records for each session include which customer data and account systems were reached, which tools were called and with what outcome, which disclosure was presented, whether a complaint or rights invocation was recognized and routed, and which policy applied to each action. Pair these with the scoping determination, vendor due diligence and monitoring outputs, and the documented controls for out-of-scope systems.

Two tradeoffs deserve a decision. Detailed records can themselves contain customer data, so retention periods, access and data residency need defining up front. And the planned interagency request for information on model risk management and AI, including generative and agentic AI, may change expectations, so scoping determinations should carry a review date.

Enforcing policy in the execution path

Scoping documents describe intent. A reviewer will also ask what the chatbot was actually allowed to do. Trussed enforces policy in the execution path on individual interactions and tool calls: allow, redact, block, or escalate for human approval. Enforcement happens at runtime, not after the fact, so a rule such as redacting account identifiers or routing a complaint to a person is applied to the interaction itself.

That enforcement generates audit-ready evidence of each decision, which supports the documented controls in a model risk management or governance file. Trussed does not replace the scoping determination, validation or vendor due diligence. It can be deployed self-managed in the customer cloud, with configurable retention and data residency.

Insurance is the regulated financial sector where Trussed is deepest. Its general production record includes 18+ months of continuous production in a regulated enterprise, 99.95% uptime on live production traffic, and 100+ models under one policy plane.