The short answer. Banks struggle to defend AI-assisted fraud and AML decisions because the evidence trail stops at the AI. An analyst can show the alert and the final decision, but not which policy or procedure the AI relied on when it summarized the case or answered the analyst’s question. The April 2026 interagency model risk guidance made this harder, not easier: it explicitly leaves generative AI out of scope, so each bank has to set its own controls for the AI its analysts use every day.
What changed in April 2026
On April 17, 2026, the Federal Reserve, the OCC and the FDIC issued revised guidance on model risk management (SR 26-2). It replaced SR 11-7, the 2011 guidance banks had built their model risk programs around for fifteen years, and SR 21-8, the 2021 interagency statement on model risk management for BSA/AML systems. It applies to banking organizations with more than $30 billion in total assets (Federal Reserve, SR 26-2, 2026).
The guidance keeps the fundamentals. It calls for documentation, a model inventory with enough information to understand model risks, validation that assesses reliability and limitations, and ongoing monitoring. For vendor models, sound practice includes understanding the model’s design and its development data.
Then it draws a line. “Generative AI and agentic AI models are novel and rapidly evolving. As such, they are not within the scope of this guidance” (SR 26-2, 2026). Banks are expected to apply their own risk management principles to tools the document doesn’t cover.
Your transaction-monitoring model sits inside a fifteen-year-old supervisory framework. The AI assistant your investigators use to read policies and draft case notes sits outside it. The examiner’s questions apply to both.
Why AI-assisted decisions are hard to defend
The answer has no source attached
Most generative AI tools return an answer, not a citation. When an analyst asks what the bank’s escalation threshold is for a structuring pattern and gets a fluent reply, there’s no record of which procedure document produced it, or whether that document was the current version. If the reply was wrong, the decision built on it is indefensible after the fact.
Access controls weren't there
IBM’s 2025 Cost of a Data Breach research found that 13% of organizations reported breaches of AI models or applications, and 97% of those lacked proper AI access controls. It also found that 63% of organizations had no AI governance policy or were still developing one, and that shadow AI added an average of $670,000 to breach costs (IBM and Ponemon Institute, 2025). In a bank, an AI tool that can reach every document can surface a suspicious activity file to someone who shouldn’t see it.
Nobody can reconstruct the path later
Examiners and internal audit ask questions months after a decision. “Which sources did the analyst rely on, and was the analyst allowed to see them?” is answerable only if the record was captured when the answer was generated.
What an examiner-ready AI workflow needs
- Every answer tied to a named internal source. If the AI can’t point to the procedure, the answer shouldn’t carry weight in a case file.
- Access rules enforced where the AI reads. An analyst’s AI should only reach the documents that analyst is cleared for.
- A way to pull outdated documents out of scope. Superseded procedures shouldn’t answer today’s questions.
- Visible disagreement. When two policies conflict, the analyst should see both, not a blend.
- A human decision on record. AI output supports the analyst. It doesn’t replace the sign-off.
- A record of how each source document was used. That’s what lets you answer the examiner’s question in minutes.
Where Mithra fits
Mithra® doesn’t score transactions or replace your fraud and AML models. It governs the generative AI your analysts use to work with the bank’s own documents: procedures, typology guides, policy manuals and prior guidance.
Documents go into permissioned Library spaces, and each one is tokenized on upload with a Token ID, a timestamp, its owner and its metadata. Space membership decides whose questions can reach which documents, so an AML investigator’s questions reach AML material and nothing else they aren’t cleared for. Every answer carries a Trusted, Partial or Untrusted label, or General Knowledge when no internal document applies, and an answer with no citation to an internal document can never be labeled Trusted. View Evidence opens each cited source with its Token ID and last-verified timestamp.
Admins control scope from the Guardrail Center. They can exclude superseded documents by metadata value, block restricted topics, require a minimum number of citations and, with “Use only your data sources” turned on, have Mithra decline questions the bank’s documents can’t answer. When two documents disagree, Mithra shows the evidence from both. Each document’s Data Lineage view logs when it was tokenized and each time it was cited or viewed as evidence.
Two limits worth stating plainly. Mithra doesn’t fact-check what’s inside a document; it shows which document an answer relied on, so owners can keep the baseline current. And a Trusted label means well supported by your documents, not verified fact. The analyst still makes the call, with the source one click away.
Frequently asked questions
Does the 2026 model risk guidance cover generative AI?
No. SR 26-2 states that generative and agentic AI models are not within its scope. Banks are expected to apply their own risk management principles to those tools.
What replaced SR 11-7?
SR 26-2, issued April 17, 2026 by the Federal Reserve, OCC and FDIC. It also superseded SR 21-8 on model risk management for BSA/AML systems.
How can a bank make AI-assisted AML decisions defensible?
Tie every AI answer to a named internal source the analyst was cleared to see, keep a record of how those sources were used, and keep a human decision on the case. The evidence has to be captured when the answer is generated, not rebuilt later.
About Mithra Technologies Inc.
Mithra Technologies Inc. is building the trust infrastructure for enterprise AI. Centered on its cryptographic Single Source of Truth (SSOT®), Mithra enables organizations to establish verifiable AI through trusted enterprise content, cryptographic proof, complete data lineage, provenance and governance. Designed to work across any AI model, Mithra helps enterprises deploy AI that is more transparent, explainable, auditable and resilient while reducing the risks associated with misinformation, unauthorized data usage, cyber threats and the emerging quantum era.