Proven with Industry Leaders
What our model risk management consulting covers
Most AI pilots that stall in model risk review stall for the same reason. The use case works, and nobody can show how it is controlled. We close that gap in four parts.
| What the risk team needs | What we deliver | The evidence it produces |
|---|---|---|
| A complete view of AI in use | An AI inventory with risk tiering | Inventory entries, owners, and tiers |
| Confidence the system behaves as intended | Validation of GenAI and agent use cases | Validation reports and findings |
| A defensible position under SR 26-2 | An evidence pack mapped to the guidance | A pack the committee can review |
| Control of what agents do in production | Agent identity, decision logs, and a stop control | System records, not documents |
The AI inventory examiners ask for
Under SR 26-2 the inventory is the central tool for risk tiering and for scaling validation and oversight to each model's materiality.
Most banks have a model inventory. Few have one that captures generative AI features inside vendor products, internal copilots, and agents that call tools.
We build or extend the inventory to cover
- Every AI use case, including AI embedded in third-party software
- The owner, purpose, data sources, and systems each one touches
- A risk tier that drives validation depth and monitoring frequency
- Agents and the tools, APIs, and data each agent is allowed to reach
The SR 26-2 evidence pack
The evidence pack is what a model risk committee reviews and what an examiner asks to see. For each in-scope AI use case it contains eight parts. For what the guidance itself says, read SR 11-7 and SR 26-2 explained.
- 01 The inventory entry and risk tier, with the rationale
- 02 The use case, its limits, and who is accountable for outcomes
- 03 Data sources, lineage, and access controls
- 04 Validation scope, results, and open findings
- 05 Ongoing monitoring: what is measured, how often, and the thresholds
- 06 Third-party and vendor model responsibilities. Reliance on a vendor does not transfer responsibility.
- 07 Agent controls: identities, permitted tools, decision logs, and the stop procedure
- 08 Change history and approvals
Agent controls: identity, decision logs and a stop control
Governance is decided in committee and enforced at runtime. For agents, four questions decide whether a control is real. Each answer has to come from a system record.
Which agent took the action?
An identity registered for that agent, not a shared service account.
On whose authority?
A delegation record that carries both the agent and the person it acted for.
What did it touch, and under which policy?
A gateway access log tied to the policy decision that allowed the call.
Can you stop it?
A revocation path that works in seconds, tested before go-live.
We implement the controls that produce those records in the gateway and identity layer. See AI gateway and agent governance for how that works, and AI and agent governance on Choreo Labs if Choreo Labs is your platform. Our practitioner writing on the topic is collected under Agentic AI Governance & Model Risk. Outside banking, the same method is set out under AI governance consulting.
AI audit services and independent review
Internal audit increasingly asks whether the AI governance program itself is sound, not only whether individual models are.
We review the program against SR 26-2 principles and your own policies: inventory completeness, tiering, validation coverage, monitoring, and agent controls. The output is a findings report with a remediation plan.
See the data behind one AI use case
A model risk committee or examiner asks what data feeds a model, where it came from, and what happens to it on the way. Our team maps it for one use case with Cloudera Data Lineage, from source systems through pipelines to the model, and produces the data sources and lineage section of the evidence pack from system metadata.
Your lawyers write the policy. The proof that it holds comes from systems. We build that proof.
- Price: free for one closely defined use case. Up to $8K for a wider scope.
- Timing: about seven working days after access is set up. Three to run the tool, four to write the report.
- Your data stays put: only metadata leaves your environment.
AI model risk management, answered plainly.
SR 26-2 and model risk
Does SR 26-2 cover generative and agentic AI?
No. SR 26-2, issued on April 17, 2026 by the Federal Reserve, the OCC, and the FDIC, expressly leaves generative and agentic AI out of its scope. Banks govern those systems under their broader risk management frameworks, and responsibility for outcomes stays with the bank.
What did SR 26-2 replace?
It superseded SR 11-7, the 2011 model risk guidance, and SR 21-8, the 2021 interagency statement on model risk for BSA/AML systems.
Does SR 26-2 leave gaps for AI agents?
Yes. Agents are out of its scope, so there is no supervisory checklist for them. The gap is closed by the bank’s own framework: an identity for each agent, policy on every tool call, decision logs, and a tested stop procedure.
What is the difference between AI governance and model risk management?
Model risk management covers models within the guidance’s definition: validation, monitoring, and governance of quantitative models. AI governance is broader. It covers generative AI, agents, and AI in vendor products, plus the runtime controls on what AI systems can do.
Validation and inventory
Who does model validation for banks?
A party independent of the team that built the model: the bank’s own model validation function, an outside firm, or both. For AI models the validator also needs to test machine learning, generative AI, and agents, not only statistical models. Duczer East performs independent AI model validation for banks, and builds validation-ready evidence for the models we design.
Who can validate generative AI models under SR 26-2?
An independent party with the skills to test the system: your own validation team, a third party, or a firm like Duczer East working independently of whoever built the system. Because SR 26-2 does not prescribe a GenAI standard, the bank sets the validation standard and has to defend it.
Can a community bank get independent validation of its AI models?
Yes. Validation depth scales with the materiality of each model, so a smaller bank can validate its highest-risk AI use cases thoroughly without building a large validation function.
Who audits third-party AI vendors for a bank?
The bank remains responsible for vendor models. We review vendor AI as part of the inventory and validation work: what the vendor discloses, what the bank can test, and which controls the bank applies around the vendor system.
What does an AI data lineage map cost, and how long does it take?
For one closely defined AI use case it is free. A wider scope runs up to $8K. Our team runs Cloudera Data Lineage against the systems that feed the use case, and the report arrives about seven working days after access is set up. Only metadata leaves your environment.