Bank AI Model Inventory: How to Build One for SR 26-2 Compliance

Defining an SR 26-2-Ready Model Inventory

A model inventory is the authoritative record of every model, algorithm, and AI agent operating in a bank's environment, along with the metadata needed to assess and monitor its risk. Long-standing model risk management guidance, most notably SR 11-7 and its OCC counterpart, OCC Bulletin 2011-12, established the expectation that banks maintain a comprehensive inventory, apply risk-based tiering to each entry, and revalidate models on an ongoing basis rather than at a single point in time. SR 26-2 extends similar oversight expectations to a broader and faster-moving category of systems, including generative AI applications and autonomous or semi-autonomous agents that can take actions, call tools, and change behavior after initial deployment. Because the specific text and scope of SR 26-2 should be confirmed against the current supervisory letter, this guide grounds its recommendations in the established model risk management principles that any AI-specific inventory expectation is likely to build on.

Why Generative and Agentic AI Change What the Inventory Must Capture

Traditional model risk inventories were built around static, versioned statistical models, such as a credit scoring model or a fraud detection model. Each entry typically recorded intended use, development data, validation status, and a fixed set of inputs and outputs. Generative and agentic AI systems do not fit this pattern cleanly. A large language model deployed as a customer-facing assistant may be reconnected to new data sources or granted new tool access weeks after its initial approval. An agentic system may be authorized to call external services, execute transactions, or hand off tasks to other agents, with the scope of that authority changing as business needs evolve. An inventory that only records design-time documentation misses these runtime changes and cannot support the ongoing oversight that model risk management guidance expects. Extending the inventory schema to capture permission scope, tool and plugin access, and logged agent behavior is necessary for entries to remain accurate over time.

Keeping the Inventory Current After Deployment

• Trigger mandatory inventory updates whenever a model is retrained, re-scoped, or granted new tool or data access
• Feed runtime telemetry, including logged agent actions and tool calls, back into the inventory rather than relying solely on point-in-time documentation
• Apply runtime policy enforcement to constrain agent permissions to what was documented and approved, reducing drift between the inventory record and actual behavior
• Maintain audit logging of agent actions so inventory entries can be traced to underlying evidence during examiner or internal audit review
• Review permission changes against the principle of least privilege before updating the inventory record

Governance, Ownership, and Practical Tradeoffs

An inventory is only as useful as the evidence behind it. Model risk management guidance has long expected that inventory entries connect to validation reports and monitoring logs, not just summary metadata, so that an examiner or internal audit function can trace an oversight claim back to supporting documentation. For AI systems, this means entries should link to logged runtime behavior, not only design documents, since agent actions and prompt-response pairs show what a system is actually doing in production. Clear accountability, an assigned owner, validator, and approver for each entry, remains a recurring expectation across model risk frameworks and should be a required field rather than an optional one. Building a more granular schema improves audit readiness but increases the maintenance burden on model owners and engineering teams; adding fields without automating data collection from production systems risks the same staleness problem as spreadsheet-based inventories. Because SR 26-2 specifics could not be independently verified at the time of writing, governance teams should confirm the citation's exact scope with legal and compliance counsel, while treating SR 11-7's ongoing monitoring and revalidation expectations as the baseline any AI-specific update is likely to extend.