Control 3.8: Model Risk Management Alignment (OCC 2011-12 / SR 11-7) — Troubleshooting
Common issues and resolution steps for model risk management compliance related to Copilot.
Common Issues
Issue 1: Model Inventory Entry Incomplete or Outdated
- Symptoms: OCC examiners flag the Copilot model inventory entry as lacking required fields or containing stale information.
- Root Cause: Model inventory was not updated after Copilot feature updates, or initial entry used incomplete documentation.
- Resolution:
- Review the latest Microsoft Copilot AI documentation from the Service Trust Portal.
- Update all model inventory fields with current information.
- Establish a trigger-based update process tied to Microsoft's Copilot release cycle.
- Assign a dedicated model owner responsible for ongoing inventory maintenance.
Issue 2: Insufficient Performance Monitoring Data
- Symptoms: Model validation reviewers report that performance metrics are too limited to assess model quality effectively.
- Root Cause: Feedback collection may be limited by low user engagement, or the audit log may not capture all relevant quality indicators.
- Resolution:
- Encourage Copilot users to provide feedback (thumbs up/down) on responses.
- Supplement audit log data with direct user surveys conducted quarterly.
- Implement structured output quality reviews by sampling Copilot-generated content.
- Consider deploying additional monitoring through Viva Insights Copilot metrics.
Issue 3: Vendor Documentation Not Available from Microsoft
- Symptoms: Required third-party risk documentation (SOC reports, AI assessments) is delayed or unavailable from Microsoft.
- Root Cause: Microsoft's reporting cycle may not align with the firm's assessment schedule, or access to the Service Trust Portal may be restricted.
- Resolution:
- Verify Service Trust Portal access for the compliance team.
- Contact the Microsoft account team to request documentation availability timelines.
- Use alternative evidence such as Microsoft's published AI principles and compliance certifications as interim documentation.
- Document the gap and timeline in the vendor risk assessment.
Issue 4: Model Inventory Classification Debate — Tier Assignment
- Symptoms: Internal model risk management team and business stakeholders disagree on the appropriate risk tier for Copilot (Tier 1 vs. Tier 2 vs. Tier 3), particularly when Copilot usage spans both internal productivity and limited client-facing activities.
- Root Cause: AI-generated content tools do not fit neatly into traditional model tiering frameworks designed for quantitative financial models. Use cases may span multiple risk levels.
- Resolution:
- Apply the higher tier when use cases span multiple risk levels — if any approved use case involves client-facing activities or lending workflows, do not classify as Tier 3.
- Reference OCC Bulletin 2011-12 and SR 11-7 on the materiality of model outputs for tier determination. The key question: could Copilot outputs in the approved use cases directly influence decisions affecting clients or regulated activities?
- Document the full approved use-case register before tier assignment — tier follows use cases, not general capability.
- Seek guidance from the firm's model risk management committee and document the outcome.
- If Copilot usage evolves to include higher-risk activities, update the tier and associated governance promptly.
Issue 5: Proportionality Determination for Mid-Size Institutions
- Symptoms: A mid-size institution (not a community bank, but not a large complex organization) is uncertain whether OCC Bulletin 2025-26 proportionality applies and what MRM tier to select.
- Root Cause: OCC Bulletin 2025-26 was written with community banks in mind, but the proportionality principle applies more broadly. Mid-size institutions may have legitimate grounds for a simplified MRM approach depending on Copilot usage scope.
- Resolution:
- Assess the actual risk profile: what use cases are approved, what is the volume and materiality of Copilot-assisted decisions, and what is the institution's overall complexity? Proportionality should reflect actual risk, not just asset size.
- Document the proportionality analysis in writing — even if the institution is not a community bank, a documented and reasoned proportionality determination is defensible in an examination.
- If Copilot is used only for internal productivity (meeting summaries, document drafting, research support) with no direct client impact, a simplified MRM approach aligned with OCC Bulletin 2025-26 proportionality is supportable. Cite the bulletin and the usage scope limitation in the model inventory.
- If examiners challenge the proportionality determination, present the documented rationale — the quality of the documentation matters as much as the tier selected.
- For institutions uncertain about the correct approach: consult with outside counsel or regulatory advisory specialists familiar with OCC model risk management examination practice.
Issue 6: Copilot Model Status Under SR 11-7 Disputed
- Symptoms: An examiner or internal legal team questions whether M365 Copilot meets the SR 11-7 definition of a "model," creating uncertainty about whether any MRM framework is required.
- Root Cause: SR 11-7 defines models as quantitative methods applying "statistical, economic, financial, or mathematical theories, techniques, and assumptions." Copilot is an LLM — not a traditional quantitative model — and there is genuine ambiguity about how the definition applies.
- Resolution:
- The consensus regulatory and industry view is to treat Copilot as a model subject to MRM (at some tier), rather than arguing it falls outside the definition. This is the safer compliance posture.
- Document the model status determination in the inventory: "M365 Copilot is classified as a vendor-provided general-purpose LLM. It straddles the SR 11-7 model definition. The institution has elected to include it in the model inventory at [Tier X] as a prudent governance approach, consistent with the 2023 Interagency Guidance on AI."
- Reference the 2023 Interagency AI Guidance, which confirms that existing risk management frameworks (including MRM) apply to AI technologies.
- Avoid the position that Copilot requires no governance — this is the posture most likely to draw examiner scrutiny.
Diagnostic Steps
- Review model inventory: Verify all required fields are populated and current, including tier classification with documented rationale.
- Check performance metrics: Run the feedback and quality metrics scripts to confirm data availability.
- Validate vendor docs: Verify that all Microsoft documentation is current and accessible.
- Assess monitoring coverage: Confirm that performance monitoring covers all approved Copilot use cases.
- Review proportionality documentation: Confirm that community banks or mid-size institutions applying a simplified MRM approach have documented rationale citing OCC Bulletin 2025-26.
Escalation
| Severity | Condition | Escalation Path |
|---|---|---|
| Critical | OCC/Fed examination finding on model risk | Chief Risk Officer + Model Risk Committee |
| High | Model validation overdue by more than 30 days | Model Risk Management team lead |
| High | Tier classification dispute with examiner | Chief Risk Officer + outside counsel |
| Medium | Performance metrics showing degradation trend | Model owner + IT monitoring team |
| Medium | Proportionality determination challenged | Model Risk Committee + compliance officer |
| Low | Documentation updates needed | Assign to model owner for next update cycle |