Skip to content

Control 2.6: Model Risk Management - Portal Walkthrough

This playbook provides portal-based configuration guidance for Control 2.6.


Overview

This walkthrough guides you through implementing Model Risk Management (MRM) governance for AI agents, including model classification, inventory management, validation configuration, and performance monitoring.


Part 1: Agent-as-Model Classification

Model Definition per OCC 2011-12

A model is a quantitative method, system, or approach that applies statistical, economic, financial, or mathematical theories, techniques, and assumptions to process input data into quantitative estimates.

Agent Classification Criteria

Criteria Model Treatment Example
Makes decisions affecting customers Yes - Tier 1 Credit recommendation agent
Provides financial calculations Yes - Tier 1/2 Investment calculator agent
Influences risk assessments Yes - Tier 1 Risk scoring agent
Customer-facing recommendations Yes - Tier 2 Product recommendation agent
Information retrieval only No FAQ/knowledge base agent
Internal productivity No IT help desk agent

Classification Form

Agent Model Classification Form

Agent Name: [Name]
Agent ID: [ID]
Business Owner: [Owner]

Classification Decision:
[ ] Model (requires MRM governance)
[ ] Non-Model (standard agent governance)

Justification:
[Explain why agent does/doesn't qualify as model]

Model Tier (if applicable):
[ ] Tier 1 - High Risk (material business impact)
[ ] Tier 2 - Medium Risk (significant but limited impact)
[ ] Tier 3 - Low Risk (minimal business impact)

Approved by: _________________ Date: _________
Model Risk Manager

Part 2: Model Selection Governance

Portal Path: Copilot Studio > [Agent] > Settings > AI capabilities > Model

Model Options

Model Availability Capability Level FSI Recommendation
GPT-5 GA (Nov 2025) Production-ready, default Approved for all zones
GPT-5.2 Preview Experimental capabilities Zone 1 only; Zone 2/3 disabled
GPT-4o GA Previous generation Approved for all zones
Custom models Varies Organization-specific Requires MRM validation

Configuration Steps

  1. Navigate to Copilot Studio > Select agent
  2. Click Settings > AI capabilities
  3. Review Model selection
  4. Document model choice in agent inventory

Zone-Specific Model Governance

Zone Allowed Models Experimental Models Documentation
Zone 1 GPT-5, GPT-4o Allowed for evaluation Minimal
Zone 2 GPT-5, GPT-4o Disabled Standard MRM
Zone 3 GPT-5, GPT-4o (approved only) Disabled Full MRM with validation

Part 3: Model Inventory Setup

Create Inventory Registry

Option 1: Dataverse Table - Power Platform Admin Center > Create new Dataverse table

Option 2: SharePoint List - SharePoint > Create list with inventory columns - Enable version history for audit trail

Required Inventory Fields

Field Description
Model ID Unique identifier
Model Name Agent/model name
Model Tier 1, 2, or 3
Business Purpose What the model does
Model Owner Business owner
Model Developer Technical owner
Primary Users Who uses the model
Data Inputs Data sources used
Model Outputs Decisions/recommendations
Implementation Date Go-live date
Last Validation Date of last review
Next Validation Due Scheduled review date
Performance Status Green/Yellow/Red

Part 4: Agent Card Template

Purpose

Agent Cards provide standardized documentation for each AI agent, capturing capabilities, limitations, training sources, and governance status.

Agent Card Requirements by Tier

Requirement Tier 3 (Low) Tier 2 (Medium) Tier 1 (High)
Agent Card Required Recommended Required Required
Full Limitations Section Optional Required Required
Bias Testing Results Optional Required Required
Performance Benchmarks Basic Standard Comprehensive
Change History Summary Detailed Full audit trail
Review Frequency Annual Semi-annual Quarterly
External Validation No Recommended Required

Part 5: Validation Framework

Validation Tiers

Model Tier Validation Requirements Frequency
Tier 1 Independent third-party Annual
Tier 2 Independent internal team Annual
Tier 3 Self-assessment + review Biennial

Validation Scope

Conceptual Soundness: - [ ] Model design is appropriate for intended use - [ ] Methodology is theoretically sound - [ ] Assumptions are reasonable and documented - [ ] Limitations are clearly stated

Data Quality: - [ ] Data sources are appropriate - [ ] Data quality is acceptable - [ ] Data preprocessing is appropriate - [ ] Data is representative of use cases

Output Analysis: - [ ] Outputs are accurate and reliable - [ ] Performance meets expectations - [ ] Outputs are consistent over time - [ ] Edge cases handled appropriately

Implementation Verification: - [ ] Model implemented as designed - [ ] Controls are effective - [ ] Documentation is complete - [ ] Users are properly trained


Part 6: Performance Monitoring

Portal Path: Power Platform Admin Center > Analytics > Copilot Studio

Performance Metrics

Metric Description Threshold
Response Accuracy Correct responses / Total >95%
User Satisfaction CSAT score >4.0/5.0
Fallback Rate Escalations to human <10%
Response Time Average response latency <2 seconds
Error Rate Failed conversations <2%
Bias Indicators Demographic disparity <5% variance

Status Thresholds

  • Green: All metrics within thresholds
  • Yellow: 1-2 metrics slightly below threshold
  • Red: Multiple metrics below threshold or critical failures

Part 7: Change Control

Change Classification

Change Type Description Governance
Material Change Affects model outputs significantly Full revalidation
Non-Material Change Minor updates, bug fixes Abbreviated review
Emergency Change Critical fix Expedited process
Prompt Change System/topic prompt updates See prompt review checklist

Prompt Change Categories

Change Type Description Review Level
System Prompt Core agent instructions Full review + testing
Topic Prompt Topic-specific logic Standard review
Fallback Prompt Unknown input handling Standard review
Grounding Instructions Knowledge source usage Full review + testing

Part 8: Multi-Agent Governance

Zone Restrictions

Zone Multi-Agent Allowance
Zone 1 No agent-to-agent delegation allowed
Zone 2 Delegation within same zone only; max depth 2
Zone 3 Cross-zone delegation with explicit approval; max depth 3

Multi-Agent Risk Considerations

  • Treat orchestrated systems as single model for validation
  • Document complete delegation chain
  • If any agent in chain is Tier 1, treat entire chain as Tier 1
  • Validate end-to-end behavior, not just individual agents

Back to Control 2.6 | PowerShell Setup | Verification Testing | Troubleshooting


Updated: January 2026 | Version: v1.2