Practical guide

Model-risk policy for AI-enabled work

A model-risk policy should define which quantitative models and model-like systems enter governance, how risk determines the required controls, who may develop, challenge, approve, use, limit, and retire them, and what records make those decisions reviewable. It should also state how the organization handles tools outside a sector rule’s formal scope without pretending that one policy covers.

By Two Prune

A model-risk policy should define which quantitative models and model-like systems enter governance, how risk determines the required controls, who may develop, challenge, approve, use, limit, and retire them, and what records make those decisions reviewable. It should also state how the organization handles tools outside a sector rule’s formal scope without pretending that one policy covers every technology or jurisdiction.

Establish purpose and scope

Policy objective: models are developed, acquired, changed, and used only for documented purposes and within risk-based limits; material uncertainty, performance deterioration, exceptions, and misuse trigger accountable review and action.

Define “model” for the organization and list adjacent systems that require another governance route. Classification should consider method, output, decision use, materiality, automation, data sensitivity, reversibility, and potential harm—not only the vendor’s label.

The 2026 U.S. interagency guidance describes risk-based model development, use, validation, monitoring, governance, inventory, documentation, and third-party oversight for banking organizations. It supersedes SR 11-7 and explicitly excludes generative and agentic AI from its model definition, while noting that organizations should govern tools outside scope appropriately. NIST’s AI Risk Management Framework provides broader voluntary outcomes for roles, documentation, oversight, monitoring, incidents, and third parties. These sources inform the policy structure below; they do not determine an organization’s legal obligations or show that its controls operate effectively. Federal Reserve: SR 26-2 revised guidance on model risk management NIST AI RMF Core

Sources: [1] [2]

Use the SCOPE policy structure

### S — Scope, classification, and inventory

The policy should require:

Unknown classification is not approval. Assign a temporary hold or restricted route until an accountable reviewer decides.

### C — Control requirements across the lifecycle

Map requirements to risk and use:

Risk-based tailoring should change the depth and frequency of controls, not erase ownership or evidence.

### O — Ownership, independence, and decision rights

Name roles rather than relying only on departments:

Define conflicts of interest and compensating review when perfect structural independence is impractical. An AI system cannot hold approval authority or accept residual risk.

### P — Performance, problems, and exceptions

Require monitoring plans to name populations, metrics, thresholds, source data, owners, frequency, and actions. Include data drift, performance deterioration, use outside scope, unexpected outcomes, vendor changes, user complaints, overrides, and control failures.

Every issue should have severity, impact, containment, decision, owner, due date, and closure test. Exceptions need scope, reason, approver, compensating steps, expiry, and trigger for reconsideration. Repeated exceptions should prompt policy or control review rather than automatic renewal.

### E — Evidence, enforcement, and review cadence

Specify the minimum record:

| Record | Minimum content | | --- | --- | | Classification | system, method, use, materiality, scope decision, reviewer, date | | Inventory | version, purpose, owner, users, data, dependencies, status, approved limits | | Challenge | tests, evidence, findings, disagreements, reviewer competence and independence | | Approval | exact version, permitted use, conditions, residual risk, approver, review date | | Monitoring | period, population, metric, threshold, result, breach response | | Issue or exception | impact, containment, authority, expiry, remediation, closure evidence | | Change or retirement | trigger, reassessment, communications, downstream action, archive |

Review the policy on a fixed cadence and when regulation, model use, technology, data, vendor terms, incidents, or organizational authority changes. Test a risk-based sample of operating records; policy acknowledgement alone is not evidence of operation.

  • a documented classification decision for in-scope models and adjacent AI systems;
  • a versioned inventory covering purpose, owner, users, methodology, data, dependencies, vendor components, status, and approved use;
  • materiality and risk tier criteria with examples and an escalation route;
  • reconciliation against deployment, procurement, workflow, and change records;
  • explicit treatment of systems under development, temporary tools, spreadsheets, vendor services, and retired models.
  • Intake: purpose, decision consequence, alternatives, data authority, and owner.
  • Development or acquisition: design rationale, assumptions, data suitability, testing, vendor evidence, and limitations.
  • Independent challenge: conceptual review, outcome analysis, benchmarks, sensitivity, boundary cases, and reviewer competence.
  • Approval and release: exact version, permitted uses, conditions, users, monitoring, and expiry or review date.
  • Use: user training, limitation disclosure, overrides, overlays, and feedback.
  • Monitoring: data and performance measures, thresholds, frequency, segmentation, and response procedure.
  • Change: triggers for reclassification, revalidation, renewed approval, and user communication.
  • Retirement: replacement, archived evidence, downstream dependencies, and access removal.
  • business owner for purpose, use, and consequence;
  • developer or supplier for design and technical evidence;
  • data owner for source authority and quality;
  • independent challenger with access, competence, and authority to disagree;
  • approver who accepts stated use and residual risk;
  • users responsible for operating within limits and reporting unexpected behaviour;
  • control owner for inventory, monitoring, findings, exceptions, and review;
  • audit or assurance role where applicable, without substituting for management controls.

Operating procedure

Requester submits purpose, decision, data, method, users, and intended release.

  • Control owner classifies the system and assigns risk tier or an adjacent governance route.
  • Owner completes required development or acquisition evidence.
  • Independent reviewer performs the risk-tier challenge and records disagreements.
  • Approver issues approve, limit, repair, hold, or reject with conditions.
  • Release owner verifies version, access, monitoring, user communication, and evidence retention.
  • Users operate within limits and report unexpected behaviour.
  • Monitoring breaches and changes enter the issue process.
  • Control owner tests closure, exceptions, inventory accuracy, and review dates.
  • Accountable governance body reviews aggregate risk, overdue findings, and material exceptions.

Worked policy test: forecast used for a new product

A synthetic planning team uses an approved demand forecast for established products with at least eight weeks of history. A new launch has only two weeks of data, but a workflow change begins routing it through the same model.

Under SCOPE, the inventory’s approved-use boundary makes the launch an unapproved extension. The user cannot treat a technically successful output as authorization. The issue owner limits the model to established products and opens a change request.

The independent challenger tests alternative assumptions, sparse-history sensitivity, and monitoring segmentation. The challenge finds that an aggregate accuracy metric masks large launch-family errors. The monitoring plan therefore adds product-family reporting and a specific escalation condition.

An urgent planning deadline leads the business owner to request a four-week exception for a manually reviewed output. The exception names the exact products, overlay cap, reviewer, user warning, expiry, and stop condition. It does not expand the permanent approval.

At expiry, the control owner checks whether the launch-specific evidence supports a new use. If not, the model remains limited and the team uses an alternative planning method. The evidence packet contains classification, inventory boundary, change request, challenge, temporary exception, communications, monitoring rule, decision, and next review date.

This example demonstrates how the policy routes a scope change. It does not show that the forecast is accurate, that the exception is suitable elsewhere, or that the organization’s control programme is effective.

Failure and hold conditions

Hold release or continued use when classification is unresolved, purpose or version is unclear, required challenge is missing, validation does not match production use, monitoring cannot be reproduced, an exception lacks authority or expiry, a material breach has no response owner, or a closed finding lacks retest evidence.

Local legal, regulatory, accounting, safety, and sector requirements may add obligations or prohibit a use. Route those conclusions to qualified reviewers. Scope statements about policy design separately from claims about sampled operation or supported effect.

Two Prune can place model-risk policy work inside a realistic AI-enabled assessment and evaluate how a participant classifies systems, navigates evidence, challenges assumptions, manages exceptions, and communicates a defensible policy. It does not validate production models, certify controls, or make employment decisions.

Sources

Board of Governors of the Federal Reserve System: SR 26-2 revised guidance on model risk management

  • National Institute of Standards and Technology: AI Risk Management Framework Core

Sources: [1] [2]

Sources

  1. 1.Federal Reserve: SR 26-2 revised guidance on model risk management · www.federalreserve.gov
  2. 2.NIST AI RMF Core · National Institute of Standards and Technology