Practical guide

How to operationalize model risk in AI workflows

Operationalizing model risk means connecting a defined use to a classified system, accountable owners, controlled evidence, independent challenge, approved limits, monitoring, and an action when conditions change. A policy can name these requirements; a workflow must make them happen before release and during use. This guide is not a universal regulatory procedure. Federal Reserve SR 26-2.

By Two Prune

Operationalizing model risk means connecting a defined use to a classified system, accountable owners, controlled evidence, independent challenge, approved limits, monitoring, and an action when conditions change. A policy can name these requirements; a workflow must make them happen before release and during use. This guide is not a universal regulatory procedure. Federal Reserve SR 26-2 applies to specified banking organizations and expressly excludes generative and agentic AI from that guidance. The NIST AI Risk Management Framework is voluntary and broader. Organizations should determine their applicable legal, regulatory, safety, and sector duties with qualified reviewers. Federal Reserve: SR 26-2 NIST: AI RMF Core

Define the control objective

The workflow objective is: no material model or adjacent AI use reaches or remains in operation without a documented purpose, scope decision, accountable owner, proportionate challenge, approved conditions, monitoring, and retained decision evidence.

Define “model,” “AI system,” “material use,” “change,” “independent challenge,” “exception,” and “retirement” for the organization. Route systems outside the model definition into another named governance path rather than leaving them unowned.

Run the ROUTE workflow

### R — Register the use and route its classification

The requester records the business purpose, decision consequence, users, affected groups, data, method, vendor components, integrations, output use, fallback, and proposed owner. The control owner checks the request against procurement, deployment, access, and change records.

The classification record assigns model status or an adjacent route, materiality, risk tier, required reviewers, and review date. Unresolved classification is a hold condition, not permission to proceed.

### O — Own the decision and map dependencies

Assign a business owner for purpose and consequence, technical owner for implementation, data owner for source authority, control owner for workflow evidence, independent challenger, release approver, and operational users. Name who can accept residual risk and who cannot.

Map upstream data, prompts or features, external services, human steps, overrides, downstream consumers, monitoring feeds, and fallback procedures. A vendor contract or system card can inform review but does not replace organizational ownership.

### U — Undertake development and independent challenge

The developer or supplier provides intended use, theory or design rationale, training or configuration evidence where relevant, data suitability, limitations, performance tests, security and privacy evidence, and change history.

Independent challenge should test conceptual soundness, data lineage, outcome analysis, benchmarks, sensitivity, boundary cases, implementation, user controls, and whether the evidence matches the proposed use. Record disagreements and their disposition. “Independent” requires sufficient authority, access, competence, and freedom from being the sole author of the work under review.

### T — Tie approval to limits, release, and monitoring

The approver selects approve, approve with limits, repair, hold, or reject. Approval identifies the exact version, permitted users and uses, prohibited uses, required human review, access controls, monitoring, issue thresholds, expiry or review date, and residual uncertainty.

Before release, verify that production configuration matches the reviewed version, user communication states limitations, monitoring receives the right population, fallback works, and evidence is retained. A technically successful deployment is not evidence of approval.

Monitoring must name source data, population, metric, threshold, frequency, owner, response, and escalation. Include data and performance deterioration, use outside scope, override patterns, unexpected outcomes, complaints, vendor changes, and control failures where relevant.

### E — Escalate changes, exceptions, issues, and exit

Changes enter triage before implementation. Define triggers such as new data, method, prompt architecture, vendor version, integration, user group, output consequence, jurisdiction, or material performance shift. The control owner decides whether to update evidence, repeat challenge, reapprove, restrict, or stop.

An exception records exact scope, reason, authority, compensating controls, evidence, expiry, and stop condition. An issue records severity, impact, containment, owner, due date, remediation, and closure test. Retirement removes access, manages replacements and downstream dependencies, and archives evidence.

Workflow evidence ledger

| Gate | Required artifact | Accountable role | Test | Allowed disposition | | --- | --- | --- | --- | --- | | Intake | use and consequence record | business owner | purpose and fallback are explicit | route / repair / hold | | Classification | status, tier, route, date | control owner | decision matches definitions | pass / escalate | | Build or acquire | design and data evidence | technical and data owners | evidence matches intended use | challenge / repair | | Challenge | tests, findings, dissent | independent reviewer | material claims and boundaries tested | pass / limit / hold | | Approval | version, use, limits, residual risk | authorized approver | conditions are implementable | approve / reject | | Release | configuration and control check | release owner | production matches approval | release / hold | | Monitor | measures, thresholds, response | operational owner | breaches are detected and routed | continue / restrict / stop | | Change or exit | reassessment or retirement record | control owner | dependencies and access resolved | reapprove / retire |

Test the workflow with samples from both approvals and stops. A completed form does not prove that challenge was competent, monitoring covers the relevant population, or limits were enforced.

Worked workflow: vendor forecast version change

A synthetic planning team uses a vendor demand forecast for established products. The approved version uses weekly sales and promotion data, requires planner review, and excludes products with fewer than eight weeks of history.

The vendor announces a new version that adds external trend data and changes the update frequency. Procurement calls it a routine upgrade. Under ROUTE, the change trigger fires because data, method, version, and monitoring timing changed. The technical owner opens a change record rather than replacing the production endpoint.

The dependency map finds that the current monitoring job runs weekly after the old forecast. With the new daily update, it would compare mismatched periods and could miss reversals. The data owner also cannot yet verify the external trend source’s coverage and refresh history.

Independent challenge runs a fixed historical sample, sparse-data cases, promotion spikes, missing-feed tests, and product-family segmentation. Aggregate error appears similar, but launch-family error worsens and the external feed fails without a visible warning. The challenger records a material disagreement with broad approval.

The approver selects limited test, not production approval. Conditions restrict the version to a shadow run, require a feed-failure alert, preserve the old fallback, prohibit launch-family use, and set a four-week review. Monitoring names population, error measures, feed availability, thresholds, owners, and the response to breach.

The evidence packet contains intake, classification, dependency map, vendor evidence, data-source gap, challenge results, dissent, limited-test approval, release configuration, monitoring definition, and review date. If the external source remains unverifiable or the alert fails, the workflow stops the change and retains the approved version.

This example shows how a workflow routes a version change. It does not demonstrate forecast accuracy, regulatory compliance, vendor safety, or control effectiveness elsewhere.

Failure and hold conditions

Hold release or continued use when purpose or consequence is unclear, classification is unresolved, ownership is missing, reviewed and deployed versions differ, data authority is unknown, challenge lacks access or competence, findings have no disposition, monitoring cannot be reproduced, an exception has no expiry, or closure lacks retest evidence.

Review the workflow on a fixed cadence and after material regulation, technology, incident, organizational, or use changes. Keep policy design, sampled operation, and supported effect as separate claims.

Two Prune can place model-risk workflow design inside a realistic AI-enabled assessment and evaluate how a participant classifies a use, navigates evidence, challenges assumptions, manages changes, and communicates an operational control. 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]