Start with intended use, not a model list
An AI inventory is useful when it explains how a capability is used in a real process. Include systems purchased by a business unit, features embedded in existing software, internally built models, and generative AI services used by staff. Record the intended purpose, business owner, affected people, data involved, supplier, connected systems, decision points, human review, fallback, and the service or control that could be affected. For a mission-critical operation, identify who owns service continuity and how work proceeds if the capability is unavailable or suspended.[2][3]
Describe the use case in operational language. “Summarize incoming case files for a licensed reviewer” is more useful than “uses a language model.” The first description makes the reviewer, workflow, source material, and limits visible. It also gives the organization something to reassess when the model, task, or population changes.
- Ask process owners to identify where AI generates, ranks, recommends, classifies, or changes information used in a decision.
- Record the business purpose and the consequences of an incorrect, delayed, unavailable, or misused output.
- Name an accountable owner who can confirm that the recorded use still matches actual practice.
Set decision rights by risk and consequence
A control framework needs to say who can approve a use, what evidence they need, and who can pause it. Set review depth according to the possible effect on people, service continuity, safety, financial or operational decisions, sensitive information, and the organization’s ability to recover. A low-impact internal drafting aid and a system influencing a consequential eligibility or safety decision should not move through identical approvals.
Keep risk acceptance with a role that has authority over the affected service and risk. Technology teams can assess architecture and performance, but they should not silently accept operational, legal, privacy, security, workforce, or safety risks for another accountable owner. Establish escalation for unresolved disagreement and define which exceptions require executive or board-level attention.[2]
Legal requirements are not interchangeable with a voluntary framework. The EU AI Act creates binding obligations for actors and systems within its scope, with responsibilities that vary by role and system category. For example, Article 26 includes human-oversight and monitoring duties for deployers of high-risk AI systems. Organizations need a fact-specific assessment of territorial scope, role, intended purpose, and classification before deciding which duties apply.[4]
- Define approval, conditional approval, exception, suspension, and retirement authority for each risk tier.
- Specify when privacy, security, legal, compliance, safety, procurement, workforce, and affected operations must review a proposal.
- Document the rationale, conditions, accountable owner, residual risks, and next review date for each decision.
U.S. banking supervision context
For U.S. banking organizations that use models, the Federal Reserve, OCC, and FDIC issued revised model risk management guidance on April 17, 2026, as SR 26-2. It supersedes SR 11-7 and SR 21-8 and describes risk-based practices tailored to a banking organization’s model risk profile, size, and complexity. The letter says it is expected to be most relevant to banks over $30 billion regulated by the Federal Reserve. This is banking supervisory guidance, not a universal AI governance obligation; each organization must assess its own scope and applicable requirements.[5]
Put controls inside the work
Governance becomes usable when it is part of existing work rather than a separate document people must remember. Add a short AI intake to project initiation, procurement, process redesign, and software change workflows. The review should identify purpose and alternatives, data restrictions, supplier commitments, required testing, human review, records to retain, monitoring, and a safe way to stop or fall back.
Controls should match the use. A staff research assistant might require approved data boundaries, user guidance, and spot checks. A capability that informs a high-impact decision may need documented validation, qualified human review, a defined override, traceable records, contingency procedures, and explicit approval before release. Do not assume that a human in the loop is meaningful unless the reviewer has enough time, context, competence, and authority to challenge the output.[2][4]
For generative AI, test how prompts, retrieved material, model updates, and output uncertainty affect the intended task. Keep the final decision with the accountable person where the process requires judgment. Preserve an alternative path for time-critical work if a service is unavailable or produces unreliable results.[3]
- Add governance questions to existing intake and change forms instead of creating an unconnected approval queue.
- Make control owners and required evidence visible in project plans, supplier reviews, and operating procedures.
- Check that frontline staff can follow the control under actual workload and service conditions.
Monitor, escalate, and learn
Approval is a decision for a particular purpose and set of conditions. It does not establish that a system will remain suitable indefinitely. Define monitoring around the service: output quality, error and override patterns, workload, service interruption, complaints, near misses, data or model changes, and relevant outcomes. Choose measures that can lead to action, assign a person to review them, and set thresholds for investigation or suspension.[2]
An escalation path should say how staff report a concern, who triages it, who can pause the use, how affected work continues, and when the supplier, risk function, executive, or regulator must be notified under applicable rules. Capture incidents and near misses in existing operational risk and incident processes so the organization can connect a model issue to its service impact.
Revisit decisions when the context changes
Reassessment is warranted when the purpose, data, population, workflow, supplier, model version, integration, level of automation, or operating environment changes. It is also warranted when monitoring finds drift, repeated overrides, an unexpected impact, a new dependency, or a new legal interpretation. Set both event-based triggers and a periodic review cadence; the right interval depends on the consequence and the speed at which the system or operating context can change.[2][4]
This is where AI governance connects to broader transformation work. Operations define where a capability belongs, technology teams make it supportable, leaders set ownership and priorities, and disciplined execution turns approvals into controls people can use.
- Which live AI uses support a critical process, and can each owner explain the intended purpose and fallback?
- Who has the authority to approve residual risk, require a change, or stop a use?
- What evidence would cause the organization to change its decision, and who reviews that evidence?
- When did a material system or workflow change last trigger a fresh assessment?
Sources
- NIST AI Risk Management Framework
NIST describes the AI RMF as voluntary guidance. Its Govern, Map, Measure, and Manage functions address risk across the AI system lifecycle.
- NIST AI RMF 1.0, Core
The framework describes governance as cross-cutting and calls for lifecycle risk management, multidisciplinary perspectives, monitoring, review, and clear roles.
- NIST AI RMF Generative AI Profile
NIST published this voluntary cross-sector profile to help organizations consider risks specific to generative AI.
- Regulation (EU) 2024/1689, Artificial Intelligence Act
The consolidated EU regulation sets obligations for covered providers, deployers, and other operators. Duties depend on scope, role, and system category.
- Federal Reserve/OCC/FDIC Revised Guidance on Model Risk Management (SR 26-2)
Issued April 17, 2026, the revised interagency guidance supersedes SR 11-7 and SR 21-8. It describes a risk-based approach tailored to a banking organization’s model risk profile, size, and complexity; the letter expects it to be most relevant to Federal Reserve-regulated banks with over $30 billion in assets.