Organisations often use "AI governance" and "AI risk management" as if they were the same programme, and the result is a committee that approves AI use cases without anyone owning the treatment of the risks it accepted. The two are linked but distinct. AI governance holds decision rights: who may approve, stop or change an AI system, against which policy and risk appetite. AI risk management is the process that identifies, analyses, evaluates and treats AI risks and gives named owners the job of carrying each treatment through. This guide separates the two by decision rights versus risk treatment ownership, maps them to ISO/IEC 42001, ISO/IEC 23894, ISO/IEC 38507 and the NIST AI RMF, shows the artefacts each produces and how they feed each other, and gives a RACI, decision criteria and a worked scenario. It is for AI governance leads, risk managers, CISOs and internal audit. The boundary between governance and security teams is covered separately in AI Governance vs AI Security.
Key takeaways
- Governance decides, risk management treats. Governance sets policy, risk appetite and who has authority to approve or stop an AI system; risk management runs the assessment and treatment process and tracks each treatment to closure.
- The standards split the same way. ISO/IEC 38507 gives guidance to the governing body; ISO/IEC 42001 sets certifiable management system requirements, including AI risk assessment and treatment in clauses 6.1.2, 6.1.3, 8.2 and 8.3; ISO/IEC 23894 is guidance on the risk process itself.
- In the NIST AI RMF, GOVERN is the governance layer. NIST describes GOVERN as cross-cutting and enabling the other functions, while MAP, MEASURE and MANAGE carry the risk work, including MANAGE 1.3 responses to mitigate, transfer, avoid or accept.
- Risk acceptance is a governance decision. The risk process can recommend acceptance of residual risk, but only someone with delegated authority under the governance model should sign it.
- They fail differently. Governance without risk management produces approvals with no evidence; risk management without governance produces registers that nobody has authority to act on.
The short answer
AI governance is the system of direction and oversight for AI: the policy, the risk appetite and tolerances, the bodies and roles with authority, the decision gates and the escalation routes. ISO/IEC 38507:2022 frames this from the governing body's point of view: it gives the members of the governing body guidance on enabling and overseeing the use of AI so that it is effective, efficient and acceptable. In the NIST AI RMF, GOVERN 2.3 asks that executive leadership takes responsibility for decisions about risks associated with AI systems.
AI risk management is the repeatable process that turns that direction into evidence and action for each AI system: context, identification, analysis, evaluation, treatment, monitoring and recording. ISO/IEC 23894:2023 provides guidance on managing AI-specific risk and integrating risk management into AI activities, building on ISO 31000. In the NIST AI RMF this is the work of MAP, MEASURE and MANAGE.
The cleanest test: if a question starts "may we" or "who decides", it is governance. If it starts "how likely, how bad, and what are we doing about it", it is risk management. See the AI Governance Guide and the AI Risk Assessment Guide for each discipline in depth.
Side-by-side comparison
| Dimension | AI governance | AI risk management |
|---|---|---|
| Core output | Decisions: approve, approve with conditions, reject, suspend, retire | Evidence and actions: risk ratings, treatment plans, residual risk statements |
| Main question | Should we, on what terms, and who answers for it? | What could go wrong, how likely and severe, and what are we doing about it? |
| Typical owner | Board or committee, executive sponsor, AI governance lead | Risk function (second line), AI system owners who own treatments |
| Cadence | Gate based (intake, go-live, material change) plus periodic oversight | Continuous and iterative across the life cycle |
| Key artefacts | AI policy, risk appetite, terms of reference, delegation of authority, decision log | Risk register, assessments, treatment plans, monitoring results |
| ISO family | ISO/IEC 38507 (governing body); ISO/IEC 42001 clause 5 Leadership | ISO/IEC 23894 (process); ISO/IEC 42001 clauses 6.1 and 8.2 to 8.4 |
| NIST AI RMF | GOVERN (6 categories) | MAP (5), MEASURE (4), MANAGE (4) |
| EU AI Act anchor | Article 17 quality management system, accountability framework | Article 9 risk management system for high-risk AI providers |
| Fails as | Rubber stamp or bottleneck | Paper register with no authority behind it |
How the frameworks divide the work
ISO/IEC 42001: one management system holding both
ISO/IEC 42001:2023 (edition 1.0, published 18 December 2023) specifies requirements for establishing, implementing, maintaining and continually improving an AI management system, and is the certifiable standard. Its structure places governance and risk management side by side. Clause 5 (Leadership) covers 5.1 Leadership and commitment, 5.2 AI policy and 5.3 Roles, responsibilities and authorities: this is the governance layer. Clause 6.1 plans the risk work through 6.1.2 AI risk assessment, 6.1.3 AI risk treatment and 6.1.4 AI system impact assessment, and clause 8 operates it through 8.2, 8.3 and 8.4. Clause 9.3 Management review closes the loop back to leadership. Annex A holds normative reference controls and Annex C lists potential AI-related objectives and risk sources. See the ISO/IEC 42001 Guide.
ISO/IEC 23894: the risk method
ISO/IEC 23894:2023 (published February 2023) is guidance, not a certifiable standard. It adapts ISO 31000 principles, framework and process to AI. An organisation certifying to ISO/IEC 42001 can use it as the method behind clauses 6.1.2 and 6.1.3. Note that it also contains a framework clause on leadership and integration: even the risk standard expects governance to exist. See the ISO/IEC 23894 Guide and the ISO 31000 Risk Management Guide.
ISO/IEC 38507: the governing body
ISO/IEC 38507:2022 sits above both. Its primary audience is the governing body, with executive managers, assessors and auditors also in scope. Use it to define what the board retains (appetite, accountability, oversight) and what it delegates.
NIST AI RMF: GOVERN versus MAP, MEASURE, MANAGE
The NIST AI RMF 1.0 was released on 26 January 2023 for voluntary use, and NIST states it is being revised as part of the White House AI Action Plan (checked November 2026). GOVERN sets the conditions: GOVERN 1.3 scales risk management activity to risk tolerance, GOVERN 1.4 establishes the risk process through transparent policies, GOVERN 1.5 plans monitoring and periodic review of that process, and GOVERN 2.1 documents roles, responsibilities and lines of communication for mapping, measuring and managing AI risks. The risk functions do the work: MAP 1.5 records risk tolerances, MEASURE 1.1 selects metrics starting with the most significant risks, MANAGE 1.2 prioritises treatment by impact, likelihood and resources, MANAGE 1.3 plans responses (mitigating, transferring, avoiding or accepting), MANAGE 1.4 documents negative residual risks, and MANAGE 4.1 runs post-deployment monitoring including override, decommissioning and incident response. NIST publishes crosswalks to ISO/IEC 23894 (revised, from INCITS/AI, August 2025) and to ISO/IEC 42001, and states that inclusion does not imply full coverage. See the NIST AI RMF Guide and NIST AI RMF vs ISO/IEC 42001.
Decision rights versus risk treatment ownership
The single most useful design step is to write down two separate lists: who holds which decision right, and who owns which risk treatment. They are different people more often than teams expect.
| Activity | Board or governing body | AI governance committee | Risk function (second line) | AI system owner | Internal audit |
|---|---|---|---|---|---|
| Set AI risk appetite and tolerances | A | R | C | I | I |
| Approve AI policy and delegation of authority | A | R | C | I | I |
| Define the risk method, scales and register | I | C | A/R | C | I |
| Run the risk assessment for a system | I | C (challenge) | A/R | ||
| Approve a use case at a gate | A for highest tier | A/R for other tiers | C | R (presents) | |
| Implement and evidence treatments | I | C (monitors) | A/R | ||
| Accept residual risk within appetite | A | C | R (requests) | I | |
| Accept residual risk above appetite | A | R | C | R (requests) | I |
| Suspend or retire a system | I | A | C | R | I |
| Independent assurance of both | I | I | I | I | A/R |
R = responsible, A = accountable, C = consulted, I = informed. Two rules keep it honest: whoever owns a treatment cannot also be the person who accepts the residual risk it leaves, and the risk function challenges the system owner's assessment rather than writing it. That mirrors the three lines approach covered in the Three Lines Model for Cyber Risk Guide. A ready-to-edit version is in the AI Governance RACI Template.
Artefacts and how they feed each other
| Artefact | Owned by | Feeds |
|---|---|---|
| AI policy and risk appetite statement | Governance | Risk criteria and tolerance thresholds used in every assessment |
| Delegation of authority matrix | Governance | Who can sign risk acceptance at each tier |
| AI system inventory | Governance (accountable), system owners (maintain) | Scope of the risk process; nothing unregistered gets assessed |
| Risk assessment and impact assessment | Risk management | Gate decision packs |
| Risk register and treatment plans | Risk management | Management review and committee reporting |
| Decision log (approvals, conditions, acceptances) | Governance | Treatment deadlines and conditions tracked by risk management |
| Monitoring and incident data | System owners, risk management | Reassessment triggers and governance decisions to suspend or retire |
The loop runs: governance sets appetite, risk management assesses against it, governance decides at a gate, risk management tracks the conditions and monitors, and the results return to governance through management review (ISO/IEC 42001 clause 9.3) or periodic review of the risk process (NIST GOVERN 1.5). The difference between a risk assessment and an impact assessment is covered in AI Risk Assessment vs AI Impact Assessment.
Threat model: what goes wrong in each function
When governance is weak
- Unclear authority: nobody can say who may approve an agent with write access to finance systems, so it is approved by whoever is asked first.
- No appetite: risk ratings cannot be judged acceptable or not, so every decision becomes a negotiation.
- Bottleneck: one committee reviews everything, teams route around it, and shadow AI grows outside the inventory.
- No stop authority: incidents occur, but nobody has a documented right to suspend the system.
When risk management is weak
- Assessments as forms: questionnaires completed once at intake, never refreshed after a model change or a new tool connection.
- Treatments without owners: conditions written into an approval but not tracked, so they quietly lapse.
- Unmeasurable risks dropped: MEASURE 1.1 asks that risks which cannot be measured are documented; weak programmes simply omit them.
- Residual risk never stated: approvals happen without a written residual risk, so acceptance is implicit and unaccountable.
Decision criteria: how to structure the two
| Situation | Structure | Reason |
|---|---|---|
| Small organisation, few AI uses, mostly bought SaaS | One owner for both, with a separate person signing high risk acceptances | Separate bodies would be theatre; the separation that matters is acceptance |
| Mid-size, several internal AI systems | AI governance committee plus AI risk folded into the enterprise risk function | Reuse existing ISO 31000 based scales and register |
| Regulated or high-risk AI provider under the EU AI Act | Formal governance under a quality management system and a dedicated Article 9 risk management system per high-risk system | Article 9 requires a documented, iterative process per system across its life cycle |
| Pursuing ISO/IEC 42001 certification | Map clause 5 to governance bodies and clauses 6.1 and 8 to the risk process; use ISO/IEC 23894 as the method | Auditors look for both and for the management review link |
| Large group with federated business units | Central appetite and method, local risk ownership and delegated approvals by tier | See Centralised vs Federated AI Governance |
| Financial services with an existing model risk function | Extend model risk management for AI risk treatment; governance stays with the AI committee and board | See the AI Model Risk Management Guide |
Trade-offs: merging the two in one person or team is fast but removes challenge; full separation adds challenge but slows decisions. A middle path that works for most organisations is a single programme with two explicit roles: the risk function owns method and challenge, the committee owns decisions, and acceptance authority rises with the risk tier.
EU AI Act: where the law draws the line
For providers of high-risk AI systems, Article 9 of the AI Act requires a risk management system that is established, implemented, documented and maintained as a continuous iterative process across the life cycle: identify and analyse known and reasonably foreseeable risks, estimate and evaluate them under intended use and reasonably foreseeable misuse, evaluate other risks from post-market monitoring, and adopt targeted measures. Residual risk must be judged acceptable, and measures follow an order: eliminate or reduce through design, then mitigation and controls, then information and training for deployers. Article 9(10) allows this to be part of, or combined with, internal risk processes required by other Union law. The governance side sits in Article 17, whose quality management system includes an accountability framework setting out the responsibilities of management and other staff. Text checked against the AI Act Service Desk consolidated version as at 27 July 2026. This is a summary, not legal advice; see the EU AI Act vs ISO/IEC 42001 guide for the mapping.
Worked scenario: an agent that drafts customer refunds
A retailer wants an AI agent that reads support tickets and drafts refunds up to a set value, which a human approves. Under governance, the AI policy classifies customer financial decisions as high tier, the delegation matrix says the AI governance committee approves high tier use cases and only the chief risk officer may accept residual risk above appetite. Under risk management, the system owner runs the assessment with the risk function challenging it: risks include wrong refunds, prompt injection through ticket text, and the agent's service credential reaching the payments API beyond refund scope. Treatments are a refund value cap enforced in the payments API rather than in the prompt, a scoped credential with only the refund endpoint, human approval above a lower threshold, and monitoring of refund rates per agent identity. The committee approves with two conditions and a 90 day review. Residual risk (some incorrect refunds below the cap) is within appetite, so the committee chair signs acceptance. At day 60, monitoring shows refund volume drifting upwards after a model update; the risk function raises it, the system owner reassesses, and the committee uses its documented suspend right until the cause is fixed. Each function did its own job, and neither could have done the other's. The agent's credential is a non-human identity: see the Identity Governance for AI Agents Guide.
Implementing the split
- Write the appetite first: a short AI risk appetite with tolerance thresholds per tier, approved by the governing body. See the Risk Appetite and Tolerance for Cyber Guide.
- Publish the delegation matrix: who approves, who accepts residual risk, who may suspend, at each tier.
- Adopt one risk method: extend the enterprise ISO 31000 process using ISO/IEC 23894 rather than building a separate AI register.
- Link inventory to register: every inventoried AI system has an assessment and named treatment owners.
- Pilot on three systems: one bought, one built, one agentic; check that decisions and treatments are recorded separately.
- Close the loop: quarterly management review of open treatments, acceptances near expiry and incidents.
- Adjust: if the committee becomes a bottleneck, delegate lower tiers rather than lowering the bar.
Detection and measurement
- Treatment closure rate: share of treatments delivered by their due date.
- Acceptance hygiene: share of residual risk acceptances signed by someone with the right delegated authority, with an expiry date.
- Coverage: share of inventoried AI systems with a current assessment.
- Reassessment latency: time from a material change (model update, new tool, new data) to a refreshed assessment.
- Decision cycle time: time from intake to gate decision, by tier, to spot bottlenecks.
Programme level indicators are covered in AI Governance Metrics.
Framework mapping
| Framework | Governance (decision rights) | Risk management (treatment) |
|---|---|---|
| ISO/IEC 42001:2023 | 5.1, 5.2, 5.3; 9.3 Management review | 6.1.2, 6.1.3, 6.1.4; 8.2, 8.3, 8.4 |
| ISO/IEC 23894:2023 and ISO/IEC 38507:2022 | ISO/IEC 38507 (governing body guidance) | ISO/IEC 23894 risk process guidance |
| NIST AI RMF 1.0 | GOVERN 1.3, 1.4, 1.5, 2.1, 2.3 | MAP 1.5; MEASURE 1.1; MANAGE 1.2, 1.3, 1.4, 4.1 |
| NIST CSF 2.0 | GV.RR-01, GV.RR-02, GV.RM-02, GV.OV-01 | ID.RA-06 |
| NIST SP 800-53 Rev 5 | PM-9 Risk Management Strategy; PM-28 Risk Framing | RA-3 Risk Assessment; RA-7 Risk Response |
| EU AI Act | Article 17 accountability framework | Article 9 risk management system |
Common failure patterns
- Committee as risk owner: the governance committee is listed as owner of every risk, so no individual delivers any treatment.
- Self-acceptance: the system owner who benefits from launch also accepts the residual risk.
- Two registers: an AI risk register separate from enterprise risk, with different scales, so the board cannot compare AI risk with anything else.
- Approval without expiry: acceptances and approvals with no review date, surviving model swaps and scope changes.
- Certification mistaken for control: ISO/IEC 42001 certification treated as proof that each system's risks are treated, when it attests to the management system.
- Agent credentials outside both: the identities AI agents use are approved by neither function; see the AI Governance vs AI Security guide for closing that handoff.
Practitioner checklist
- Document AI risk appetite and tolerances per tier, approved by the governing body.
- Publish a delegation matrix for approval, residual risk acceptance and suspension rights.
- Name an individual owner for every risk treatment, never a committee.
- Separate the person who accepts residual risk from the person who owns its treatment.
- Run AI risk through the enterprise risk method, extended with ISO/IEC 23894.
- Give every acceptance and approval an expiry date and a reassessment trigger list.
- Map ISO/IEC 42001 clause 5 and clauses 6.1 and 8 to named bodies and processes before audit.
- Report treatment closure, acceptance hygiene and coverage at each management review.
Standards and references
- ISO/IEC: ISO/IEC 42001:2023 Artificial intelligence, Management system
- ISO: ISO/IEC 23894:2023 Artificial intelligence, Guidance on risk management
- ISO: ISO/IEC 38507:2022 Governance implications of the use of artificial intelligence by organizations
- NIST: AI Risk Management Framework
- NIST AIRC: AI RMF Core (GOVERN, MAP, MEASURE, MANAGE)
- NIST AIRC: AI RMF Crosswalks
- European Commission AI Act Service Desk: Article 9 Risk management system
- European Commission AI Act Service Desk: Article 17 Quality management system
Related NHI Mgmt Group resources: AI Governance Guide, AI Risk Assessment Guide, AI Governance vs AI Security, ISO/IEC 23894 Guide, AI Governance RACI Template.