AI governance policy and accountability register

AI governance policy and accountability register

AG-01 — AI governance policy

[Organisation] governs AI activities according to their intended use, authority, information processing and consequences. The policy applies to the development, acquisition, integration, use, change and retirement of AI capabilities within [scope]. The AI owner maintains the inventory, risk decisions and evidence. Existing SSDLC governance remains applicable to AI-related software, infrastructure, suppliers and operational work.

Governance considers technical reliability and security together with effects on people and organisational objectives. The organisation uses the NIST AI RMF functions Govern, Map, Measure and Manage to organise relevant work. These functions support the operating records; they do not replace the system life cycle or establish that an AI capability is trustworthy without evidence.

AG-02 — Inventory and intended use

The owner records the AI capability, provider or development source, model and service identifiers, system interfaces, information used and the people responsible for operation. The record distinguishes model development, system integration, service acquisition and assisted work. It identifies the decision or task supported, intended users, affected people, operating conditions and known limitations.

Permitted use and authority are stated explicitly. The record identifies outputs treated as advice, outputs used in decisions and actions the system may execute. Changes in audience, purpose, autonomy, information access or connected tools are assessed as changes to the operating context. A capability approved for one task is not assumed to be approved for every task it can technically perform.

AG-03 — Accountability and oversight

The role register identifies responsibility for use-case approval, data, technical integration, evaluation, risk decisions, operation and incident handling. Oversight arrangements state what a reviewer can observe, what decision they make, when intervention is possible and what authority they hold. A nominal human review step is assessed for whether the person has enough information, time and competence to perform it.

The organisation records the responsibilities retained when using an external model or service. Provider information, local evaluation and operating evidence are considered together. The owner identifies what cannot be inspected or controlled and the consequence for reliance. Supplier assurances are not represented as evidence of properties outside their scope.

AG-04 — Map context and risk

The assessor identifies the purpose, stakeholders, information flows, dependencies, likely uses and reasonably relevant misuse or failure conditions. Analysis considers inaccurate or misleading outputs, security compromise, privacy effects, unavailable service, inappropriate reliance and impacts on affected people. Consequences are described in the actual context rather than inferred only from the model category.

The assessment records whether performance differs across relevant populations, languages, conditions or uses and how those differences affect stakeholder needs. It considers the ability to contest, correct or recover from an adverse outcome. Technical and stakeholder concerns remain connected to requirements and risk decisions, with uncertainty and evidence limitations preserved.

AG-05 — Measure and decide

The evaluation owner establishes criteria and methods for the intended use before judging the capability. Evaluation includes relevant functional performance, security, privacy, robustness, operational reliability and stakeholder effects. Results identify the model or service version, configuration, test data, conditions and uncertainty. Aggregate measures are supplemented where they conceal consequential failures in a particular condition or group.

The decision authority reviews benefits, alternatives, evaluation results, residual risks and limitations. The decision states the permitted scope, conditions, monitoring and reassessment triggers. It may authorise a bounded use or trial where appropriate to the authority and risk decision. The record distinguishes approval to evaluate from approval to operate and preserves unresolved issues.

AG-06 — Operate and manage change

The operations owner monitors the criteria identified for the use case and receives reports of harmful, incorrect or unexpected behaviour. The plan identifies how the capability can be restricted, suspended, replaced or returned to a previous configuration where technically possible. Actions follow the relevant authority and change arrangements. A fallback is assessed for whether it is actually available and suitable.

Changes in model, provider behaviour, prompts, retrieval sources, training or evaluation data, tools and operating population are assessed for their effect on the approval basis. The owner records version visibility limitations for remotely managed services. Incidents and significant findings feed into requirements, supplier management, evaluation and management review.

AG-07 — Management-system review

At [schedule], management reviews the AI inventory, objectives, resources, responsibilities, risk position, evaluation coverage, incidents and corrective actions. The review considers whether policies and competence remain appropriate as use changes. Decisions identify owners, resources, deadlines and effectiveness evidence. The organisation retains the relationship between management promises and the operational assessments that test them.

Where the organisation considers ISO/IEC 42001, it records the scope of that consideration and the sources used. This policy provides original operational content and does not assert full clause-by-clause conformity or certification. A crosswalk involving a draft ISO edition is identified as such and is not silently treated as a verified mapping to the final standard.

AI governance record: [AG ID; capability; owner; organisational role]
Intended use: [Task; users; affected people; conditions; limits]
Authority: [Advice/decision/action; permissions; oversight; intervention]
Information and supply chain: [Data; models; services; dependencies]
Risk and evaluation: [Context; criteria; results; uncertainty]
Decision: [Authority; scope; conditions; review; monitoring]
Management review: [Findings; resources; actions; effectiveness]