# ai-policy.yaml: the AI policy as data, the single source for two views. # ai-policy.md the prose policy people read and approve (one section per key below) # ai-policy.rego the executable skeleton a policy engine evaluates (rules[].rego) # Change the policy here, then regenerate or hand-edit both views in the same pull request, so the # prose and the code cannot drift. Illustrative, not a claim of conformity, and not legal advice. # Part of https://aigovernanceengineer.com/resources/templates # This work is licensed under CC BY 4.0. Attribution: Jorge García Aibar. policy: id: ai-policy version: "1.0.0" title: AI policy owner: ai-governance-committee approved_by: board-risk-committee effective_from: "2026-10-01" review_by: "2027-10-01" evidences: - ISO/IEC 42001 5.2 - ISO/IEC 42001 A.2 - NIST AI RMF GOVERN 1.2 - NIST AI RMF GOVERN 1.4 - EU AI Act Art. 17 scope: applies_to: - every AI system, model and agent the organisation builds, buys or runs - generative AI tools used by staff and contractors for work excludes: - AI used only in research sandboxes with synthetic or public data and no production access definitions: ai_system: A system that infers from its inputs how to generate outputs such as predictions, content, recommendations or decisions. agent: An AI system that can take actions through tools on its own initiative. record: A structured document that validates against a schema in the templates library. principles: - id: earliest-block statement: Build each control at the earliest point where it can block. - id: teeth statement: Give every control teeth; a control that cannot fail a build or stop an action is advice. - id: register-first statement: Register and bound every actor before it acts. - id: self-proving statement: Instrument the build to produce its own proof. - id: harm-first statement: Start every control from a named failure mode or harm. - id: governed-path statement: Make the governed path the easiest path. roles: - id: ai-governance-committee duty: Approves this policy, high and critical tier use cases, exceptions and residual-risk acceptance above the owner's authority. charter: committee-charter.md - id: system-owner duty: Accountable for one AI system through its lifecycle and for every record filed against it. - id: ai-governance-engineer duty: Implements this policy as code, runs the gates and maintains the evidence store. - id: privacy-office duty: Owns DPIA addenda and the lawful-basis checks in dataset cards. - id: security duty: Owns agent identity, adversarial testing and AI security incidents. - id: model-risk duty: Challenges models independently and signs test reports for high and critical tiers. - id: internal-audit duty: Tests that the controls run, from the evidence store. rules: - id: register-before-production statement: No AI system or agent reaches production without a current register entry naming an owner, a declared scope and an expiry. applies_to: [ai_system, agent] enforcement: [deploy] rego: deny records: [ai-system-register-entry, agent-register-entry] evidences: [EU AI Act Art. 49, ISO/IEC 42001 A.4, NIST AI RMF GOVERN 1.6] - id: classify-at-intake statement: Every new use of AI is recorded at intake, classified, and routed to the assessments its class requires before build work starts. applies_to: [ai_system, agent] enforcement: [deploy] rego: deny records: [use-case-record] evidences: [EU AI Act Art. 6, ISO/IEC 42001 A.5, NIST AI RMF MAP 1.1] - id: impact-assessment-before-use statement: Each assessment the intake routes a system to (AI impact assessment, DPIA addendum or FRIA) is approved before first use. applies_to: [ai_system] enforcement: [deploy] rego: deny records: [impact-assessment] evidences: [EU AI Act Art. 27, GDPR Art. 35, ISO/IEC 42005] - id: data-admission statement: No dataset enters a training, evaluation or retrieval pipeline without a dataset card and an admit decision from the admission gate. applies_to: [dataset] enforcement: [pre_merge] rego: deny records: [dataset-card, dataset-admission-record] evidences: [EU AI Act Art. 10, ISO/IEC 42001 A.7] - id: eval-gate statement: A release ships only if every blocking suite in its test plan has a passing result for that exact version. applies_to: [ai_system, agent] enforcement: [pre_merge, deploy] rego: deny records: [test-plan, eval-result, test-report] evidences: [EU AI Act Art. 15, EU AI Act Art. 9(8), NIST AI RMF MEASURE 2.3] - id: go-no-go statement: High and critical tier releases need a signed go or conditional-go decision for that exact version. applies_to: [ai_system, agent] enforcement: [deploy] rego: deny records: [go-no-go] evidences: [EU AI Act Art. 17, NIST AI RMF MANAGE 1.1] - id: agents-bounded-spend statement: Agents act only through registered tools and operations, while active, and within their spend limit; anything above the limit needs a human approval. applies_to: [agent] enforcement: [runtime] rego: deny, require_approval records: [agent-register-entry, policy-card, evidence-record] evidences: [EU AI Act Art. 14, ISO/IEC 42001 A.9, NIST AI RMF MANAGE 2.4] - id: oversight-competence statement: Only people with a current training record for a system may approve, override or stop its outputs. applies_to: [ai_system, agent] enforcement: [runtime] rego: deny records: [training-record] evidences: [EU AI Act Art. 4, EU AI Act Art. 26(2), ISO/IEC 42001 7.2, NIST AI RMF GOVERN 2.2] - id: sanctioned-genai-tools statement: Staff use only approved generative AI tools, and never enter confidential or restricted data into a tool not approved for that class. applies_to: [staff_use] enforcement: [runtime] rego: deny records: [training-record, evidence-record] evidences: [ISO/IEC 42001 A.9, ISO/IEC 42001 7.3] - id: vendor-due-diligence statement: No third-party AI product is bought or renewed without an assessed due-diligence response that is still in date, with its conditions written into the contract. applies_to: [procurement] enforcement: [periodic] rego: deny records: [vendor-due-diligence-response] evidences: [EU AI Act Art. 25, GDPR Art. 28, ISO/IEC 42001 A.10, NIST AI RMF GOVERN 6.1] - id: incident-handling statement: Any harm, near miss or out-of-policy action involving an AI system is recorded as an incident within one working day and assessed against every reporting regime. applies_to: [ai_system, agent] enforcement: [periodic] rego: null # a process control: evidenced by incident records, not decided by the engine records: [incident-record, risk-register-entry] evidences: [EU AI Act Art. 73, NIST AI RMF MANAGE 4.3] - id: decommission-by-runbook statement: Systems are retired by runbook, with identity revoked, the register entry retired, evidence archived and data disposed of. applies_to: [ai_system, agent] enforcement: [periodic] rego: null # a process control: evidenced by the runbook's step records records: [decommissioning-runbook] evidences: [NIST AI RMF GOVERN 1.7, EU AI Act Art. 18] exceptions: requested_by: system-owner approved_by: ai-governance-committee max_duration_days: 90 record: An exception is a signed verdict override with an expiry, filed in the evidence store. review: cadence: annual triggers: - a change in law or in a standard the organisation relies on - a serious incident - an internal audit finding against this policy - a new class of AI use (for example the first agent with write access to money or personal data)