AI governance frameworks: Compare

A private certification standard, not law: voluntary to adopt, and the Artificial Intelligence Underwriting Company, which maintains it with a consortium, is the only body that issues the certificate.

Agentic AI systems only, built or deployed by any organisation in any country; the organisation chooses which agents are certified, and may start with one high-risk agent.

Separates the agent developer, which sets and documents secure defaults, from the deployer, which configures the agent securely for its own context; one organisation can be both.

Each agent gets a risk taxonomy built from its capabilities and deployment context (C001); requirements apply by capability tags such as universal, automation or code generation. No legal risk tiers.

Logs of agent processes, actions and outputs (E015) and a record of data processing locations (E011) are mandatory; a repository of model cards and datasheets is supplemental (E017).

A risk taxonomy (C001) and internal pre-deployment testing across risk categories (C002); the standard's own ISO/IEC 42001 crosswalk grades the impact assessment clauses as a partial gap.

Human review of flagged high-risk outputs (C007) and real-time feedback and intervention (C009) are supplemental, not mandatory; E004 names an accountable lead for each change that needs approval.

Users must be told when they deal with an AI system rather than a person (E016, mandatory); a transparency policy with model cards and datasheets is supplemental (E017).

A whole domain of eight mandatory requirements: input and output data policies (including training on customer data), limits on agent data access, and safeguards against leaks of personal data, IP, customer data and secrets.

Mandatory technical evals, unique on this wheel: third-party adversarial testing (B001) and third-party tests of harmful and out-of-scope outputs, hallucinations and tool calls at least every three months (C010 to C012, D002, D004).

Failure plans for security breaches, harmful outputs and hallucinations (E001 to E003) and logging of third-party access (E009); risk monitoring is supplemental (C008). No duty to report to a regulator.

An accredited auditor audits the controls and the publisher runs the technical tests; a certification committee issues a certificate valid for 12 months, kept valid by technical testing at least every three months.

Heavy: evidence across legal, operational and technical controls plus evals. The standard's pages give four to eight weeks (certification page) or five to ten weeks (FAQ), then quarterly testing and a yearly re-audit.

Released each quarter on 15 January, April, July and October; the current release is dated 2026-07-15, the site was last updated 2026-09-17, and the next release is due 2026-10-15. No legal dates.