Skip to content

AI governance: committees, risk tiers and ISO 42001

How to set up AI governance: who sits on the committee, how to tier use cases by risk, and what ISO/IEC 42001, 23053 and 22989 each cover.

6 min readBy AI Guru AcademyFacts last checked

By AI Guru Academy

In short

AI governance is a small, named group with the authority to decide which AI uses are allowed, how closely each one is checked, and who answers when something goes wrong. A practical set-up has three parts: a cross-functional committee, risk tiers based on impact and reversibility, and a register of approved, blocked and under-evaluation tools. ISO/IEC 42001 is the international standard for running this as a management system.

Who should sit on an AI governance committee?

Include everyone whose decisions an AI use case can affect, not only technical staff. A committee made only of engineers misses legal, people and customer risk. A committee made only of lawyers slows everything and misses how the tools work.

A workable membership looks like this (an illustration, not a mandated structure):

Who should sit on an AI governance committee?
RoleWhat they bring
Executive sponsorAuthority to approve, fund and stop use cases
Business owner of the use caseThe outcome and who is affected
Legal and complianceRegulatory and contractual duties
Security and privacyData handling, access, incident response
Technical leadWhat the system can and cannot do
People or HR representativeTraining and staff impact

The AWS Cloud Adoption Framework groups its capabilities into six perspectives: Business, People, Governance, Platform, Security and Operations. That spread is a useful check on your membership. If one of those perspectives has nobody speaking for it, ask why.

How do you tier AI use cases by risk?

Tier each use case by two questions: how much harm a wrong output could cause, and how easily the result can be undone. Impact alone is not enough. A wrong draft slogan that a person reads before use is low impact and fully reversible. A wrong decision sent automatically to a customer may be hard to reverse.

The tiers below are an illustration of one way to apply the two questions. Neither AWS nor ISO prescribes them.

How do you tier AI use cases by risk?
TierTypical impactReversibilityExample (invented)Oversight
1 LowInternal convenienceEasy to undoSummarising public meeting notesTeam lead sign-off
2 MediumAffects customers or staff indirectlyMostly reversibleDrafting replies a person reviewsOwner approval, spot checks
3 HighAffects money, access or rightsHard to reverseAutomatically declining a service requestCommittee approval, human review of each decision, ongoing monitoring

Decision rule: start at the tier that matches impact, then move up one tier if the output goes out without a person checking it, or if mistakes cannot be corrected afterwards.

Worked example (illustration): a regional training company wants a chatbot to answer course questions. It uses only published course pages, so impact is low. But it may quote refund terms, and a wrong answer could create a dispute. The committee places it in tier 2, requires a human-approved list of answerable topics, and reviews a sample of conversations each month.

What controls support each tier?

Match controls to the tier, and use the platform's own safeguards where they exist. Amazon Bedrock Guardrails, for example, can apply content filters and denied topics to model inputs and outputs. The content filter categories are Hate, Insults, Sexual, Violence, Misconduct and Prompt Attack. Guardrails can also block or mask sensitive information, such as personal data, in user inputs and model responses. Two limits matter for governance: this filter does not evaluate tool-use (function calling) inputs or tool results, and masking does not change the original request recorded in model invocation logs.

A contextual grounding check scores a model response for grounding (is it supported by the reference source you supply?) and relevance (does it answer the user's query?). It needs both a reference source and a query. The grounding and relevance thresholds are configurable, so the tier can set how strict it is.

For evidence, Amazon Bedrock model invocation logging can record request data, response data and metadata to an Amazon S3 bucket or a CloudWatch Logs log group in the same account and Region. Logging is off by default, so a governance policy has to require turning it on where the tier demands it.

Under the AWS shared responsibility model, AWS is responsible for the security of the cloud, and the customer for security in the cloud, which AWS describes as including managing its data, classifying its assets and using IAM tools to apply appropriate permissions. What the customer must configure depends on the services it chooses. Governance does not transfer to the provider.

How does an approved, blocked and under-evaluation register stop shadow AI?

A tool register gives staff a clear answer to "can I use this?", which is the best defence against shadow AI, meaning AI tools used without the organisation's knowledge. People turn to unapproved tools when the approved route is unclear or slow.

Keep the register short and public inside the organisation:

  • Approved: the tool, the permitted data types, the tier it is cleared for, and the named owner.
  • Under evaluation: tools being assessed, with an expected decision date and a rule that no confidential data goes in meanwhile.
  • Blocked: tools refused, with the reason in one line, so staff do not keep asking.

Checklist for each entry: owner, purpose, data allowed, highest tier permitted, review date. A tool not on the register is treated as under evaluation by default, never as approved.

What do ISO/IEC 42001, 23053 and 22989 each cover?

They cover three different things, and the descriptions below come from ISO's catalogue.

What do ISO/IEC 42001, 23053 and 22989 each cover?
StandardWhat it is
ISO/IEC 42001AI management system
ISO/IEC 23053Framework for AI systems using machine learning
ISO/IEC 22989AI concepts and terminology

ISO/IEC 42001 is a standard organisations can be certified against. AWS notes that 42001 certification means an independent third party has validated how an organisation manages AI risk, and that customers are not certified by association with AWS. AWS describes the standard as an international standard that specifies requirements for establishing, implementing, maintaining and continually improving an AI management system. In plain terms, it asks you to run AI governance as an ongoing process, not a one-off policy document.

23053 helps teams describe the parts of a machine-learning-based system in a shared way. 22989 gives everyone the same vocabulary, which removes many arguments caused by people using the same word differently.

One caution: this article relies on AWS's description of 42001 and ISO's catalogue titles. It has not reviewed the full text of the standards, so check the standard itself before building a certification plan.

How does governance relate to AWS adoption frameworks?

Governance is one perspective of the AWS Cloud Adoption Framework, alongside Business, People, Platform, Security and Operations. Do not confuse these perspectives with stages. AWS also publishes a generative AI maturity model with four levels: Envision, Experiment, Launch and Scale. AWS lists establishing foundational and governance frameworks among the Experiment level's criteria, so a team at that level already needs a register and tiering. Shadow AI can start during experimentation.

Key takeaways

  • Governance needs a named, cross-functional group with authority to approve and to stop use cases.
  • Tier use cases by impact and reversibility, and raise the tier when no person checks the output.
  • Match controls to tier, and require logging where the tier demands evidence.
  • A short approved, blocked and under-evaluation register is the most practical shadow AI control.
  • ISO/IEC 42001 is the AI management system standard. 23053 and 22989 cover system framework and terminology.
  • Using a cloud provider's safeguards does not hand over your governance duties.

FAQ

What is ISO/IEC 42001 in simple terms?

It is an international standard for an AI management system. AWS describes it as specifying requirements for establishing, implementing, maintaining and continually improving that system.

Do I need a committee if my organisation is small?

You need the decisions the committee makes, even if three people make them. Name who approves use cases, who owns security and privacy, and who can stop a tool.

Is there an official list of risk tiers?

The tiers in this article are an illustration. Choose tiers your organisation can apply consistently, and document how each is decided.

Does using Amazon Bedrock Guardrails complete my governance?

No. Guardrails apply filters and checks to model inputs and outputs. Policy, ownership, approval and review remain with your organisation.

Sources

Last checked 2026-11-16

AI Guru Academy is an independent training provider and is not affiliated with or endorsed by Amazon Web Services. AWS service names are trademarks of Amazon.com, Inc. or its affiliates.

Next step: take the free AIB-C01 diagnostic at /diagnostic, then see the course page at /courses/aib-c01-exam-prep.

AI governance: committees, risk tiers and ISO 42001 | AI Guru Academy