What is AI governance?

AI governance is the set of roles, decisions, controls and records through which an organisation manages the AI systems it builds, buys and uses. It answers four questions for each system: what it is, who owns it, whether it is acceptable and under what conditions, and what evidence shows the conditions are being met in practice.

It is broader than a policy document and narrower than all of IT governance. A policy says what is allowed. Governance makes sure each AI system has been looked at against that policy, that someone accepted the remaining risk, and that the decision is applied where the system actually runs.

What an AI governance programme covers

Most programmes, whatever framework they cite, end up with the same eight components.

  1. Inventory. A register of AI systems in use, built from declarations and from what is observed in traffic, with vendor, models and purpose.
  2. Ownership. A named business owner for each system, accountable for its use and for keeping the record current.
  3. Risk classification. A consistent way of placing each system in a risk tier, with reasons that a reviewer can read.
  4. Reviews. Privacy, security and vendor reviews proportionate to the tier, each with findings and conditions.
  5. Approval. A recorded decision by someone with authority, with conditions and an end date.
  6. Controls in operation. Technical controls that enforce the decision: which models, which data, which teams, what budget.
  7. Monitoring. Watching use and changes over time, and reopening a decision when the facts behind it change.
  8. Evidence. Records that show what was decided, on what basis, and what actually happened, in a form an auditor can check.

The main frameworks

Three references come up in almost every enterprise programme. They do different jobs and are often used together.

FrameworkWhat it isBinding?
EU AI Act, Regulation (EU) 2024/1689A risk-based EU regulation that sets obligations for providers, who develop AI systems, and deployers, who use them under their authorityYes, within its scope
NIST AI Risk Management Framework 1.0A framework published by NIST in January 2023, organised in four functions: Govern, Map, Measure and ManageNo, intended for voluntary use
ISO/IEC 42001:2023An international standard that specifies requirements for establishing, implementing, maintaining and continually improving an AI management systemNo, adopted by choice

In the NIST AI RMF, Govern is cross-cutting: it sets the culture, processes and structures. Map establishes context and intended purpose. Measure analyses and monitors risks. Manage allocates resources and plans responses. NIST also published a Generative AI Profile, NIST AI 600-1, in July 2024.

On EU AI Act timing, prohibited practices have applied since 2 February 2025, general-purpose AI model obligations since 2 August 2025 and Article 50 transparency obligations since 2 August 2026. The Digital Omnibus on AI moved high-risk obligations for Annex III systems to 2 December 2027 and for Annex I to 2 August 2028. Whether and how the Act applies to a given system is a determination for your organisation and its counsel.

Who owns what

RoleTypically ownsTypically decides
CISOSecurity reviews, gateway and access controls, incident handlingWhether a system meets security conditions
DPOPrivacy reviews, DPIA where required, records of processingAdvises on whether personal data may be used, and how
RiskThe classification method, risk appetite, reporting to the boardWhich tier a system sits in and what each tier requires
LegalRegulatory scope, contracts with AI vendors, terms of useInterpretation of obligations for the organisation
Business ownerThe system's purpose, users, data and day-to-day useRequests approval and accepts the conditions attached

The requester should not approve their own request. Separating the person who asks from the person who approves is one of the simplest controls a programme can have.

Common failure modes

  • An inventory on paper, disconnected from traffic. The register lists what people declared a year ago. The gateway logs show something else. Nobody reconciles the two.
  • Approvals that never expire. A system approved for one use, with one model and one dataset, keeps its approval after all three have changed.
  • Risk scores nobody can explain. A system is rated 62 out of 100, and no one can say which answer moved it or what would change the rating.
  • Policy without enforcement. Conditions are written into the approval but nothing at the point of use applies them.
  • Evidence assembled before the audit. Screenshots and spreadsheets are collected after the fact, which shows the intent but not what happened.

A six-step way to start

  1. Name one accountable owner. Give one executive the authority to decide, and a small group that includes privacy, security, risk and legal.
  2. Build a first inventory. Combine a short declaration form with proxy or DNS logs matched against known AI providers.
  3. Agree a classification method. Use a few factors, such as data sensitivity, autonomy and who is affected, and record a reason for each.
  4. Set up intake and approval. One route to request a system, proportionate reviews, an approver who is not the requester, and an end date on every approval.
  5. Route traffic through one control point. Apply approved models, data rules and budgets where requests actually pass.
  6. Keep evidence as you go. Record each decision and each enforced request at the time, and map them to the frameworks you report against.

A short checklist

  • An accountable owner and a written AI policy
  • An inventory that includes observed as well as declared systems
  • A business owner for every system in it
  • A classification method with readable reasons
  • Approvals with conditions and an end date
  • Controls applied where traffic passes
  • Evidence recorded at the time of each decision

How MERIDVAR fits

MERIDVAR is self-hosted software that supports the programme from inventory to evidence, on your own network.

  • Inventory. AI systems observed in traffic or declared by a person, with owner, purpose, vendor, models and lifecycle status. An observation becomes a record only when a person confirms it.
  • Classification and reviews. Rules-based risk tiers from Low to Critical with a reason for every factor and no numerical score, plus privacy and security reviews with findings, conditions and a review date.
  • Approval. By an administrator who did not make the request, with conditions and an end date at most three years ahead. Approvals read "review required" when facts change. Nothing is approved automatically.
  • Controls and evidence. An admission gateway applies policy to routed traffic, such as approved model lists, prompt DLP and team budgets, and an append-only, hash-chained evidence log records governance decisions. Control mappings with computed status cover the EU AI Act, NIST AI RMF and ISO/IEC 42001.

MERIDVAR does not certify compliance, controls apply only to traffic routed through the gateway, and the product is in beta. See AI governance for the details.

Sources: EU AI Act, Article 3 definitions; NIST AI Risk Management Framework; NIST AI RMF Core; ISO/IEC 42001:2023.