What Is AI Governance?

AI governance is how an organization stays in charge of its AI. It sets out who may approve a new use, which risks must be assessed first, which data and models are allowed, where people must review outputs, and what evidence has to exist afterwards. Its scope is wider than any one model: it covers systems built in-house, systems bought from vendors, and AI features that arrive inside software already in use.

A practical definition has four parts. Policies state what is permitted and what is not. Roles name the owners, reviewers and approvers, from an executive sponsor down to the engineer who ships a model. Processes such as intake, risk classification, testing and incident response make the policies repeatable. Technical controls enforce the rules inside the systems, so a ban on sending customer records to an external model becomes a blocked request instead of a sentence in a handbook.

Three pressures have turned governance from an ethics exercise into an operating requirement. Regulation is in force: the EU AI Act entered into force on 1 August 2024, its prohibitions have applied since 2 February 2025 and its rules for general-purpose AI models since 2 August 2025 (European Commission overview, verified September 2026). Adoption is decentralized, so AI reaches the organization through many teams and suppliers at once. And AI systems increasingly act, which moves the risk from a wrong answer on a screen to a wrong change in a production system.

Governance is also what lets adoption scale. Without an agreed intake and approval path, every new use case turns into a one-off negotiation between the business, security, legal and IT. With one, low-risk uses move quickly and scrutiny goes to the few cases that deserve it.

The NIST AI Risk Management Framework (AI RMF 1.0), released on 26 January 2023, is voluntary guidance built around four functions: Govern, Map, Measure and Manage. NIST added a Generative AI Profile, NIST AI 600-1, in July 2024, and its site notes that AI RMF 1.0 is being revised under the White House AI Action Plan. Most teams use it as a shared risk vocabulary and a catalogue of practices.

ISO/IEC 42001:2023, published in December 2023, specifies requirements for establishing, implementing, maintaining and continually improving an AI management system within an organization. Because it is a management-system standard, accredited bodies such as DNV certify organizations against it, which helps when customers or auditors want independent assurance.

The EU AI Act, Regulation (EU) 2024/1689, is binding law rather than guidance. It sorts AI systems into four levels: unacceptable risk (prohibited), high risk, transparency risk and minimal risk. Most of its provisions have applied since 2 August 2026. Amendments known as the Digital Omnibus on AI took effect on 27 July 2026 and pushed the high-risk obligations back to 2 December 2027 for stand-alone systems listed in Annex III and to 2 August 2028 for AI built into regulated products (Commission notice).

The OECD AI Principles, adopted in 2019 and updated in May 2024, are the intergovernmental baseline: five values-based principles, among them transparency, robustness and accountability, plus five recommendations to governments. They are short and non-binding, and many national approaches trace back to them.

The four fit together rather than compete. The OECD principles describe what good looks like, the NIST AI RMF describes risk practices, ISO/IEC 42001 provides an auditable management system, and the EU AI Act sets legal duties for systems placed on the EU market or used in the EU. The framework for regulated industries maps those duties to specific controls.

Organizational governance is where most programs begin: an AI policy, an inventory of AI systems, a risk committee, an approval matrix and a training plan. It settles who decides and on what basis. Its weakness is distance from the systems. A policy can require human review for high-risk uses, but it cannot make a deployed system pause for that review.

Technical governance closes the gap by enforcing decisions where the AI runs. Typical controls include an allowlist of approved models, data-classification rules that decide which requests may leave the network, per-role access to agents and tools, approval gates on consequential actions, rate and budget limits, and logs detailed enough to reconstruct a decision. Runtime enforcement also produces the evidence: a record of what actually happened, not only what the policy said should happen.

Mature programs join the two halves. The inventory kept by the governance team is the same registry the runtime checks, and a policy change becomes a configuration change that takes effect everywhere at once.

Agents plan, call tools and change records, which raises questions a chat assistant never did. Which tools may this agent call, and on whose authority? Which actions need a person to approve them first? Can the organization reconstruct a multi-step run, including every retrieval and tool call, weeks later? Can it stop a misbehaving agent at once?

Those questions have their own treatment in the AI agent governance framework, with controls and a checklist for agent deployments. Standards such as the Model Context Protocol make agents far easier to connect to enterprise systems, which is exactly why tool access now needs the same care as data access.

Most organizations move through four stages. Ad hoc: teams adopt AI with no shared inventory or policy. Defined: there is an AI policy, an accountable owner, an intake process and a register of AI systems with a risk class for each. Enforced: the rules exist as controls in the platforms where AI runs, including model allowlists, data boundaries, tool grants and approval gates. Evidenced: logs and reports can show a regulator, auditor or customer what each system did and why, and incidents follow a tested playbook.

The step from defined to enforced is where programs stall most often, because it takes engineering work rather than documents. A good first move is to pick one valuable use case, write down the controls it needs, and implement them end to end before scaling the pattern. AI governance tools differ mainly in which of these stages they automate, and the governance framework lesson walks through setting up roles and approvals.

Four common reference points and what each is used for. Dates checked against primary sources in September 2026.