What is AI governance?

What it covers, and what it is not

Three terms get used interchangeably and mean different things. Keeping them apart is most of what makes a programme legible to the people who have to run it.

• AI ethics asks what a system should be permitted to do. It sets the intent.
• AI risk management asks what could go wrong, how likely it is, and what it would cost. It sizes the exposure.
• AI governance decides who is accountable, encodes the limits, enforces them, and keeps the evidence. It is where the other two become operational.

What a governance programme actually produces

A programme that cannot point at these has produced documentation rather than governance. Each is a record somebody can be asked for, which is the test that separates the two.

• An inventory of the AI systems in use, including the ones nobody registered.
• A named owner per system, and a stated purpose that later decisions can be measured against.
• A permission boundary — what the system may access and do, and under which conditions.
• An enforcement point that stops a prohibited action, as distinct from a policy that discourages it.
• A durable record of what was decided, by which rule, and when.

What changes when the AI takes actions

Governance built for models governs training data, evaluation and bias — the question is what the model outputs. That framing holds while a human reads the output and decides what to do with it.

It stops holding when the system holds credentials, calls tools and acts in production. The risk is no longer what the model said but what the software did, and the governing question changes from "is this output acceptable" to "should this action be allowed, right now, on behalf of this person". That is a question about authority and enforcement, and it has to be answered before the action rather than reviewed after it.

Why programmes stall

The common causes are organisational rather than technical, and they repeat across organisations that otherwise have little in common.

• Ownership sits with a committee, so no individual can be asked why a system exists.
• Scope is drawn around approved systems, which leaves the unapproved ones ungoverned by definition.
• Controls terminate in documentation, with nothing in the running system that enforces them.
• Evidence is assembled for an audit rather than produced continuously, so it describes a moment rather than a practice.

Common questions

Is AI governance the same as AI ethics?

No. Ethics sets the intent — what a system should be permitted to do. Governance decides who is accountable for that intent, encodes it as limits, enforces those limits, and keeps the record. A programme can have well-articulated ethics and no governance at all.

Who owns AI governance in an organisation?

The programme usually sits with risk, security or a dedicated AI governance function, but the part that matters is per-system: every production AI system needs a named individual accountable for its purpose and its authority. Ownership held by a committee is ownership nobody can be asked about.

Where do you start?

With an inventory, because you cannot govern what you have not found, and discovery usually reveals systems nobody had registered. Ownership and purpose come next, then permission boundaries, then enforcement. Starting with enforcement means blocking things you cannot yet see.

Does AI governance apply to agents differently?

The accountability and evidence questions are the same. What differs is that an agent acts — it holds credentials and calls tools — so the limits have to be enforced at the moment of the action rather than assessed in review. Governance that only inspects outputs does not cover an agent.