AI governance is the organisational machinery that answers four questions about every AI system you run: who approved it, what is it allowed to do, who is accountable when it misbehaves, and how would you prove all of that to someone outside the room. The definition matters because the word gets used for two thinner things — a written policy nobody operates, or a procurement checklist — and neither survives contact with a real incident. Governance that works is a running system: decision rights with names attached, an inventory that stays current, and controls that produce evidence as a side effect of operating.
The discipline grew up around models that produce content, so its classic concerns are data, bias, transparency, and acceptable use, organised by frameworks like NIST's AI Risk Management Framework and ISO/IEC 42001 and sharpened by regulation such as the EU's AI Act. That layer remains necessary. What it did not anticipate is software that *acts* — agentic AI holding credentials, calling tools, and writing to systems of record. An acting system needs governance borrowed from operations, not just ethics: identity, change management, autonomy levels, runtime enforcement. The practical translation of all of it into steps is the governing agentic AI guide; the framework selection problem has its own walkthrough.
Why the inventory carries the whole structure
Every governance question above silently assumes you know what AI you run — which is why maintaining an AI inventory is not bookkeeping but the load-bearing wall of responsible governance. The inventory is what turns "we govern AI" from an aspiration into a checkable claim: each entry names an owner, a purpose, the systems touched, and the controls attached, and each governance decision lands on an entry rather than into the air. Organisations discover the dependency in reverse, during an audit or an incident, when the first question is "how many of these do you have?" and the honest answer is a shrug. Building the inventory is deliberately the first guide on this site; everything else attaches to it.
Governance and its neighbours
Two boundaries keep the concept sharp. Governance is not security: security bounds what an attacker can make your AI do; governance decides what the organisation lets it do on purpose — each is unfinished without the other, and owning one does not discharge the other. And governance is not the framework document: NIST, ISO, and their kin organise the work and prove it to outsiders, but the work itself is decision rights, inventory, tiers, and enforcement. A certified management system wrapped around an ungoverned agent estate is paperwork with an incident inside.