AI governance is not a policy document. It is a chain of evidence.
Most organisations can produce an AI policy on request. Very few can produce, for any given model, the evidence that the policy was followed. That gap is what governance actually has to close.
Ask a board what its AI governance looks like and you will usually be handed a document: principles, a committee, a risk appetite statement. Ask which models are in production, who approved each one, against what tests, and what would trigger a withdrawal — and the room goes quiet. The document exists; the chain that connects it to a running system does not.
Governance is a property of systems, not of paper
Governance is present when three things are true for every model that touches a decision about a person or an asset. First, someone with authority and without a delivery incentive approved it. Second, the approval rested on evidence that can be re-examined later — test sets, thresholds, results, the version of the model that was tested. Third, there is a defined condition under which the approval lapses, and a mechanism that notices when it does.
Everything else — committees, charters, training — is scaffolding for those three facts. Useful scaffolding, but scaffolding.
Why the evidence chain breaks
It breaks for ordinary reasons. Models are retrained and the evaluation is not re-run. A vendor updates a foundation model underneath a product and nobody is told. The person who approved a system leaves and the approval outlives their understanding of it. A threshold is quietly relaxed to ship on time. None of this is malice; all of it is invisible to a policy document.
A policy tells you what should have happened. Evidence tells you what did.
What good looks like
A register of models in production, each with a signed assessment that names its tests, its results and its version hash. A separation between the people who build and the people who approve. Monitoring wired to the same thresholds the approval was granted against, so drift is a governance event, not a data-science curiosity. And an independent party — internal audit or an external validator — who can reproduce the tests without asking the builders for help.
That is a smaller list than most frameworks suggest, and a harder one to fake. It is also the list a regulator will work through when something goes wrong.
