AI Governance Is Enforcement, Not Intent

Every organisation I talk to has AI principles. Fairness, transparency, accountability — nobody argues with them, which is the first clue that they are not doing much work. Ask a narrower question and the room goes quiet: when an engineer ships a model on a Friday afternoon, what in your organisation actually stops them if it should not go out?

That silence is the whole subject. A principle that cannot point to the moment it says no is not governance. It is a statement of preference, and most organisations have several.

So the test I apply to any governance claim is: show me where it bites. A control is enforced by a system or by a person with the authority to refuse, it triggers at a specific point in the workflow, and it leaves a trace. “We value transparency” is an intention. “No model reaches production without a recorded evaluation and a named owner” is a control. The first one survives a board deck. The second one is what is true at four o’clock on a Friday.

Governance versus velocity is the wrong argument

The binary everyone reaches for is governance against delivery speed — the committee that says no to everything on one side, the free-for-all on the other. It is the wrong axis. What actually determines whether governance costs you anything is where the control sits and what it is priced against.

A control built into the delivery path costs real money to build and maintain — an evaluation harness is a system, with its own failures and its own exceptions to handle — but what it does not charge per change is coordination. A control that sits beside the path, waiting for someone to remember it and find a slot in a diary, bills every single change a scheduling round-trip instead, and that is the cost teams learn to route around — a control that is routed around is not slow governance, it is no governance with extra reporting. The variables that matter are the same three every time: how reversible the decision is, how far its blast radius reaches, and how sensitive the data underneath it is.

Where governance sits organisationally matters less than people expect. The shape I keep arriving at is a small central function owning standards and hard escalations, with delivery teams applying those standards in context. The centre sets the rails; the teams run on them. Treat that as a preference I can defend rather than a finding — there is no survey behind it, and the right split moves with how much delivery capability the teams already have. That is a structural question with its own mechanics — decision rights, evidence, enforcement points — and it deserves more room than a section here.

Data governance is the floor, not a sibling track

Here is the part that gets skipped for being unglamorous, and it is the one that decides whether any of the above is achievable: you cannot govern AI on top of data you do not understand.

Every hard AI question resolves down to a data question. Is this fair — trained on what. Is this private — collected under which basis. Can we explain it — derived from which records. Are we even allowed to use it — consented by whom, retained how long, and was it ever supposed to be in the training set.

Governance without data lineage is an audit with no ledger. The auditor can ask any question they like; nothing in the building can answer it. That is not a documentation gap you close during an incident — the record either exists from the beginning or the answer does not exist at all.

Organisations that treat data governance as a lower-priority track end up bolting fairness and privacy onto a foundation they cannot see into, and the first serious incident is where they discover it. Get the data layer right and a large amount of AI governance becomes enforceable almost mechanically, because you finally know what you are working with. That is the single highest-leverage thing on this list, and it is the one that never makes the strategy slide.

Price the gates by reversibility

Not every AI decision needs a human in it. Insisting otherwise is how governance becomes a synonym for slow, which guarantees it gets worked around. The skill is putting human gates exactly where they earn their cost.

Sort by stakes and reversibility, the same way you would sort actions for an autonomous agent. A recommendation that is easy to ignore or undo can run on its own. A decision that materially affects someone — credit, employment, healthcare, anything legally consequential, anything you cannot take back — gets a human who can genuinely understand the recommendation and overrule it, with that override logged and reviewed.

The failure mode is the reviewer who approves everything, because they do not understand the model and will not be thanked for slowing delivery down. A gate that always says yes is theatre with a headcount. If you are going to put a person in the loop, give them the information and the standing to actually be in it.

Proportionality is not an invention of mine. The EU AI Act (opens in a new tab) sets its rules by risk category rather than applying one standard to every system — that is the Regulation’s own stated approach, and it is the part worth copying whatever jurisdiction you are in. A model drafting internal meeting summaries does not warrant the scrutiny of one deciding who gets a loan. Spend the scrutiny budget where the blast radius is real.

I would not pretend the tier boundaries are obvious in practice. They move as a system gains reach — an internal summarisation tool that starts being pasted into customer emails has changed tier without anyone filing anything — which is why the tier belongs on the deployed system and gets re-asked when its use changes, not stamped once at approval.

Make the governed path the easy path

The naive version of governance is a tax on everything: every project waits for the committee, the committee becomes the bottleneck, the business routes around it, and you end up slow and ungoverned.

The way out is to make compliance the path of least resistance. Evaluation harnesses in the deployment pipeline so checks run without anyone convening. Pre-vetted components and approved patterns, so the compliant choice is also the fastest one. Templates that carry the requirements with them. When the safe option is also the quickest option, people take it without being policed.

The three questions

Stop asking whether you have AI principles. Almost everyone does, and they are almost interchangeable. Ask instead: when this goes wrong, who is accountable, what stops it happening again, and can you prove what the system did?

Clear ownership, enforced controls, an audit trail you would stake your reputation on. Answer all three and you have governance. Answer none and you have an aspiration with good production values — and the work of closing that gap happens in the delivery pipeline and the data catalogue, not in the document.