You cannot lead what you do not measure

Say the word governance in a leadership meeting and watch what happens to the room.

Shoulders drop slightly. Someone checks their phone. The word has been so thoroughly captured by the compliance function that most executives now hear it as a synonym for restriction — a committee, a register, a policy document that someone will circulate and nobody will read, existing mainly so that the organisation can demonstrate it existed if anyone asks.

That version of governance is real, and it deserves the reaction it gets. But it is a degraded form of something else, and the something else is the reason this matters now.

Governance, done properly, is not a brake. It is an instrument panel.

It is the mechanism by which a leader can see what is actually happening in their organisation's data, hold specific people to specific standards, and know — rather than hope — whether the system they just deployed is doing what they were told it would do. Without it you are not leading the AI programme. You are receiving reports about it.

What governance is actually for

Strip away the vocabulary and governance answers four questions that a leader cannot function without.

• Who decides what this number means.
• Who is accountable when it is wrong.
• How would we know if it were wrong.
• What happens next when we find out.

That is the whole thing. Everything else — the councils, the charters, the maturity models — is machinery for producing reliable answers to those four questions at scale. When the machinery becomes the point, governance turns into theatre. When the questions stay the point, it turns into control.

The reason this became urgent rather than merely advisable is that AI changes the consequence of not knowing. A dashboard with a wrong number is read by a person who brings judgement and context and often catches it. A model with a wrong input produces thousands of decisions at machine speed, each one individually plausible, and nobody reads them at all. The error does not announce itself. It accumulates.

So the tolerance for ambiguity that most organisations have lived with quite comfortably for twenty years — three definitions of active customer, a field everyone assumes is populated, a pipeline that depends on one helpful person — stops being a mild inefficiency and starts being a live exposure.

The first thing governance produces is sight.

Ask most senior teams what data their organisation holds, where it lives, who can reach it, and how good it is, and you will get an answer that is confident at the top and vague within two questions. Not because anyone is hiding anything — because nobody has ever been made responsible for maintaining the picture, and the picture decays continuously as systems change, people leave, and workarounds accumulate.

This is the difference between an inventory and a map. An inventory is a list of systems, and most organisations have one somewhere. A map tells you which data feeds which decision, how it gets there, what condition it is in, and where it breaks. Almost nobody has that, and it is the thing you need before you can make any sensible judgement about what is buildable.

The practical effect of having it is that conversations get shorter. When a proposal arrives, the question "do we have the data for this" has an answer in the room rather than a three-week investigation. Programmes that would have failed get stopped in the meeting rather than in month nine.

The second thing governance produces is a name next to a number.

Most organisations have accountability for systems and almost none for data. There is an owner of the CRM. There is rarely an owner of what "customer" means inside it. So when the definition drifts — and it always drifts, quietly, as new fields are added and new teams start using it for new purposes — there is nobody whose job it is to notice.

Naming an owner sounds bureaucratic and is in fact the single highest-leverage act available. It changes drift from something that happens to something someone is answerable for. It gives every downstream user a person to ask. And it surfaces, immediately and sometimes uncomfortably, which of your critical data assets currently have nobody attached to them at all.

The third thing governance produces is the least tangible and the most valuable.

Every AI system eventually faces a moment where its output contradicts what an experienced person believes. That moment decides whether the system is adopted or quietly ignored. And what determines the outcome is not the model's accuracy — nobody in that room can assess the model's accuracy. It is whether the organisation has any basis for believing the inputs.

If lineage is documented and definitions are owned and quality is monitored, the disagreement becomes an investigation. Someone traces it back and either finds a data problem or confirms the system saw something the person missed. Either way the organisation learns and confidence compounds.

If none of that exists, the disagreement becomes a matter of who is more senior. And the model loses, every time.

The experienced person has a track record. The system has an interface. This is what people are really describing when they say a deployment failed on adoption — it rarely failed on user experience. It failed because there was no way to establish whether it should be believed, so the safest thing for everyone was not to rely on it.

Governance that gets used

The failure mode is well documented and easy to walk into. A framework is adopted wholesale, a council is convened, an ambitious catalogue project begins, and eighteen months later there is a great deal of documentation and no observable change in how decisions are made.

What works better is narrower and less satisfying to announce.

Start with the decisions that matter, not the data estate. Pick the handful of numbers your leadership team actually runs the business on and govern those properly — definition, owner, lineage, quality monitoring, a route for raising a problem. Five done thoroughly beats five hundred catalogued shallowly, and it produces a working example you can point at.

Make it visible where the work happens. Governance that lives in a separate portal is governance nobody consults. It has to appear at the point of use, in the report and the dashboard and the model documentation, or it functions as an archive.

Measure the governance itself. How many critical elements have named owners. How long a data issue takes to resolve. How often definitions change without notice. If you cannot report on whether your governance is working, you have reproduced exactly the problem you set out to solve.

Give it teeth proportionate to the risk. A model informing a marketing sequence and a model informing a credit decision should not face the same gate. Uniform process is easier to write and reliably produces either a bottleneck or a rubber stamp.

Vivienne Umusu is Co-founder and Director of Kenvitek, a vendor-neutral AI data readiness and transformation practice working across Africa and the United States.