A data platform finds the fault. It does not change behaviour.

A data contract names an owner. A test on the platform reports when reality drifts away from the agreement. And then an alert sits there, asking whether anyone will act on it.

A platform can find every deviation. It cannot force anyone to fix them.

That is not a technical problem. It is a matter of change.

A test finds the fault; it repairs nothing

Testing happens on the data platform: in notebooks, in dbt, in Databricks. Those tests find problems reliably. But a problem found is not yet a problem solved.

A test can report the same deviation day after day. If nobody is moved to act, the alert simply stays there. People get used to it and stop looking. The platform keeps doing its job; nothing changes.

Every step works as designed. Only the last step, the fix, depends on a person who is not held to account for it.

One fault does not stay one fault

Until recently, a human read the wrong number. That person could frown, hesitate, ask around. An agent does not. Agentic AI does not only look at data, it acts on it: it orders, books, decides, passes on.

A wrong number then carries exactly the same confidence as a right one. And where the data is incomplete or contradictory, the model fills the gap itself. What comes out sounds plausible and rests on nothing.

Agents also work in chains. One agent's output is the next one's input. A single wrong definition at the bottom does not stay one fault; it travels through every step that trusts it. That is how a small defect in the data grows into a large, automated, confident failure. And the human who used to catch the odd number is no longer in the chain.

With no human in the chain, a small defect in meaning becomes a large automated decision.

The speed and autonomy that make agents valuable are what make bad data dangerous. Under agentic AI, governance of meaning is not a luxury. It is the precondition.

A vitamin moves nobody; a painkiller does

As long as data quality is "good for the organisation", it stays a vitamin. Everyone nods, nobody feels the cost of postponing it today. It is useful, and therefore easy to ignore.

Data quality only becomes a painkiller when ignoring it hurts. When a wrong number lands with the person who is judged on it. People act on pain they feel, not on a virtue they endorse.

The same subject, two positions in the organisation. Only the right-hand column moves.

It becomes a task once it counts

Data quality turns from vitamin into painkiller the moment it is part of the job. Not something added when there is time left over, but a fixed part of the role, visible in performance management.

As long as an owner exists only inside a contract, ownership is paper. Once guarding a definition weighs into how someone is reviewed, ownership becomes real. That link to performance management is what turns governance from a schema into behaviour.

Platform on one side, people on the other

Data quality that lasts asks for two sides. On one side the platform: the tests, the contracts, the enforcement. That is process and platform, and it is where Datahub focuses. On the other side the people: the roles, the incentives, the behaviour. That is people and process, and there are partners who specialise in it.

They meet at the process. The platform makes the agreement measurable and enforceable. The partner makes sure people start playing their part in it. Neither side gets there alone.

On the left what you can install, on the right what you have to organise. Measurable and important are two different things.

What is left

A platform makes data quality measurable. It does not make it important. Whether it becomes important is decided where a wrong number becomes someone's problem: in the job, in the review, in the work itself.

That is not an installation. It is change. A platform makes data quality measurable. Only an organisation makes it important.