Most decision latency is an argument about whose number is right

Sit through a decision that takes three meetings instead of one and the pattern is usually the same. Two teams arrive with different numbers for the same thing. Neither is wrong exactly; they are measuring slightly different populations over slightly different periods. Nobody present has the authority to declare which figure governs, so the decision is deferred and someone is asked to reconcile them. That reconciliation is the delay, and it is a data ownership problem wearing the costume of an analytical one.

Ownership assigned to a committee is ownership assigned to nobody

The test of whether ownership is real is simple: when a dataset is wrong at two in the morning, is there a name attached to it? A working group is not an answer. Neither is a team, unless one person within it holds the decision.

Stewardship also has to come with authority. A steward who cannot reject a change, hold a release, or declare one figure canonical is performing documentation rather than governance, and the role degrades into maintaining a catalogue nobody consults. Decision rights are the substance; the title is not.

When a dataset is wrong at two in the morning, is there a name attached to it? A working group is not an answer.

Draw the boundaries along the business

Ownership drawn along system boundaries fragments the moment a platform is replaced, and platforms get replaced several times over the life of the data. Drawn along business domains — customer, product, transaction, entitlement — it survives the replacement, which is the entire point.

This is also what makes ownership legible to the people who need it. "Who owns customer data" has an answer a director can act on. "Who owns the customer table in the warehouse" has an answer only the data team can use, and only until the warehouse changes.

Lineage is what ends the argument

Ownership says who decides. Lineage is what lets them decide quickly. When two figures disagree and the derivation of each is visible — source system, transformations applied, filters, refresh time — the reconciliation takes minutes rather than a week, because the difference is visible rather than hypothesised.

That is the practical argument for treating lineage as an operational property rather than documentation. Documentation describes what should be true. Lineage captured from the pipeline describes what is true, and only the second one settles a disagreement between two directors.

What ambiguous ownership costs to run

Unclear ownership is expensive in a way that never appears as a line item, because it is paid in analyst time and duplicated infrastructure. When a team cannot establish whether an existing dataset is authoritative, the rational move is to build their own — their own extract, their own transformation, their own refresh schedule. It is faster than the alternative, which is finding somebody willing to vouch for the existing one.

Repeat that across a few years and the organization is running many pipelines computing approximately the same thing, each with its own compute cost, its own storage, its own on-call burden, and its own capacity to disagree with the others. The disagreements then generate the reconciliation work described above, which is the recurring labour cost sitting on top of the recurring infrastructure cost.

This is a more tractable place to start than a catalogue programme. Take the handful of measures that appear in board reporting, establish a named owner and a canonical derivation for each, and count how many pipelines currently produce a version of them. The redundant ones can be retired with a specific saving attached, and the exercise generates its own evidence for extending the approach.

The AI reason to do this now

Every AI programme eventually becomes a data programme, and the organizations that get there faster are the ones that already knew who owned each dataset, where it came from, and whether it could be trusted. A model grounded on a corpus nobody owns will produce confident answers derived from documents nobody has checked — and unlike a dashboard, it will do so in fluent prose that invites belief.

The governance question, therefore, is not new. It is the same ownership and lineage work that would have shortened decision cycles anyway, now with a second and more urgent reason to fund it. That is a more honest business case than presenting it as an AI prerequisite, and it survives better if the AI programme is paused.