A data team is asked to enable self-service analytics while ensuring numbers are trustworthy. How should the boundary be drawn?

The boundary

Govern the small set of metrics that carry consequence; leave everything else self-service and explicitly ungoverned.

Governed: the numbers in board reporting, regulatory filings, team objectives and customer-facing commitments. Typically twenty or so, defined once in a semantic layer with a named owner.

Ungoverned: everything else — exploration, hypothesis testing, one-off questions. Fast, flexible, and clearly labelled as not authoritative.

Why attempting to govern everything fails

It becomes a bottleneck and is abandoned. Every new question requires a definition, a review and an implementation, so analysts route around the platform — and the governance then covers a shrinking share of the analysis that actually happens.

Governance that is routed around provides less assurance than none, because it exists on paper and not in the system.

What makes the ungoverned tier safe

• Clear labelling, so a number from an ad-hoc query is visibly not the governed metric.
• Access to the same underlying data, so the ungoverned tier is not working from a worse copy.
• A promotion path: when an ad-hoc metric becomes important, it is adopted into the governed set with a definition and an owner. Without that path, important metrics stay ungoverned because governing them is a project.
• Cost quotas, since analytical workloads have no natural ceiling and the ungoverned tier is where the expensive queries live.

The failure that governance is supposed to prevent

The same metric reported differently by three teams, each confident, with no way to resolve it by inspecting the numbers — because the definition lives in three places and each encodes slightly different assumptions.

That is what the governed tier exists for, and it is why the governed set should be defined by which numbers cause that argument rather than by which are most queried.

The honest note

Some disagreements are genuinely about the business question. If finance and product mean different things by "active customer", the answer is two governed metrics with distinct names — not one definition imposed on both, which produces a number neither team trusts and which reappears as a shadow metric within a quarter.