An interviewer says - our data governance programme is eighteen months old. We have a catalogue, a glossary, twelve part-time stewards and a monthly council. The CFO asked last week whether it is working and I could not answer. What would you measure, and what would you tell leadership if the numbers are bad?

What the interviewer is testing

Whether you can tell the difference between a programme's outputs and its outcomes, and whether you will say an uncomfortable number out loud. Almost every candidate lists artefacts: catalogue coverage, glossary term count, stewards appointed, policies published. Those are counts of work done, and a programme can increase all four while nothing about the business's numbers changes.

The clarifying questions that change the answer

• What decision went wrong that started this? A programme funded after a misstated regulatory return needs different evidence from one funded after an argument about active users.
• Is there an enforcement point anywhere? A policy no system checks cannot have an effect, so the measurement question is premature until one control is binding.
• Who is allowed to be told the programme is failing? If nobody is, the measurement will be gamed within a quarter, and you should say so.

A strong answer's arc

Four measures, all cheap, all sampled rather than reported by the people being measured.

1. Time to find the authoritative number. Take five real questions the CFO has actually asked, hand each to an analyst who has not seen it, and time how long until they produce a number plus the asset it came from. A working programme puts this in minutes. Eighteen months in, two hours or "I asked Priya" is the common result, and it is the single most persuasive number you will have.
2. Definition drift. Sample 20 glossary terms, read the SQL that actually implements each one, and count how many still match. Twenty terms at roughly 30 minutes each is 10 hours a quarter. Expect between a third and half to have drifted if nobody has been re-reading them, because a definition is code and code changes while prose does not.
3. Traceability of asserted numbers. Take the last three board decks and count what share of the figures can be traced to a governed definition. This converts governance from an IT activity into a statement about what leadership is making decisions on.
4. Standing entitlement. Count permanent read grants on the estate's restricted tables and the median age of those grants. A programme with no effect on this number has not touched access at all.

What to tell leadership if the numbers are bad

Say the numbers, then propose cutting scope rather than adding governance. The honest framing: twelve people at 20% is 2.4 full-time equivalents against a 1,200-table estate, which is about half a day per table per year — enough to maintain roughly 30 to 60 definitions properly and nothing like enough to maintain a catalogue of everything. Pick the 25 numbers the business actually asserts, give each one owner and one implementation that every tool reads, and let the rest be explicitly ungoverned.

Common weak answers

• "Measure catalogue coverage and glossary completeness." Both rise as a harvester runs and neither is evidence of anything. A catalogue at 95% coverage of stale descriptions is worse than one at 20%, because it teaches people that the catalogue is wrong.
• "Survey the stakeholders." Satisfaction tracks how recently someone was helped, not whether numbers reconcile, and it moves when the programme runs a roadshow.
• "We need a maturity assessment." A score out of five measures resemblance to a framework.

What a strong answer adds

The organisational layer: governance that has no consequence attached is a documentation project with a committee, so the first thing to build is one binding control — usually a single governed metric set that the BI tool refuses to let anyone redefine. Also the reversal condition: if the estate is small enough that one person holds every definition in their head, the right recommendation is to stop the programme and keep the owner, because the measurement cost alone would exceed the benefit.