Five versions of the truth

The results meeting starts at nine sharp. At 9:05, sales puts up the month's revenue. At 9:07, finance raises a hand: their number is 8% lower. Operations brings a third. The CEO remembers a fourth, from Friday's email. At 9:45 they're still trying to work out why the numbers don't match. Nobody has yet talked about what to do.

The scene, a composite but a familiar one, has an uncomfortable twist: each team may be answering a different question. Data can also be incomplete or cut off at different times, so check that before choosing a figure.

Three questions sharing one name

Sales counts what was invoiced. Finance counts what's booked, net of credit notes. Operations counts what was shipped. In this scene, all three queried consistent data and answered different questions. The problem is that the company never decided what, exactly, “revenue this month” means. In real operations, also check data quality, cut-off times and filters.

If two people can read the same metric two ways, the metric doesn't exist yet. A word does.

The cost goes beyond the lost meeting. Decisions get postponed “until we have the right number”. Teams start distrusting each other and building reports to defend themselves rather than to decide. And leadership, tired of the argument, ends up deciding on gut feeling.

Anatomy of a good definition

A useful definition is more boring than it sounds, and that's where its value lies. It says what's counted and what isn't, with the edge cases written down: returns, credit notes, sales between related companies. It says which system it's calculated from and when the number is considered closed. A named business owner approves changes; authorized team members can implement them with review, versioning and a record. It should support a concrete decision, task or obligation; if it has no clear use, question it.

Why AI makes urgent what used to be important

For years companies lived with implicit definitions because experienced analysts knew them. A good analyst may know from context that “sales” in the leadership meeting means net sales. An AI agent doesn't have that implicit agreement. Ask it for the number without a written definition and it may choose an interpretation and present it with apparent confidence.

Language models can speed up writing an answer, but they cannot replace defining and verifying the figure. If the question remains ambiguous, producing more answers only multiplies the review needed. You don't need to start with every metric: agree on the five most disputed ones first.

The scenes in these episodes are composites of common situations, not customer cases. Any figures in them are illustrative.