Three revenue numbers can all be correct

It’s the monthly review. Sales puts up the value of signed orders. Operations shows fulfilled orders. Finance shows recognised revenue. Three slides, three totals, and within ten minutes the meeting has become a debate about whose system is broken.

Nobody’s system is broken. Each team counted something real, using a rule they can defend. The trouble is that all three called it “revenue”.

No dashboard settles that argument, however good it looks. Neither does a language model. Ask an AI assistant for revenue without saying what you need it for, and the most fluent answer may simply hide which definition it picked.

Start with the decision, not the metric

The instinct is to standardise: one number, one truth, done. Resist it, at least at first. A sales capacity decision needs a view of demand. A fulfilment decision needs committed work and delivery dates. Financial reporting needs the accounting treatment finance owns. Collapse all three into one universal measure and you lose distinctions people genuinely rely on.

Instead, give each measure a precise name, an owner and a boundary. Write down what event is being counted, which date applies, how cancellations are treated, which currency basis is used and which entities are in scope. That detail belongs next to the number, not in a spreadsheet one analyst maintains and nobody else can find.

The UK’s Government Data Quality Framework describes quality as fitness for purpose. It’s a useful lens for a management meeting: yesterday’s complete finance extract can be exactly right for one decision and wrong for another.

Draw the bridge between the totals

Once the definitions are written down, reconciliation stops being an argument and becomes an explanation. Signed orders include deliveries scheduled for next quarter. Fulfilled orders include items still waiting for an accounting event. Returns, cancellations and exchange rates account for the rest. Show those categories, and keep the underlying records so anyone can dig in.

Don’t sweep what you can’t explain into a line called “other adjustments”. Give each unresolved difference an owner and keep their notes as the investigation goes. A visible exception is far more useful than a tidy total produced by a correction nobody can trace.

It also helps to separate errors from choices. A duplicated order is an error, so fix it. Which period a sale belongs to is a choice, so agree it. Send both to the engineering team and you’ve created a backlog they have no authority to clear.

A new definition is a business change

Acquisitions, new sales channels and revised commercial policies all quietly change what a measure means. Record when the change took effect and why. Decide whether past periods are restated or stay under the old definition. Skip that step and next year’s trend line will show growth that is really just a change in scope.

Lineage is what lets a reviewer trace a figure back through its transformations to the source records. The W3C’s provenance work describes how to record the entities, activities and people involved in producing data. For an executive the point is simpler: the history behind a number should be readable without opening a pipeline.

Give the AI the same definitions

A conversational interface should use the agreed measure and say which period and scope it used. When a question is ambiguous, a clarifying question is part of a good answer, not a failure. And if an answer blends incompatible definitions, a citation underneath won’t rescue the reasoning.

This is why a governed knowledge layer sits at the centre of Akashic. Agreed meaning has to travel with the records into Business Intelligence and Ask Akashic. More access to data only helps if the meaning survives the trip.