Newsletter Article
WHEN THE SAME METRIC MEANS DIFFERENT THINGS
Why “revenue,” “demand” and “abandonment” don’t mean one thing in your business — and why that’s not the data problem it looks like.
Every Monday morning, some version of this conversation happens in contact centres and boardrooms everywhere:
Sales says revenue is up 12% for the quarter.
Finance says it’s flat.
Operations says demand spiked last week.
Marketing says it barely moved.
Nobody in the room is lying, and nobody’s system is broken. They’re just answering different questions with the same word.
This is one of the quieter problems in enterprise data — quieter than a broken pipeline or a missing integration, because nothing actually fails. Every number on every dashboard can be completely accurate and the organisation can still walk away from a meeting with three different pictures of the same month.
The problem isn’t bad data. It’s undeclared context.
Different “versions of correct”
Ask sales for revenue and they’ll often give you booked revenue — deals signed, commitments made. Ask finance and you’ll more likely get recognized revenue — the portion actually earned under accounting rules for that period.
Ask an operations team for “demand” and they may count interactions. Marketing may count individual contacts. Customer service may count unique cases.
None of these numbers is wrong.
Each answers a real question that a real part of the business needs answered. The trouble starts the moment those numbers get compared, combined, or dropped into the same slide without anyone flagging that they were never meant to be the same measurement in the first place.
For years, this was manageable because a small number of trusted analysts sat between the raw data and the business. They knew which table finance actually used, which figure the board expected, and why last December’s number couldn’t be compared with January’s without an asterisk. That knowledge rarely lived in a document. It lived in people.
Natural-language analytics and AI assistants remove that buffer. When anyone can ask a question directly and get an instant answer, the institutional knowledge that used to sit between the question and the number needs to be encoded somewhere machines can use it — or the same undeclared-context problem simply happens faster, and at a scale no single analyst can catch.
It’s not just what the metric means. It’s who’s asking.
There’s a second layer to this that’s easy to miss: the same metric, with the same definition, can still require a different answer depending on who’s asking and what they’re about to do with it.
Take a question as ordinary as “which region generated the most revenue from our highest-value customers this quarter?”
A CFO asking that question wants a defensible, board-ready number — recognized revenue, reconciled, with no surprises if someone asks a follow-up question in the audit committee.
A regional sales director asking the same question wants something else entirely: booked revenue and pipeline detail, broken down enough to know which accounts to call this week.
A customer success lead, hearing the word “highest-value,” is really asking a different question underneath — which of those customers are also flagged as at-risk, because that’s the decision they’re actually making.
Same metric. Same underlying data. Three different — and equally correct — answers, because the role asking the question determines what “correct” actually means.
This shows up constantly in operational settings too. Take abandonment rate, a number that exists on nearly every contact centre dashboard. A team leader watching it in real time needs queue-level detail they can act on in the next ten minutes. A workforce planner needs a rolling weekly trend to get next month’s staffing right. An executive reporting to the board needs one clean, aggregated monthly figure with no queue-level noise attached.
None of them is misusing the metric. They’re each asking a version of the question their role is actually responsible for answering. A system — human or AI — that hands all three the same answer has technically retrieved the right number and still failed all three people.
Good data in, good insight out
There’s an old rule in data quality: garbage in, garbage out. It’s still true, but it’s no longer the whole story.
You can have clean, accurate, complete, real-time data — genuinely good data — and still get a bad or misleading insight out the other end, if that data arrives without the context that explains what it means and who’s asking. Good data is necessary. It has never been sufficient.
The more useful version of the old rule, for the AI era, is this:
Good data in, plus good context in, is what produces good insight out.
“Good context” here means three things arriving alongside the number, not bolted on afterward: an agreed definition of what the metric actually measures, an understanding of how that measurement relates to the other numbers around it, and awareness of who’s asking and what decision sits behind the question. Strip out any one of those three and the “good insight” part quietly stops being reliable, even though the underlying data never changed.
This is exactly the pattern showing up in recent research on AI and enterprise data. Testing across multiple frontier AI models found that giving a system explicit business semantics — rather than just the raw database schema — improved answer accuracy by 17 to 23 percentage points. The data being queried didn’t change between those two tests. Only the context around it did.
Turning this from a philosophy into a practice
None of this means every metric needs to be forced into one universal definition — that’s neither realistic nor actually useful. Booked revenue and recognized revenue are both legitimate; sales and finance are both allowed to be right.
What it means is that the definition needs to become explicit, and the role asking needs to become part of how a system decides which explicit definition to serve up.
The question worth asking inside your organisation isn’t “which number is correct?”
It’s “which definition is appropriate for this decision, asked by this person, right now?”
The question worth asking inside your organisation isn’t “which number is correct?”
It’s “which definition is appropriate for this decision, asked by this person, right now?”
Practically, that shows up as a small set of habits:
- Name the version, every time. “Revenue” should never appear on a report without a qualifier — booked, recognized, forecast — even if that feels repetitive at first.
- Document the difference once, centrally. The explanation for why sales and finance revenue diverge should exist in one place both teams can point to, not be re-explained in every meeting.
- Design for the question behind the question. Before building a report or an AI-ready dataset, ask what decision the requester is actually about to make — that’s usually a faster way to the right definition than starting from the data model.
- Treat role as an input, not an afterthought. The same query, asked by different roles, should be expected to return different levels of detail and different — equally correct — framings.
Get this right, and “how are we performing this month?” stops being a question that starts an argument about whose number is real.
It becomes a question that gets answered correctly, the first time, for whoever is actually asking.
This is part of The Data Advantage’s ongoing look at what it takes to turn enterprise data into something AI — and people — can actually trust. Read the full August issue, “From Data to Context: The Next Enterprise Advantage,” or explore how AskEmite is built to serve the right answer to the right role, every time.
