Ask three people at a PE-backed multi-location company how many active customers it has, and you have asked them a policy question without meaning to. The finance lead answers with the people billed this month. The head of operations answers with the people who showed up. The location managers answer with whatever their dashboard says when they open it, which depends on a default somebody set years ago and nobody remembers setting.
Each of those answers is defensible. Each one is a decision about what the business counts. Until somebody makes that decision on purpose, the company has several numbers sharing one name, and a board meeting every month where they collide.
Two Reports, One Metric, Two Answers
Picture the week before a board meeting at a sponsor-backed company with locations across several states. The CFO pulls the active customer report: 4,100. The COO pulls hers: 4,650. (Illustrative numbers; the shape is the point.) Both can walk you through their logic line by line. They spend the first forty minutes of the prep meeting reconciling, settle on a footnote, and agree to sort it out before next month.
Next month, the same forty minutes.
Underneath a gap like that, finance usually counts a customer as active when there is a paid invoice in the period. Operations counts anyone with a scheduled visit and no cancellation on file, which sweeps in free trials and people who signed up last week for a service that starts next month. Both definitions are reasonable. They answer two different questions: who is paying us, and who are we serving? The trouble starts when both land on the same slide under the same label.
The board reads the disagreement as a data quality issue and asks for a new dashboard. The new dashboard inherits whichever definition its builder happened to pick, and now there are three numbers.
Why a KPI Definition Is a Policy Decision
A metric definition reads like a technical spec: a filter, a date field, a join. Every line of it is a business choice. Does a trial count? Does a customer who prepaid for a future period count today? Does “as of today” mean this moment or last night’s data load? Does a customer who visits two locations count once or twice?
Each answer changes the number, and each one encodes a view of how the business works. Counting trials says this report cares about the funnel. Counting a customer once across locations says the company is one brand; counting them twice says each location runs as its own P&L. Those are management decisions. Translating them into SQL is the last step, and the analyst writing the query should be the last person asked to make them.
This is what I mean by merging the data model and the business model. The data model is the business model written down precisely enough that a computer applies it the same way every time. When nobody has settled the business question, the analyst settles it by default, usually without knowing a decision was on the table.
Acquisitions make this louder, which is why the case for settling a common metric layer at the first acquisition is so strong. The same war breaks out inside a single company on a single system, though, because the disagreement lives in people’s heads before it lives in a database. Two leaders who came up through different functions carry different mental models of the same word. A location manager built a habit around the report that makes her numbers look steady, and she would like to keep that report, thank you.
Tier One: A Metric Alignment Glossary for the Contested Metrics
The first tier is the one you can see coming: active customer, retention, revenue per location, utilization. If two leaders have ever argued about a number in a meeting, that metric goes on the list.
For each one, put the people who own the outcome in a room: usually the CFO, the COO, whoever runs the locations day to day, and whoever builds the reports. Then write down five things.
- The definition in plain language. One or two sentences a location manager would recognize as true.
- The edge cases. Trials, prepaid future periods, pauses, transfers between locations, refunds, the customer who cancels and comes back within the month.
- The time basis. Active during a period and active as of a specific date are both legitimate, and they produce different numbers. The slide has to say which one it shows.
- The system of record. Where the number comes from, and which system wins when two of them disagree.
- The decision it drives. The question operators skip most often. If nobody can name a decision that changes when this number moves, it comes off the board pack.
Get sign-off before anything gets built. The glossary is the artifact, and the data model follows it.
Expect dozens of definitional questions on even a short list. Most are small. A few turn out to be surprisingly political, and those are the valuable ones. When the COO and the CFO disagree about whether a prepaid customer is active, they are disagreeing about how to run the business. Better that it surfaces in a conference room than in the board meeting.
Tier Two: The Definitions Only a Wrong Number Reveals
A good glossary catches the arguments you can anticipate. Most definition projects stop there and miss a second tier that never makes the list, because nobody knew those definitions were contested until the data made them visible.
The method for finding them is simple and a little uncomfortable. Put a real number on screen in front of the person who knows the business best, and watch their face. Three patterns come up often enough to plan for.
The count that comes back several times too high. A year-to-date customer count lands far above anything the operator believes. The trace shows a returning customer counted as new every time they renewed for another period. Nobody wrote “count each customer once across the year” in the glossary, because nobody imagined counting them any other way. The wrong number forced the definition into existence.
The active figure that moves with a date filter. The active count swings depending on whether people signed up for a future period are included. Both versions hold up, and they differ materially. Toggling the filter live and asking which number matches the operator’s mental model of the business settles it.
The ratio with the wrong numerator. A revenue-per-service-hour metric looks fine until someone notices a month where revenue arrived with almost no hours behind it: sign-up fees and deposits collected before any service was delivered. The timing side, where cash collected and revenue earned run on different clocks, is its own subject. The definitional side is what belongs in the ratio at all, and the answer is revenue earned from delivered service.
The most useful sentence in any of these sessions is the operator saying “that can’t be right.” That is the moment a hidden policy question becomes visible enough to answer. Treat it as progress.
Running the Live Reconciliation Loop
The practical implication is a budget line. Put the second tier on the calendar on purpose, so it surfaces in a working session and stays out of the board deck.
Build a first version against real data quickly, while the glossary is still fresh. Then schedule working sessions where the domain expert reviews numbers on screen, metric by metric and location by location, with the person who knows a given location can’t possibly have that many customers in February looking at the actual figure.
When a number looks wrong, trace it to the definition that produced it, decide, and write the decision into the glossary with the date and the role of whoever made the call. Six months later, when someone asks why returning customers count once, the answer is on paper.
Repeat until the numbers stop surprising the people who run the business. Each round is shorter, because each session retires a few more hidden questions. Once the definitions hold, they can travel: cascading board-level KPIs down to the measures each location manager owns only works when the word at the top of the cascade means the same thing at the bottom.
Why AI Makes the KPI Definition Problem Worse
Before an AI tool can answer questions about your business, the business has to agree on what its metrics mean.
Ask an AI assistant connected to your data how many active customers you had last quarter. It finds the invoice table and the schedule table, picks one, and answers 4,100 or 4,650 with no hint that the other number exists. It applies that definition consistently and presents it with the same polish whether it matches your business or contradicts it. The forty-minute reconciliation at least made the disagreement visible. The assistant resolves it quietly, and nobody questions the number.
Built on a signed glossary and a data model that encodes it, the same assistant answers the way your leadership team would. The glossary is the business half of what data teams call a semantic layer: the definitions an AI tool, a dashboard and an analyst all read from. An AI answer is only as accurate as the business logic it inherits, which is why you have to fix your data before implementing AI. Every company planning to put AI on top of its reporting is about to run a definition project, deliberately or one wrong answer at a time.
Settle the Predictable Questions, Then Find the Rest
The messy middle after a close is full of numbers that almost agree. Most of the time, the gap between them is a decision nobody made. Make the predictable decisions first, in a room, on paper, with sign-off. Then put real numbers in front of the people who run the business and let them find the rest.
One exercise to run this week: take the metric that opens your board deck and ask three leaders to define it in writing, separately, without comparing notes. If the three answers match, move to the next metric. If they differ, you have the first entry in your glossary.
Alex Escoriaza helps PE-backed and independently sponsored companies write their metric definitions, build the data model that follows them, and run the live sessions that catch the definitions nobody saw coming. If your board deck opens with a number two leaders would define differently, start there.
