Nobody can say how many customers we have
August 22, 2026
Not because the data is missing. Because there are three answers, each of them defensible, and no agreed place where the real one is written down.
Three answers, all correct
Someone senior asks how many active customers the business has. Sales says 4,100, counting anyone with an open opportunity or a contract in the last eighteen months. Finance says 3,600, counting anyone who has been invoiced this fiscal year. The product team says 5,200, counting anyone who logged in during the last quarter.
Nobody is wrong. Each number is a faithful answer to a slightly different question, computed correctly from a real system. What does not exist anywhere is a sentence saying which question the organization means when it says “active customer”. Because that sentence does not exist, the meeting spends twenty minutes on arithmetic instead of on the decision it was called for.
This is the elephant we call Fuzzy: no trustworthy, machine-legible truth. Noisy decisions, no data integrity, and no structure a crawler can parse. It is the quietest elephant in the herd, because at no point does anything visibly break.
The evidence
A bill nobody invoices, estimated by the people paying it
$5M+
a year — what more than one organization in four estimates poor data quality costs it
7%
put the bill at $25M a year or more
IBM Institute for Business Value, 2025.
Why is a data-quality bill so hard to see?
Because nothing ever arrives asking to be paid. There is no invoice, no outage, no line in the budget. The cost is disbursed in twenty-minute increments across every meeting that has to establish whose number is being used, in decisions deferred until someone reconciles, and in the quiet discount every experienced person applies to any figure they did not compute themselves.
Which makes the IBM figure interesting for a reason beyond its size. These are organizations estimating a cost that nobody bills them for. More than one in four of them still puts it above five million a year, with seven per cent at twenty-five million or more. That is not a long tail of unusually broken companies. That is the ordinary condition, priced by the people living in it.
Self-reported estimates deserve the usual caution: nobody audited these numbers, and an organization that has just been through a bad quarter may over-attribute. But the direction is not in doubt, and the interesting thing about a self-reported cost is that it is the cost people can already see. It is a floor, not a ceiling.
What does “quality” actually mean here?
Not accuracy in the abstract, and this is where most data-quality programmes go wrong. Chasing correctness for its own sake produces a two-year cleansing project with no named beneficiary. That is a Mirage, the elephant that solves the wrong problem perfectly.
The useful definition is narrower: is there one agreed statement of what a thing means, does it hold still long enough to be relied on, and is it written somewhere a machine can read it? Three answers to the customer count is not an accuracy failure. Every number was accurate. It is a legibility failure. There was no single definition to be accurate about.
That reframing matters because it changes what you fix. You do not need every record clean. You need the handful of concepts your decisions actually turn on to have one owner, one definition, and one place where that definition lives. And you need that place to be readable by systems, not just by the person who wrote it.
How do you know the Fuzzy is in your room?
Nothing breaks, so look for the compensating behaviour rather than the failure.
- The same question gets a different answer depending on who you ask, and everyone considers this normal.
- Experienced people quietly rebuild figures themselves before trusting them, and nobody treats that as a symptom.
- A definition exists in a slide from two years ago, and that slide is the authority.
- Reports carry a caveat about which source was used, because the caveat is doing real work.
- Nobody can say who owns the meaning of a core term, only who owns the system it lives in.
What has any of this got to do with AI assistants not finding us?
It is the same elephant wearing a second costume, and the connection is worth making explicit because the two halves are usually owned by different departments.
Inside the organization, the missing artifact is an agreed, machine-readable statement of what is true. That is why the customer count has three answers. Outside it, the missing artifact is exactly the same thing: an assistant asked about your company reads what you publish and answers on your behalf, and it can only do that where the claims are self-contained, sourced, and structured underneath. A page that gestures gets skipped in favour of one that states.
So an organization that cannot state its own numbers cleanly to itself is rarely able to state anything cleanly to a model either. The two symptoms, the meeting that argues about arithmetic and the assistant that does not know you exist, have one cause. Fixing the internal definition is what makes the external claim publishable, and both are cheaper together than either is alone.
How do you get the Fuzzy out of the room?
The same three moves we bring to any elephant, pointed at this one.
Map
Do not inventory the data. Inventory the disagreements: the terms that produce more than one answer, and which decisions each of them feeds. That list is usually shorter than everyone fears: a dozen concepts, not a warehouse. It is the only part worth fixing first.
Prove
Take the single most-argued term and settle it properly: one owner, one written definition, one place it lives, and a rule for changing it. Then check whether the argument actually stops. If it does not, the disagreement was never about data, and you have learned something more valuable than a clean field.
Scale
Make the definitions machine-readable and let them serve both audiences at once: the internal report and the published page drawing on the same stated truth. What earns its keep is not a governance programme. It is that the answer has one home, and that home is legible to software.
Find out which of your numbers has three answers
The Elephant Safari names your herd in ten questions and ranks them by what they cost you. Or start from the business problem instead: buyers asking an AI assistant, and it not knowing you exist. Get the Fuzzy out of the room. State the number once.
Every figure in this article traces to a primary source. See it in Knowledge