Ask three departments what a customer is. Sales will count anyone who signed. Finance will count anyone who paid. Service will count anyone who called. Each answer is reasonable, each produces a different number, and every month somebody spends days reconciling figures that were never meant to agree.
That is not a data problem. The data is usually there and the systems usually work. What is missing is an agreed, owned answer to a prior question: what does this word mean here, and what decision is it for?
What the data quality studies do and do not show
Data quality is the field that has measured this most carefully, and its results are sobering. In a study of 75 organisational assessments, Nagle, Redman and Sammon had managers and their teams examine 100 recently created data records each and mark critical errors. On average 47% of records had at least one critical error, and only about 3% of the assessments scored at the 97% level that most people would call acceptable.1
Read that as a statement about accuracy, because that is what it measures. It leaves a question open that comes before accuracy: accurate against which definition? A record can be correct for sales and wrong for finance without containing a single typo. Cleaning it harder does not settle which one is right. Someone with authority has to say.
Meaning is use
Philosophers got to this point long ago. In the Philosophical Investigations, Wittgenstein warns against the picture in which every word simply labels an object, and proposes instead that for a large class of cases the meaning of a word is its use in the language (§43).2 The qualification matters and scholars still argue about how far the claim reaches. For an organisation the practical reading is modest and useful: look at what each team does with the word, and you have found what it means to them.
Sales uses "customer" to trigger a commission. Finance uses it to raise an invoice. Service uses it to open a ticket. These are three practices, and so they are three meanings. None is a mistake. The mistake is letting all three travel under one label into a board pack.
Why it matters more now
For twenty years the industry answered meaning problems with technology: another warehouse, another lake, another dashboard. Now it reaches for AI. Our view, and it is a view rather than a measured result, is that this makes the problem worse before it makes it better. A language model will answer a question about "customers" fluently, quickly and confidently from whichever definitions its sources happen to contain. Inconsistent definitions in, inconsistent but convincing answers out.
What settling meaning looks like
It is not a glossary project. Every term that a decision depends on gets three things, and we put them on one line:
- Defined. A sentence that someone outside the team could apply.
- Owned. A named person who can change it, and who is asked before it changes.
- Dated. When it was agreed, and when it will next be looked at.
Then, and only then, the technology goes underneath so that one canonical result can be shown to every chair at the table in a form they recognise. The argument at the meeting moves from "whose number is right" to "what should we do", which is the argument the meeting was supposed to have.
Coherence, in this sense, does not mean everyone agreeing about everything. It means everyone meaning the same thing when they use the same word.