The meeting has a familiar shape. Two teams present figures for the same metric and the figures differ. The discussion that follows is about data quality, source systems and pipeline reliability, and it concludes with an agreement to invest in better infrastructure so this does not happen again.
It happens again. It happens again because the numbers were never the disagreement.
The dispute in almost every case of this kind is definitional. What counts as an active customer, and after how long does one stop counting. Whether revenue is recognized at contract or at delivery. Whether a churned account that returns is retained or reacquired. These are not data questions. They are decisions about how the business describes itself, and they have consequences for whose numbers look good, which is why they never get made.
Infrastructure as a way of not deciding
Buying a warehouse, a semantic layer or a governance platform is genuinely useful and does nothing about this. A definition layer encodes definitions; it does not produce agreement about what they should be. Organizations that install one without doing the definitional work simply relocate the argument into a configuration file, where it is now also invisible.
The reason the negotiation gets avoided is that it has losers. A definition of qualified lead that survives scrutiny will make a marketing team's performance look different than the definition currently in use. A retention definition that counts accurately will move a number that someone is compensated on. The technical framing is attractive precisely because it lets everyone treat a political problem as an engineering backlog.
This is the same avoidance visible in meetings convened to align on something that a single decision would settle. The activity substitutes for the resolution, and it is more comfortable because it never requires anyone to say that a colleague's number was wrong.
The consequences compound as more gets built on top. Every system consuming a metric inherits its ambiguity, and organizations deploying AI systems on business processes are discovering this acutely, since a model cannot be evaluated against a target nobody has defined. The evaluation gap in enterprise AI is substantially a definition gap wearing newer clothes: you cannot measure whether the system is right without first agreeing what right means, and that agreement was missing long before the model arrived.
The fix is unglamorous and cheap and almost nobody does it. Write down the twenty metrics the company actually runs on. For each, name a single owner with authority to define it, publish the definition in language a new employee could apply, and require that any dashboard using the term use that definition or explicitly declare a variant. This is a week of difficult meetings and it costs nothing.
The same failure explains why so many organisations cannot say whether a programme worked. A mental health benefit evaluated against undefined engagement, or a hiring initiative measured by a quality-of-hire metric nobody wrote down, produces a debate rather than a finding, every time, for the same reason.
It does not get done because it is not a project, it does not have a vendor, and its output is a document rather than a system. Companies are far better at buying tools than at making decisions, and definitional clarity is entirely the second kind of work.
