It nearly always starts the same way. Someone presents a number, someone else has a different number, and the rest of the hour goes on working out which one is right. Two weeks later the same thing happens on another topic.
The usual conclusion, our data quality is poor, is sometimes right: missing, duplicate or wrongly joined records do exist and have to be ruled out first. But often both numbers are correct and answer a different question. One counts customers per contract, the other per organisation. One excludes cancelled orders, the other does not. Nobody is wrong then, and that is exactly why measuring better does not fix it.
The real problem is that the definition is written down nowhere and the origin cannot be traced. As long as nobody can show how a number came about, every discussion stays a matter of who defends their version most forcefully.
Why it persists
This is an organisational defect, not a technical one. Every system is well set up on its own, every department has good reasons for its own definition, and nobody owns the whole. Anyone who tries to solve it by putting everything in one data warehouse discovers that the conflicting definitions simply move along with it.
It becomes more visible as soon as AI enters the picture. A model trained on loosely defined data produces outcomes whose meaning stays disputable: validation says little when it is unclear what was measured. Then the project stops, not because the model is poor, but because nobody is willing to sign off on it.
How this plays out in government and non-profit is on the sector page.
