A services group asked us to help them “do something with AI” on top of their operations data. The ambition was reasonable; the foundation was not there. The same customer existed under three spellings across three systems, key fields were filled in differently by different teams, and two reports of the same number rarely agreed.
Every analysis started with an argument about whose figures were right, and ended before it got anywhere useful. No model built on that would have meant anything — it would have learned the mess.
The challenges we had to solve
- The same entities appeared inconsistently across systems, so nothing could be joined up reliably.
- There was no agreed definition for basic measures, so two true reports could still disagree.
- The client expected a visible result, and data plumbing is not visible — we had to show progress people could feel.
- Whatever we fixed had to stay fixed, not decay the moment we left.
How we approached it
We were honest that the real project, for now, was the groundwork — and that it was worth doing whether or not a model ever followed.
We worked with the teams to agree single definitions for the entities and measures that mattered, reconciled the records so one customer was one customer, and built the connections and checks that keep data consistent as it comes in rather than cleaning it up after.
We treated data readiness as the precondition it is, not a chore to rush past on the way to something shinier.
Because plumbing is invisible, we tied the work to something the client could see: a small set of plain reports they had never been able to trust before, now reconciled and agreed across the business.
That gave them an immediate return and gave us a way to measure the groundwork against something real — whether the numbers two teams produced finally matched. People still own and interpret those numbers; we made them trustworthy, not automatic.
Where it stands
Meetings no longer open with an argument about whose figures are correct, because there is one set everyone accepts. The reporting alone repaid the work, before any model entered the picture.
Whether the group goes on to build something more involved is now a real choice rather than a fantasy resting on data that could not bear the weight — and if they do, the foundation is finally there to hold it.