The unglamorous prerequisite for enterprise AI: clean data
Every AI project that fails in pilot fails for the same reason. The model was fine. The data underneath it was not. We share the data readiness framework we use with clients before a single GPU spins up, and the four questions that decide whether an initiative is ready to leave the lab.
The pilot graveyard is full of clean models
Walk into any enterprise that ran an ambitious AI program in 2024 or 2025 and ask what the model did. You will get a confident answer. Then ask what the inputs looked like and the room gets quiet.
The pattern is consistent. The model was the easy part. The data was a mess of half-owned tables, undocumented joins, and definitions that drifted across teams. The pilot worked. Production did not.
Four questions before a single GPU spins up
Who owns this data? Not the platform team. The business owner whose name goes on the slide when something is wrong.
What does each field actually mean? Write it down. If two teams disagree on the definition of an active customer, your model will pick a side, and you will not like which one.
Where did it come from and where can it go? Lineage is not a compliance checkbox. It is the only way to debug a bad prediction six months later.
What is the freshness contract? A model fed yesterday's data and today's questions will be wrong in confident, expensive ways.
The framework we run with clients
We score each data domain on four axes: ownership, definition, lineage, and freshness. Any domain that scores red on more than one axis is not ready to feed a production model, and we say so even when nobody wants to hear it.
The remediation work is unglamorous. Dictionaries, contracts, retention rules, owners. It is the part that does not make the press release. It is also the part that makes the press release a year later look earned.
The shortcut that is not a shortcut
There is a temptation to skip the data work by reaching for a retrieval-augmented setup. RAG is a useful pattern, but it does not absolve you of definition discipline. A retrieval system pointed at messy sources will quote that mess back to you with footnotes.
Do the work. The teams that do are the ones whose AI investments compound. The teams that do not are the ones writing this same article next year.
