You do not need a lakehouse to start. You need three datasets cleaned, labeled, and exportable. Here is the 30-day audit we run before any AI engagement.
Week 1: Inventory
List every dataset relevant to the proposed AI project. Owner, system of record, refresh cadence, last documented schema change. The output is a one-page table. The act of producing it surfaces 80% of the data-readiness issues.
Week 2: Sample and label
Pull 200 rows from each candidate dataset. Have a domain expert label them against the task you intend to solve. This becomes the seed of your eval set. It also reveals whether the data actually contains the signal you need.
Week 3: Plumbing
Can you export each dataset to a flat file or an API your AI system can consume? In what latency? With what authorization? "We have the data" usually means "the data exists somewhere"; the audit asks whether it can move.
Week 4: Governance
Who owns each dataset? Is there a retention policy? Are there PII fields that need masking? Is consent in place for the use you intend? The answers determine whether you can ship.
What this audit catches
- Datasets that don't exist. Surprisingly common. "We have all our customer interactions" turns out to mean "we have last week's."
- Datasets you can't legally use for the intended purpose. Better to know now.
- Datasets too dirty to use without months of cleaning. Rescope or descope.
- The fact that the model is the easy part. Most AI projects are data projects in disguise.
No data audit, no project. We have learned this the hard way enough times to make it a precondition.