If the first question of your AI roadmap is “which model should we pick?”, the second question is about to be expensive. Every enterprise AI program we inherit is bottlenecked on the same thing: a data platform that was not engineered to run models in production.
The three missing primitives
- Golden retrieval — governed, versioned corpora with lineage and access control, not notebooks dumped into a vector store.
- Evaluation as a product — harnesses owned by a team, with SLAs for time-to-evaluation that match your release cadence.
- Model risk as code — policy, sign-offs and audit trails that meet model risk committees halfway, not at the last mile.
When these three are in place, the models stop being the hard part. Teams shift from demos to outcomes, and model-risk committees move from gatekeepers to partners.
“The fastest AI teams we see are not the ones with the best models. They are the ones whose platforms made the models boring.”
A 90-day foundation
Start with the data you would defend in an audit. Build a semantic layer, define lineage, and stand up an evaluation harness that runs on every merge. Only then open the door to product teams.