Teams often buy model access before they can answer basic questions: What is the source of truth? Who owns freshness? What happens when fields are missing? Weak foundations do not always crash demos, they quietly produce answers nobody should trust.
Data readiness for AI is practical. Inventory the systems that will feed retrieval or fine-tuning. Measure completeness and latency. Decide what must never leave your boundary. Design eval sets from real operator questions, not marketing prompts.
Only then does model choice matter. A strong model on stale, contradictory sources will still disappoint. A modest model on clean, permissioned data often wins in production.
If you are unsure where to begin, start with one workflow and the five fields that decide its outcome. Make those reliable. Then widen the aperture.