Data Quality
Critical fields complete, recent, standardised, and credible for the median customer. Aggregate numbers can look healthy while the middle 60% of accounts have thin coverage.
Median-customer population rate > 80%
Usable Volume
Volume is relative to approach: zero-shot, RAG, classifier, and recommender each need a different data profile. Measure history, cold-start duration, and bootstrap plan.
Cold-start mitigation plan exists
Access & Privacy
The team can explain what data flows where, under what terms. DPAs, residency constraints, enterprise procurement rules, and provider retention policies are confirmed.
Processing rights documented and confirmed
Evaluation & Monitoring
Task-specific evals are defined. Edge cases are captured. Failures are loggable and reviewable. Users have correction paths. Clicks and impressions are not substitutes for output quality.
Output quality definition agreed before launch