Why AI Data Analysis Stalls at 80% Accuracy

Julie Zhuo, co-founder of Sundial, ran a 50-question accuracy test on AI-powered data analysis and hit 80% out of the box. After feeding in canonical metric definitions and business context, that number jumped to 98%. The gap wasn't model intelligence - it was missing institutional knowledge like which metric definition is official when three exist, whether a revenue dip was a bug or a pricing change, and how analysts actually work backward from business decisions. The practical takeaway lands well: golden metric sets, machine-readable change logs, documented investigation playbooks. Teams at Airbnb, Stripe, and similar data-heavy orgs have been circling these same problems for years. The bottleneck for trusted AI analytics is organizational documentation, not model capability.