How do you prevent AI hallucinations or math errors when using AI to generate financial reports like P&L or NOI?

The approach described is to never ask the AI to perform math itself. Automatic transaction categorization uses Plaid's AI-based categorization layered with a real-estate-specific chart of accounts and mapping, plus a robust rules engine (assign categories, subcategories, and vendor assignments, with hierarchies) — none of which is AI, just data design. When a report like a P&L or NOI needs to sum categories, the system calls existing database functions to perform the calculations rather than having the AI compute them. This was described as a major improvement for avoiding hallucinations and wrong math, and the underlying calculations can be checked both on the back end and visibly on screen. The system is designed to function completely without AI; AI is only an add-on layered on top of the existing data structure.