How Schema Validation, Not AI Classifiers, Should Route Supplier Invoice Extraction

A healthtech billing system design uses a small language model for a first-pass extraction of supplier invoices, escalating to a larger model only when the output fails JSON Schema validation. The routing logic is built around invariant checks — such as line-item totals reconciling and supplier tax IDs matching known records — rather than an upfront AI-based difficulty classifier. Common failure modes include silently truncated line items, decimal format mismatches on European invoices, hallucinated SKUs, and duplicate writes from crashed workers. The duplicate-write problem is solved with idempotent upserts keyed on a document hash, while the remaining correctness issues are caught by the validator. Any extraction that fails validation twice is routed to a human review queue instead of being written to the ledger.