Multi-Agent Vision-Language Pipeline for Template-Agnostic Invoice Extraction and Validation

Traditional accounts payable systems remain dominated by template-based extraction tools that fail on new vendor layouts, by manual data entry whose error rate sits in the five-to-ten-per-cent range, and by approval cycles that routinely span twelve to fifteen days. We present a multi-agent invoice processing pipeline that addresses these limitations through three coordinated mechanisms. First, all agents -- document preprocessing, invoice extraction, validation, general-ledger and tax-code mapping, and a multi-agent supervisor -- are implemented as explicit LangGraph state machines with named nodes, conditional routing, and observable state. Second, invoice extraction is performed in a template-agnostic manner using multimodal Vision-Language Models operating under a strict, schema-bound prompt over more than one hundred structured fields. Third, validation combines structured business rules with sentence-embedding-based semantic matching to perform automated two-, three-, and N-way reconciliation against purchase orders and goods receipt notes. We describe the architecture, position it against the relevant literature in document understanding and agentic workflows, identify the design targets that follow from established industry baselines for manual accounts payable, and discuss the empirical evaluation that remains to be performed against a public benchmark.