Financial AI systems need to answer decision‑bearing questions about payments, collections, cash flow and compliance while guaranteeing numerical accuracy, traceability and auditability. Large language models (LLMs) enable natural‑language interaction but cannot alone ensure reproducible calculations or evidence‑based recommendations. We therefore propose a production‑grade architecture composed of three components:
- A financial knowledge graph that captures enterprise relationships, temporal context and provenance;
- A deterministic optimization engine that performs exact financial calculations and decisions;
- A multi‑agent LLM layer that interprets intent, orchestrates tasks, generates explanations and validates claims.
A coordination layer called context engineering selects, ranks, composes and validates query‑specific context across ERP systems, tools and tenants. Observations from roughly eleven months and fifteen deployments show that single‑agent pipelines exhibit hallucination rates of about 7%‑9% (5%‑10% across query types), often inventing invoices or unsupported financial statements. Graph‑backed multi‑agent validation reduced observed rates to below 1%, indicating a strong mitigation effect despite being observational. Data quality in ERP systems and tenant‑specific schemas remain major implementation challenges. Reusable mappings, graph templates and validation policies cut deployment time from roughly six weeks to two weeks.
We argue that financial context engineering goes beyond text‑centric retrieval‑augmented generation, requiring structured relationships, temporal validity, exact numerical evidence, access controls and validation against source records and solver outputs. Finally we outline research challenges in evaluation, reproducibility, privacy, context quality and cross‑tenant generalization.
Review: The proposed three‑engine design demonstrates a practical path to combine structured graph knowledge with deterministic solvers, offering concrete evidence for improving reliability in financial AI applications.