Abstract
Enterprise knowledge is commonly fragmented across operational systems, process models, and documents, limiting its use for traceable automated reasoning. This paper introduces the Business Graph, a formal model that represents an organization as a typed, temporal, and evidence-aware graph. Its meta-ontology and relation algebra establish consistent semantics for querying and inference, while temporal, epistemic, and provenance information makes results traceable to their sources and derivations. The proposed architecture separates language understanding and presentation from symbolic execution through explicit contracts, BGQL, and lineage-preserving reasoning. A Python reference implementation validates graph and contract semantics, executes queries over in-memory and persistent snapshots, and produces evidence-bound results with immutable execution identity. The concept and its implementation are examined through synthetic delayed-order and payment-service scenarios that exercise explanation, dependency traversal, missing-knowledge handling, and grounded result presentation. Automated tests verify the implemented semantics, while provider-neutral rendering and retrieval protocols are exercised with interchangeable external language and retrieval models rather than tied to a particular model family. The experiments show that contract validation prevents unsupported references and that adding local graph structure improves retrieval coverage, while also revealing that grounded generation may remain incomplete and that relevant retrieval does not by itself establish answer sufficiency. These results support the Business Graph as a computable foundation for auditable AI-assisted business reasoning, while larger independent evaluations remain future work.