Governed Agentic Analytics: A Systems Architecture for Trustworthy Natural-Language Access to Enterprise Data

Abstract

Governed Agentic Analytics (GAA) is a systems architecture for trustworthy natural-language access to enterprise data. It separates probabilistic model reasoning from deterministic authorization, validation, and least-privilege execution. The architecture combines identity-aware access, retrieval-grounded context, bounded agent planning, policy-enforced tool execution, positive data allow-lists, provenance capture, observability, evaluation gates, and human oversight according to action risk. This preprint describes the GAA architecture and a production-oriented reference implementation over a heterogeneous, multi-region enterprise data environment. Model-generated actions are treated as untrusted proposals; deterministic policy and infrastructure independently determine whether an action is authorized, safe, within scope, and executable. Evidence boundary: This manuscript is the architecture and reference-implementation source. Corrected comparative functional results are archived separately in the Governed Agentic Analytics Experimental Artifact, DOI 10.5281/zenodo.22105797. Quantitative claims derived from that benchmark should cite the experimental artifact rather than this manuscript. The manuscript reports architecture and systems-validation properties without organization-specific identifiers, proprietary datasets, customer information, credentials, internal URLs, or sensitive business logic.