A Governed Approach Makes Enterprise Analytics Reproducible

Summary

This study examines a governed approach to enterprise analytics in which a language model interprets a user’s question, while deterministic policy selects and runs a pre-approved analytical program. The program returns both the analytical result and supporting evidence. Within a defined analytical class, the approach supports relational operations, aggregation, comparison, window functions, ranking, and similarity, while fixed rules for meaning, policy, data, and execution make results replayable. The evaluation covered 440 runs using three 8B models for runtime SQL generation and tool selection; Qwen3-8B was also used for intent interpretation in the policy-executed setup. None of 330 runtime-planning episodes satisfied the complete answer-and-evidence contract across all test datasets. By contrast, the policy-executed analyzer satisfied it in all 110 evaluated episodes. The authors present this as a configuration-specific result, not proof that runtime agents cannot succeed under other system designs.