The problem. Executives ask "why did West margin fall?" and get a dashboard, or an AI answer with numbers nobody can check. On governed data, an answer is only useful if every figure can be traced to the query and rows it came from, and if the agent never reads data the asker isn't cleared for.
The idea. A team of agents answers a plain-English business question with a short brief: the cause, the evidence and 2-3 actions.
- A Planner splits the question into 2-4 sub-questions.
- A Data scout per sub-question picks tables from the Master Catalog, seeing only what the caller's clearance allows.
- A SQL analyst writes one read-only SELECT, runs it, and retries with the error.
- A Context agent pulls the policy and pricing passages that apply (RAG over a knowledge base).
- A Recommender writes the brief, citing evidence ids for every claim.
- A Critic, deterministic code rather than a model, removes any number that doesn't match the cited rows.
- Discover mode: the swarm scans the data and writes its own questions, such as "Why did West freight per unit rise 60%?".
Why SQL and RAG together. SQL gives the numbers: margin, discount and freight are computed from the governed tables, never guessed. RAG gives the reasons: the policy documents explain why those numbers moved (the West Q3 volume push, the new freight contract). The model ties the two into one brief. Without SQL you get a nice explanation with no proof; without RAG you get accurate numbers with no reason.
Governance is built in. Read-only, enforced at the engine: DELETE, DROP, ATTACH and PRAGMA are refused. Column sensitivity comes from the catalog, so a restricted column such as contact_email is withheld instead of leaked. Every number in the brief links to its SQL and rows.
Result so far. Asked "Why did West margin fall since July 2025?", the prototype answered in 11.4 s: gross margin fell 13.3 pts (38.9% to 25.6%), driven by discount +13.3 pts and freight per unit +60.1%. 10 claims verified, 0 rejected, every figure linked to its query.
Oracle AI Data Platform pieces. AIDP Agents (agent flow with a SQL tool and a RAG tool), Master Catalog (tables, column descriptions, sensitivity), Knowledge Bases (company policy documents), Spark compute, and Autonomous AI Lakehouse. The prototype runs on local NVIDIA GPUs today; the model layer is built to swap in OCI Generative AI or NVIDIA NIM.
