Same AI model, same 40 business questions — once with the warehouse schema and SQL, once through Bellaso’s MCP — graded against answers our team verified by hand.
92.5% vs 22.5% of questions answered correctly — about 4x the raw-data score: 37 of 40 vs 9 of 40, majority of 3 runs each. Measured November 2026.
Results
Each figure is the lower of the two raw baselines, rounded down.
10 of 10 — questions never used to tune the setup, answered correctly (raw data: 6 of 10).
nearly 4x fewer — tokens per answer (about 75% fewer).
about 3x lower — model cost per answer (list-price estimate, model tokens only).
about 3x fewer — tool calls per answer.
about 2.7x faster — answers, end to end (measured at matched benchmark load).
about 4x less — warehouse data billed per answer.
How we tested
Same AI model on both sides.
40 real business questions from one client’s marketing and sales data, graded against answers our team verified by hand.
Each question asked 3 times, scored by majority.
Raw data: the model gets the warehouse schema and can write SQL.
Bellaso: the model connects to Bellaso’s MCP server — business definitions, the semantic model, and curated data marts.
10 questions were held out — frozen before tuning and never used to adjust the setup.
Caveats
One client pilot, 40 questions, one model. We are running the same benchmark for more clients — this is the first published result, not a universal claim.