Enterprise AI systems can stumble on ordinary database queries even when public benchmarks look strong, Postgres inventor Michael Stonebraker argues in a sponsored interview published by The Register. His proposed answer is to use language models for text, then pass structured data to database tools built for the job.
Watch Desk analysis
What happened
Stonebraker describes testing text-to-SQL against MIT’s warehouse, which contains 1,400 tables. His team says its first attempt achieved zero per cent accuracy on real queries. Adding retrieval and prompt engineering moved results into the teens; supplying table and join details raised them into the 30s. He says the team’s Beaver benchmark has not yet been beaten beyond that range.
His group’s Rubicon system uses models to handle text, turns the results into tables, then leaves the remaining work to a tabular engine. Stonebraker says it was ten times cheaper and considerably more accurate than model-centred agent frameworks in the team’s testing. Those are his team’s reported results, not a general guarantee for enterprise workloads.
Why it matters
The argument is less “AI cannot work with company data” than “do not make a language model do every job”. Enterprise databases contain structured, often poorly documented information, and real queries can combine several data sources. Putting conventional database machinery in the loop may be more reliable and less costly than building the whole system around an agent.
That is a useful architectural distinction for anyone deciding where to put AI in a business system. A fluent answer is not much comfort if the joins underneath it are wrong.
Our read
Stonebraker’s case is a welcome reminder that the sensible AI system may involve less model magic and more dependable software doing the work it was designed for. The MIT result is striking, but it comes from one team’s tests, and enterprise databases vary enormously. The practical takeaway is to test against real queries and data, not assume a strong public benchmark will travel neatly into your warehouse.
What to watch
- Whether Beaver’s results are independently reproduced or extended to other enterprise datasets.
- How Rubicon performs on real deployments, including cost and accuracy.
- Whether AI tools make it easier to document the schemas and data that models need to use.
Discussion spark: For enterprise AI, should language models stay in charge of the workflow, or should databases and other conventional tools do most of the heavy lifting?
Sources and evidence
- PARTNER CONTENT: Built to Bend: Postgres’s Inventor and a Steward on What AI Really Needs – The Register (29 October 2026, 15:00 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.