Database inventor says AI that answers business questions still fails

Michael Stonebraker, who introduced the Postgres database at Berkeley in 1986, says AI that answers business questions in plain English is still not ready. His team tried it on MIT's own data warehouse and got zero percent accuracy at first. The account comes from partner content, paid for by database company EnterpriseDB (EDB), in its AI & Data Horizons series.

What the MIT test found

The idea is called text-to-SQL. A manager types a question, and the AI turns it into a database query. Public benchmarks called Spider and BIRD report accuracy above 80 percent. MIT's warehouse holds 1,400 tables on students, faculty, buildings and salaries. Using real past queries, Stonebraker's team scored zero. Extra AI techniques lifted that into the teens. Describing every table and join term reached the 30s. He called that "still not ready for prime time."

Why real company data trips AI

• The data is not on the public internet, so the AI has no training examples.
• It is odd. MIT buildings have numbers, not names.
• Table and column names have become cryptic over the years.
• Real questions are harder than benchmark questions.

Stonebraker's team published its own benchmark, called Beaver. He says nobody has moved past accuracy in the 30-percent range on it.

A different approach

Stonebraker says language models handle text well but are poor at combining structured data, which is most of what companies hold. In work with Munich's transport department, one citizen complaint about crosswalk timing meant pulling together regulations, CAD drawings, trolley schedules and school-zone maps. Using a model to run all that was "ridiculously expensive," he said. His team's system, Rubicon, uses AI only on text, turns the results into tables, and lets a standard table engine do the rest. He says it is 10 times cheaper and more accurate than agent tools built around a model.

Why he distrusts the hype

He expects agentic AI, where software takes actions on its own, to fall into a "valley of despair" before views balance out. Businesses dislike unpredictable answers. Approve a loan, then reverse it on review, and "that’s called a lawsuit." He notes databases already have access controls, complete logs and triggers that can stop a transaction. He calls AGI, human-level AI, "basically marketing hype."

Why Postgres, and what PostgreSQL 19 shows

Stonebraker praises Postgres because no single company controls it. Bruce Momjian, a cofounder of the PostgreSQL Global Development Group who works at EDB, says the project has no office, assets or legal entity. He says it favors flexibility over small speed gains. PostgreSQL 19 pulled several previewed features, including SQL property graph queries and partition merging and splitting, because reliability comes first. The August 2026 patch release for PostgreSQL 18 fixed 28 security flaws, which the article ties to AI-assisted review.

The text provided cuts off where the two experts disagree on how fast AI will write database queries, so their positions on that are not covered here.