TL;DR

Data scientists at BIST, the government department covering business, innovation, science and trade, have shown that AI can convert plain English questions into queries for a graph database. The catch is that a query can run perfectly and still answer the wrong question. The team’s conclusion favours a hybrid, with a bank of approved queries for routine questions and AI generation kept for the unusual ones.

The bottleneck being tackled

Graph databases store how things connect, such as organisations, projects and suppliers, and are usually queried in specialist languages like Cypher. Analysts on the department’s Global Supply Chains Intelligence Programme have depended on data scientists to turn each business question into code, which slows everything down.

Larry Adeyemi, Emily Lambert and Joe Hearnshaw, the engineers behind it, built a lightweight tool that takes a question such as “Which organisations are connected to this project?” and writes the query itself. To test it without touching live data, they built a generator for synthetic graphs at scale, then scored the output with the DeepEval framework plus targeted human review.

What they learned

Getting syntactically valid queries turned out to be the easy part. The difficult part was making sure the query matched what the analyst actually meant, since a well-formed query built on a misreading returns confident but wrong results.

That finding generalises well beyond graph databases. Any team putting natural language interfaces over structured data, from finance systems to CRM platforms, faces the same failure mode, and it is invisible unless someone checks results against intent. The team’s preferred design reflects that: deterministic, pre-approved queries where reliability matters most, generative AI where flexibility is worth the risk.

The project also left the department with reusable in-house tools for synthetic data and evaluation that can support other AI work.

The Whitehall context

Science now sits in BIST’s title alongside business and trade; Andy Burnham abolished DSIT on taking office this summer. Like MHCLG’s shared AI gateway, this is a small, internally built project rather than a procurement, and it is being written up openly.

Looking forward

Next steps include scaling the tool towards possible production use, making the synthetic test data more realistic, explaining results visually to non-technical users, and trialling agent-driven testing with human oversight. For UK businesses considering AI over their own data, the honest lesson from this work is to budget for checking intent, not just for generating answers.