An AI agent that asks before it queries your data

A user types "show me last quarter's churn by cohort" in plain English and gets a dashboard back. The language model never executes a query on its own, and the authors' own ceiling is just as clear: the results are small, DataBrain-specific, and do not establish production readiness, general text-to-SQL accuracy, or an efficiency advantage over manual dashboard construction.

That is the news in arXiv preprint 2610.06914, a DataBrain prototype called Text2Dashboard. The model proposes actions; deterministic software decides what runs. Every query passes read-only SQL enforcement, schema-constrained typed tools, and named Hooks for exact-call approval, audit logging, checkpointing, and recovery. A browser-level inspection pass checks the result before a human sees it.

The pipeline runs entity resolution, metadata discovery, read-only SQL, dashboard composition, static checks, dynamic preflight, and browser inspection, with a ReAct-style decision/action/observation loop and persistent state at each stage. On the authors' own evaluation: metadata selection 4 of 4, SQL semantic match 4 of 4, exact output-column contract 2 of 4 (six of eight strict). All four single-panel dashboard tasks passed, plus one two-panel task and one refinement; one parameterised task exceeded its step limit. Ten controlled fault scenarios met their specified outcomes without unapproved external side effects.

Model inference still accounts for over 97% of the observed runtime in every reported group. On the DataBrain About page, the company positions the work as part of its governed-data stack, and the prototype is one worked example of how a language model can sit in front of company data without holding the keys.