AI Just Got Better at Answering Business Questions From Messy Data

If you've ever opened Power BI or Tableau and spent an afternoon hunting for the right spreadsheet, matching customer IDs across tabs, and cleaning columns before you could finally ask "which region sold best last quarter?" — this paper is about automating exactly that misery. Researchers collected real dashboards from public sources and pulled out real questions with known correct answers, creating the first large test of whether AI can handle the whole business-intelligence job from start to finish.

The honest finding: it can't, yet. Even the most advanced AI models answered fewer than half the questions correctly when left to figure out the data prep on their own. That matters because the prep is the slow part. As one researcher put it, the hard work isn't the question — it's finding the tables, linking them, and reshaping them first.

So the team built BI-Agent, an AI helper that breaks the job into smaller tasks: search for relevant tables, join them together, transform the data, then answer. Think of it as an assistant who reads the file room before answering your question, rather than guessing. That structure alone improved accuracy by up to 40 percentage points over a plain AI model. They then trained it further on thousands of real examples, adding another 30 points.

THE CATCH: This is a research paper, not a feature in the tools your company already uses. The agent still gets plenty of questions wrong, and results come from a curated test set — real corporate data is messier, with inconsistent naming, missing values, and access restrictions. Still, the direction is clear: the tedious 80% of reporting work is a target, and it's getting hit.