From a tax ID to a finished report
Starting point
The data is public, but scattered across many sources.
Registries, the VAT whitelist, beneficial owners, related entities, financial statements as PDFs from the registry. Every source had to be opened separately. A few hours of analyst work for one report, and the quality depended on who wrote it.
What we built
Three bots and one model.
The first bot checks the tax ID and queries every source at once, then assembles the answers into one whole — with the option to re-run at any time to check whether the data has changed. The second reads financial statements, calculates profitability, liquidity and debt ratios, and draws charts into a PDF.
The third hands the collected data and documents to an LLM. The model writes the financial health assessment, a description of the ownership structure, a list of risks, and conclusions. The report is generated in the background, and the user gets the complete result.
In the end
The model gets the data, it doesn't look for it.
The model has a narrow task and works only on documents for the given company. It doesn't search for data itself — the bots hand it over. That means the report can be shown to a client without checking every sentence, guarding against AI hallucinations.
Results
A report that used to take half a day.
We don't name the company. Numbers come from the audit and process measurement at the client.