New AI Framework Lets Banks Double-Check Its Answers

If an AI tells your bank how much profit your company made last year, would you trust it without proof? A new AI framework from researcher Sergiy Lunyakin makes AI answers auditable — meaning every fact comes with a clear source, just like a well-documented Excel spreadsheet.

This matters because AI in finance currently works like a fast but forgetful intern: it can give great answers, but there’s no easy way to check if the numbers are real or hallucinated. The new system, called KDAF, builds a knowledge map of finance terms and facts, then retrieves evidence using a method that tracks where each piece of data came from. For example, if it says revenue grew 8%, it can show which SEC filing, internal report, or analyst note it pulled that figure from.

In a test against 145 real finance questions, KDAF scored just as well as other AI retrieval systems on correctness. But it won decisively on auditability — getting the highest score for linking answers to real sources. That’s important because in finance, trust isn’t optional. Regulators, auditors, and boards demand proof. Without it, AI recommendations can’t be used in real decision-making.

The system also prevents AI from making up connections between unrelated companies or events — a common flaw in simpler AI systems. Every fact must trace back to a real document, and it can’t borrow evidence from outside the specific question. That level of discipline could significantly reduce errors in financial forecasting, risk analysis, and compliance reporting.