Retrieval-augmented generation for finance, tax and audit
Retrieval-augmented generation appears on 165 recent papers, 171 of 178 such papers in AI journals, and only 12 touch finance, legal, tax, accounting, audit or compliance questions. Banks, tax advisers and auditors are deploying it over filings, rules and contracts. How often does it cite the right source, and what happens when the source changes?
What prompted this
Since October 2025, 178 papers carry a retrieval-augmented keyword (171 in AI journals, 5 in management): "retrieval-augmented generation" on 165, "graph-based retrieval-augmented generation" on 8. Only 12 of the 178 also carry a finance, bank, legal, regulation, accounting, tax, audit, compliance, economics or business keyword, and those few include legal question answering and tax return filing.
Questions
- On questions about accounting standards, tax rules and loan agreements, how often does a retrieval system answer from the right passage, from a superseded version, or from nothing?
- When rules change, how long do deployed systems keep answering from the old text, and who notices?
- Do professionals check the retrieved sources, and does showing them reduce errors or only slow the work?
Data that could help
Question sets with gold answers and sources built from accounting standards, tax codes and standard contracts, including versions over time.
Next step
Build a test set of 300 questions over the current and prior versions of one accounting standard and measure how often retrieval systems cite the wrong version.