I took a close look at CFO Connect’s new report on The State of AI in Finance. The adoption story is strong, but the reality is more grounded—AI is still mostly used for basic tasks, while bigger challenges like fragmented data and output reliability remain. Surprisingly, the report doesn’t meaningfully address hallucinations at all.

This post explores how ChatGPT and a specialized accounting agent handle increasingly complex US GAAP questions—from basic revenue recognition to nuanced bill-and-hold scenarios—highlighting differences in accuracy, source interpretation, and the importance of "trust but verify" when using AI in financial analysis.