AI finance assistants answer a different question with data numbers

Researchers studied how AI models answer finance questions that do not give all facts. One example is what a figure measures. A second is which fund variant is meant. They tested Indian mutual funds, where variants and tax rules have formal definitions. The fault has the name frame collapse. The answer has a number from the data, but it answers a different question. The authors say they show how it occurs, not how frequently.

What they tested

Each study tests one challenge. With an unclear label, the models used the figure, frequently with the label of the user. With all labels given, they named all of them, but the layout set the sequence.

The results

When a fact about the investor was missing, weaker agents assumed a default or an expired tax rule. A rule that required the fact did not make them use it. An AI judge found mislabelled figures only when it knew what each field means.

What is not known

The paper gives no rate of fault. It proposes controls for finance assistants, but this text gives no results for them.

Insight: A check against the data cannot find a question that is different

In October 2026, ResolveLab graded agents from server records and not from the claims of the agent. A Japanese test found that six models fact-check primary claims less correctly than side claims. A check against the data accepts an incorrect answer, and the party that writes the check controls what is correct. A leader needs a test of the question, not only of the number.