An LLM predicted how individual CFOs would answer economic surveys.

Economists want to track business sentiment, which is how optimistic executives feel about the economy. Gathering this data is slow and expensive because surveys only reach a few hundred companies every so often. To fix this, researchers tested whether artificial intelligence could act like corporate leaders. They prompted a large language model to role-play as the chief financial officer of a specific firm at a specific point in time. The model then answered an economic-optimism survey question. It worked: the AI's predicted optimism closely matched the real executive's actual answers. Supplying extra background information about the executive's past choices and company made the predictions even more accurate. In the future, conditioned AI could provide fast, low-cost economic data.

Why it matters

For decades, measuring business sentiment meant surveying a few hundred CFOs quarterly. That data is slow, costly, and coarse. This paper shows that an LLM can replicate those responses, potentially giving economists and policymakers a way to run the same survey every week instead of every three months — at near-zero cost.

Questions answered