A startup CFO used ChatGPT to build an FP&A tool

There’s been an ongoing debate over OpenAI’s ChatGPT, from its inaccuracies to some banks banning it to those who believe generative A.I. is part of the future of finance. One CFO set out to prove the value of the technology.  

“The amazing thing with these large language models like ChatGPT is they can take plain natural language queries and respond to them as a person would,” says Glenn Hopper, a startup CFO for more than 20 years at primarily private equity-backed companies, and a financial consultant. He’s also the author of Deep Finance.

I sat down with Hopper to discuss his project, “Application of ChatGPT to Build an FP&A Tool,” which resulted in 30 pages of documentation. “I consider myself a tech evangelist,” Hopper tells me. “I actually just registered the domain name Robo.CFO.AI.”

Hopper says he understands the vital role that financial planning and analysis professionals play in the CFO organization. 

“I guess I had two purposes with it,” Hopper explains. “One was to show finance people, ‘Hey, we’ve all been clamoring for tools like this.’” Tech like ChatGPT not only can respond in natural language, but it could also write code, he says. “I think most FP&A people probably can write SQL (structured query language) queries at this point, and they have data science skills,” Hopper says. “But they’re not developers.”

The second purpose of the project was to create a “proof of concept,” so they could show the developers of their company and say, “‘Look, here’s what we could do, and these are the advantages if we could do it,” Hopper says. 

What was the process? “I took three years of financial statements from a fictional company and prompted ChatGPT to write code that would import those financial statements into database tables,” Hopper explains. In analyzing the data, he wanted to reflect on the process of “when you do the close, and you start comparing your performance to budget and your performance to the prior month or the year before,” for example. So, he created key metrics. 

“When I realized I wouldn’t be able to get ChatGPT to query the data directly, I had it write code that created a chatbot that let users ask questions in natural language about the data,” he says.

“The bot will respond to your queries as if you were talking to an analyst. You can get information from the tables without having to write SQL queries. You can just…