On The Heels of Bloomberg, FactSet Launches Its Own AI Earnings Tool

A chatbot, search function, summaries, and integration with other FactSet data will accelerate in-depth research, the company says.


Illustration by II

In late 2022, after OpenAI’s ChatGPT captured the attention of finance professionals, many asset managers and other organizations began to develop, or accelerate their efforts to leverage, generative artificial intelligence. Now, more of those upgrades are coming to one of the most popular third-party investment data and research firms.

FactSet said Tuesday that its “Transcript Assistant,” an AI-powered chatbot that helps users search earnings transcripts and summarize calls, is now available to its clients, just weeks after Bloomberg announced that it added similar features to the terminal. More than 4,000 FactSet users are already actively employing the tool to help with their research, the company said.

The goal of these tools goes beyond sending a buy, hold or sell signal to investors. Like Bloomberg, FactSet and other platforms say the AI improvements will accelerate the in-depth research and analysis of company earnings, saving analysts time they can use to potentially uncover new information and make more informed decisions.

FactSet’s chatbot can analyze earnings calls in moments, then users can ask it bespoke questions or pre-populated ones. In response, the tool will summarize specific sections of an earnings report and comments from speakers, as well as highlight key data and surface related trends and themes. FactSet’s other data and research are woven into the summaries and cited.

“The solution fosters enhanced questioning and heightened confidence ahead of an earnings call. Ultimately our users are spending less time sifting through transcripts and more time on their strategic and analytical workflows,” said Emily Geer, senior director of the corporate business unit at FactSet.

Kristina Karnovsky, executive vice president and chief product officer at FactSet, said Transcript Assistant is a great example of a large language model already disrupting a part of research that was ripe for innovation. “There’s just such a great opportunity with the vast quantity of unstructured data and transcripts,” she said.


The tool was in development for about a year and human experts at FactSet have been helping to optimize it.

FactSet users have understandably been asking questions about the AI tool’s reliability and security. The chatbot is powered by ChatGPT 4, but the model only uses FactSet’s data and research and the company does not train it on other inquiries. That means the chatbot won’t consider the questions that other users have to inform whatever it generates.

Some asset managers and other clients were comfortable and eager to leverage an AI tool from FactSet. Others knew less about AI and required a lot of education about it.

“It was all over the map at the beginning of the process. People have ramped up very quickly. But I still think that there’s a spectrum of understanding and comfort across the different types of clients that we serve. What’s important to remember about the product is that we’re using a third-party large language model, but it’s a private instance of that. I think that was a confusion in the beginning,” Karnovsky said.

The chances of hallucination — when an AI model uses what it has gathered to fabricate answers to questions — is unlikely and greatly diminished compared to public generative AI tools because FactSet’s version is not training with its users’ inputs.

“We’re not doing unsupervised learning on the questions that people might be asking about the transcript. They can rate the answer by giving a thumbs up or a thumbs down and insert a comment like my question was unanswered. And our product managers can review that feedback, but those are people in the loop,” Karnovsky said.

FactSet has plans to incorporate AI into its other tools. In a public roadmap, the company said AI could put at the fingertips of users “mile-wide discoverability” and “mile-deep depth” of information.