FactSet has added conversational AI to the transcripts of conference calls on its financial analysis platform. The feature appeared unannounced late last week.

The application leapfrogs Bloomberg by offering analysts the ability to query transcripts with custom questions. Bloomberg announced a similar product five days earlier.

The dual announcements in quick succession suggest the battle to provide ChatGPT-style technology to Wall Street is heating up.

Bloomberg and FactSet took more than a year to adopt generative AI, an indication they are prioritizing accuracy over speed. As market leaders, it is a luxury they can afford.

Bloomberg fired the first shot seven days ago with a function that allows clients to generate summaries of selected transcripts for recent conference calls.

The Bloomberg application uses pre-set prompts, so all the summaries have the same format. It does not include what’s known as conversational AI, or the ability to freestyle questions.

Factset upped the ante by including, along with predefined queries, the ability to input hand-crafted questions.

FactSet’s document search has preset questions such as “Summarize the high and low points of the Management Discussion Section” and “List all the guidance points presented in the call.”

The screen indicates FactSet is using OpenAI’s GPT4 large language model.

The product works well enough, though it runs slow because it waits to generate the entire answer before displaying results. The conversational AI functions that are publicly available in Microsoft’s Bing and Google’s Bard show each line as it is unspooled.

Like Bloomberg, the FactSet application is limited to transcripts, a useful but fairly narrow corner of the data universe that both companies collect.

Neither player has turned ChatGPT loose on their vast array of data. In part that’s because it would make errors more likely. It is also expensive to run ChatGPT-style queries.

(Note: FactSet released a Beta version of an AI product for junior bankers in December, but the press release is such a word salad it’s hard to say what it actually does.)

It’s unclear whether Bloomberg or FactSet plan to charge more to use the features.

Bloomberg’s press release also wasn’t clear if it’s using its own model or one from OpenAI.

The Bloomberg approach limits costs because the prompts are hard coded and the results can be cached. So, once someone runs the prompt for Tesla, the same answer can be delivered for other clients.

An unresolved issue for all financial companies using generative AI is privacy.

FactSet’s application includes a disclaimer with this language: “Your queries and associated response will be logged to improve FactSet’s products and services and your user experience in accordance with industry best practices.”

In plain English, that means FactSet will log all the queries and use them to train models.

Wall Street firms tend to have a high level of anxiety about how their data is collected and used.

They will probably have questions.