Inside the Bloomberg Terminal’s AI

If a piece of technology wants to stay relevant for decades, it needs to constantly evolve. The 43-year-old Bloomberg Terminal, which feeds a constant stream of real-time financial data to Wall Street traders and others, is a prime example of this—especially as Bloomberg integrates AI into its feature set.

AI isn’t that new to the Bloomberg Terminal, which released a machine learning-based market sentiment model in 2009. Over the past two years, the company has begun to even invest more in generative AI and large language models for production. Amanda Stent, the head of AI strategy and research for Bloomberg, described AI as a way to help financial professionals who are overloaded with information.

“The volume of information that users have to stay on top of is really overwhelming, and so we need to give tools to help them adapt to that reality, to allow them to think and respond faster,” said Suzanne Szur, the research product manager at Bloomberg. “It’s really about transforming hours or days of manual information gathering and synthesis into minutes.”

Szur was a part of the team that rolled out Bloomberg’s first AI solution, an earnings transcript summarization: Someone could open an earnings transcript on their terminal and receive a summary. From there, Bloomberg moved onto other features such as AI-generated answers to research questions, which allows the company to leverage its significant subject-matter expertise.

The customer is always right. To figure out the right choices for AI implementation, Bloomberg places a significant amount of weight on client feedback, including challenges and pain points.

Stent pointed out that the company adheres to four pillars of responsible AI: protection, transparency, reproducibility, and robustness.

What does that actually mean? In terms of “protection,” Bloomberg claims it doesn’t sell client or contributor data. On the “transparency” front, it offers up links to original sources when its AI tool provides a generated response to a research question.

When it comes to reproducibility, Stent said, “We have a shared technology stack and we develop evaluation sets, and we’re continually evaluating and testing so that we can reproduce the outcome even as the technology changes so that we’re providing our clients with a consistently good experience.”

The last pillar, robustness, addresses financial professionals’ fears that they won’t be able to access something on the terminal in time to make decisions. None of Bloomberg’s AI tools rely on a single instance. As Stent put it: “We are here to be up all the time.”

I’m not hallucinating, right? Hallucinations remain a concern for companies using AI tools. Bloomberg counters that risk through entailment and factuality checks, as well as transparent attribution. There’s also a human in the loop for Bloomberg’s AI solutions, with the exception of processes such as intraday bond valuation (IBVAL), which prices bonds every 15 seconds. “Too fast for humans,” Stent said.

There’s also regular, ongoing testing to ensure that the model isn’t drifting into inaccurate outputs. The best way to test for accuracy for something like IBVAL, Stent added, is to wait until a process is complete and then perform a check:

“You wait till the future becomes the past, and then you use that to evaluate and so we have real ground truth.”

When we look to the future…Stent described Bloomberg’s use of AI as “an evolution, not a revolution” over the past decade.

“With every generation of AI, you have the potential to bake the AI into the existing way that people work, or think about how AI can validate and support new ways in which people can work,” Stent said. “If we think about our clients over time, their work also changes as technology changes.”

“There’s a real recognition from everyone that we have an opportunity to really reimagine the terminal in the age of AI, and so that is really the path that we are on,” Szur said.