What is AI’s strategic play in finance?

Bloomberg Professional Services

What lies ahead for generative AI? How can finance leaders balance speed and safety in enterprise adoption of LLMs? 

At Bloomberg’s 2025 Sell-Side Leaders Forum in New York, Mandeep Singh, Global Head of Technology Research at Bloomberg Intelligence spoke with Shawn Edwards, Chief Technology Officer at Bloomberg, Ather Williams III, Senior Executive Vice President, Head of Strategy, Digital, and Innovation at Wells Fargo, and Ali Hirsa Professor of Professional Practice in the Department of Industrial Engineering and Operations Research, Columbia University about the opportunities and considerations surrounding AI integration, and how to scale innovation across complex ecosystems. 

In focus 

Featured insights from the discussion panel: 

On LLMs as orchestration layers 

Shawn Edwards: LLMs are not a replacement for your data warehouse. That’s not how anybody should think about it. You can teach it, you can train it all you want, but you need your trusted data sources to get the facts out of it. You have to be very careful when you ask an LLM to do something. And it’s also not a replacement for all your calculators and your services. You don’t ask LLMs to do math. They are actually really bad at it. They’re not going to do bond calculations. It’s really about having built and continuing to curate your data sources. It’s continuing to build that up and expanding it. It’s about teaching your AI systems, how to use these tools, how to reach out to a data lake, how to reach out to a calculator, how to use these things and synthesize answers. And in some sense, the AI system becomes a coordinator.*

On access to quality data

Ali Hirsa: When it comes to the pillars of AI, no matter what we’re talking about, the very, very first one has to do with data. It’s quality. And definitely, in most cases, we may not have enough—especially when it comes to relevance, tail risk, or rare events. We need to bring in deep learning techniques for generating synthetic data. These are non-parametric methods. But that opens a whole other field: what’s the best way to create synthetic data that still captures the stylized facts of real-world data? That’s where a lot of current research is headed, and it’s where we’ve been focused.

On the value of synthetic data

Ather Williams III: The generation of really good synthetic data…it’s been a game changer. Because as we start going through fundamental platform rebuilds, you can use a Copilot or a GitHub Copilot to do the code. But we also need to generate test scripts, and then we need to run those tests. But the cycle time for us doing regression testing…has gone from six – eight weeks to days. Because we can generate synthetic data, we can automate test cases, and we can run it through. We can get through a lot more cycles… which has been absolutely game-changing in terms of the speed with which we can bring new platforms to market.

*Quotations have been edited for brevity and clarity.

Interested in learning about Bloomberg solutions for sell-side leaders? Click here.