Three Ways Investors Are Using AI Now
Snowflake’s Jonathan Regenstein on how banks, hedge funds and asset managers are putting LLMs to work.
Two summers ago, when I started AI Street, I remember getting some confused looks when I told some old sources what I was covering. One said: “So, you’re writing about ChatGPT and Wall Street?”
I thought the technology had a real shot at becoming widespread across Wall Street. Or, if not, it was going to be a great big dud worth covering.
Early on, I spoke to Snowflake’s Jonathan Regenstein, who was also optimistic that AI would have a big impact on how the investing world works. I interviewed him in the fall of 2024, and much of that conversation focused on a basic question: how do you even use this technology for investing? Does it make sense to? What about hallucinations?
When we spoke again recently, Regenstein said the skepticism has largely disappeared, at least among the most sophisticated funds and asset managers. The question now is how to build efficient processes.
Regenstein is a former attorney and self-taught coder who now leads wealth and asset management at Snowflake, where he works with banks, hedge funds and asset managers building AI systems on top of their own data. Snowflake does not build frontier models. It works with providers and platforms including OpenAI, Anthropic, Mistral, Meta and Hugging Face, which gives Regenstein a view into how firms are putting different models into production and where they still struggle.
He says investors are using AI in three main ways:
Chat: using the model conversationally to test a hypothesis or get a first read on what’s happening.
Text analytics: “Go examine this 10-K for me” — or, at asset-manager scale, turning years of filings, research papers, and sell-side research into an “engine of insights.”
Coding assistant: “Write the code for me” — helping quants and investors code up hypotheses, reproduce papers, build dashboards, or call validated models.
Regenstein has also published a new book, Large Language Models: The Hard Parts: Open Source AI Solutions for Common Pitfalls, coauthored with Thársis Souza PhD, whom I’ve previously interviewed on my podcast here. The book is for non-experts who want to understand LLMs well enough to use them: builders, go-to-market salespeople and strategy leaders. I’m looking forward to getting my copy in the mail!
This interview has been edited for clarity and length.
Matt: Tell me about your role.
Jonathan: I work at Snowflake, which is a super interesting place to work right now because we don’t build frontier models. We partner with OpenAI, Anthropic, Mistral, Meta and Hugging Face. We partner with all of them.
As this world evolves, we’re sitting at the center of it, even though we’re not really as high up in all the headlines. We see it all, deal with it all, work with them all and optimize for all of them. I get to see some trends emerging around who uses which tool for what.
There are three different ways that LLMs are being deployed now. One is the chatbot. People like the chatbot. The second is text analytics: “Go examine this 10-K for me.” The third is a coding assistant: “Write the code for me.”
Those are three very different things. The models can do all of them, but they’re not the same thing at all. They’re not for the same person or the same use case. They’re really totally different. It’s starting to break down along those lines, and there’s a lot underneath each one.
How has AI adoption in finance evolved since we first spoke in the fall of 2024?
Jonathan: There have been a few big shifts. Back in 2024, I think you and I were both believers in this. But at that time, with some exceptions, if you had gone into a room full of quants or portfolio managers, I think you would have gotten a lot of crossed arms: “I’ve been doing this a long time. I understand, and I’m sure AI is helpful, but it’s not going to revolutionize how I do my job.”
We’ve now come to a place where there is no more skepticism, certainly at the most sophisticated shops and asset managers. The question is: “How can we optimize this?”
It’s becoming a recruiting and retention tool for funds and asset managers: “You should come work here as a quant. Why? The AI harness we’re going to give you is better than you’re going to get anywhere else. If you don’t have that, you cannot win in this world.”
Once you’ve come to a fund or asset manager and they’ve built you that harness, you don’t want to leave because you can’t take it with you. It’s intellectual property. It’s becoming a recruiting and retention tool. It’s a complete shift.
How are firms building governance into those AI harnesses and testing whether they actually work?
Jonathan: This is an area where I work directly. Snowflake doesn’t build its own models, so what are we doing? We’re trying to give you the governance architecture that sits underneath them.
When you ask for data, we need to make sure that what you asked for is what you actually got and that it is underpinning your answers. That’s a ton of the work we’re doing right now: validating your context and making it scalable to monitor and evaluate what you’re doing.
Skills are becoming a huge thing. What skills do is bring determinism, as much as possible, to the way large language models work because you’re giving them a recipe: “Do this. Follow these steps. Bang, bang, bang, bang, bang.”
One of those recipes can be a quantitative model that I’ve built somewhere. Large language models don’t run a stress test or even run a regression for you. They could write the code to run the regression, but I don’t want the model to write the code on the fly. I want it to call the model I’ve already built and validated.
Let’s say I have a phenomenal modeling team and they’ve built these models. We’ve validated them and have them in our ecosystem. This is something you can do with Snowflake: You can store all your models in one place.
Now the LLM needs to choose the right model, not run the model. Where we have a whole model-testing framework that we’ve validated over the last 20 years, the LLM can call the right model. That’s what I need to get right more than I need to get an LLM to build a model on the fly for me.
You can do that for research. Part of the research for your model garden can be having an LLM build a model for you. But when it’s actually time to make a decision, the LLM needs to call a validated, back-tested model.
