Snowflake Summit 2026 | John Heisler, Snowflake & Liam Hynes, S&P Global Market Intelligence

In this interview from Snowflake Summit 2026, John Heisler, head of AI and financial services at Snowflake, joins Liam Hynes, global head of new product development for public markets at S&P Global Market Intelligence, to talk with theCUBE's Dave Vellante and Rebecca Knight about how AI is transforming qualitative investment analysis and democratizing financial intelligence. Hynes traces the evolution from Michael Burry's manual approach — reading millions of words in SEC filings to predict the 2008 financial crisis — to S&P Global's AI-powered rebuild of the "Lazy Prices" framework, which uses Snowflake Cortex and CoWork tools to identify genuine year-over-year shifts in corporate risk disclosures and concentrate alpha in a short book. Heisler, drawing on a neuroscience background, explains how the brain's architecture of question, memory and action maps directly onto how modern agentic systems should be designed. The conversation also explores the distinction between democratizing data — which the internet enabled — and democratizing solutions, which AI now makes possible. Hynes details how a workflow that once required months of manual reading can now deliver an equity analyst a ranked view of S&P 500 companies with new incremental risks in minutes, collapsing the gap between academic research and real-world value. The discussion shifts to determinism: Hynes explains how Snowflake CoWork's stored-procedure functions ensure agents return consistent, repeatable outputs rather than probabilistic guesses — a critical requirement for production-grade financial applications. Heisler argues that governing data is governing AI, pointing to role-based access control and semantic commonality as the foundation for enterprise-wide intelligence. From building a clear lineage from business strategy to data strategy to AI strategy, to knowing precisely when not to deploy AI at all, both guests outline a disciplined framework for financial services organizations ready to move from experimentation to trusted, scalable agentic systems.

John Heisler, Snowflake & Liam Hynes, S&P Global Market Intelligence

In this interview from Snowflake Summit 2026, John Heisler, head of AI and financial services at Snowflake, joins Liam Hynes, global head of new product development for public markets at S&P Global Market Intelligence, to talk with theCUBE's Dave Vellante and Rebecca Knight about how AI is transforming qualitative investment analysis and democratizing financial intelligence. Hynes traces the evolution from Michael Burry's manual approach — reading millions of words in SEC filings to predict the 2008 financial crisis — to S&P Global's AI-powered rebuild of the "Lazy Prices" framework, which uses Snowflake Cortex and CoWork tools to identify genuine year-over-year shifts in corporate risk disclosures and concentrate alpha in a short book. Heisler, drawing on a neuroscience background, explains how the brain's architecture of question, memory and action maps directly onto how modern agentic systems should be designed. The conversation also explores the distinction between democratizing data — which the internet enabled — and democratizing solutions, which AI now makes possible. Hynes details how a workflow that once required months of manual reading can now deliver an equity analyst a ranked view of S&P 500 companies with new incremental risks in minutes, collapsing the gap between academic research and real-world value. The discussion shifts to determinism: Hynes explains how Snowflake CoWork's stored-procedure functions ensure agents return consistent, repeatable outputs rather than probabilistic guesses — a critical requirement for production-grade financial applications. Heisler argues that governing data is governing AI, pointing to role-based access control and semantic commonality as the foundation for enterprise-wide intelligence. From building a clear lineage from business strategy to data strategy to AI strategy, to knowing precisely when not to deploy AI at all, both guests outline a disciplined framework for financial services organizations ready to move from experimentation to trusted, scalable agentic systems.

Share this session

Clips
More from Snowflake Summit 2026

John Heisler

Head of AI for Financial Services

Snowflake

Liam Hynes, PhD

Global Head of New Product Development for Public Markets

S&P Global Market Intelligence

play_circle_outline From Burry to AI: Flagging Year‑Over‑Year 10‑K/10‑Q Risk‑Section Changes as Investment Signals

play_circle_outline Augmenting Risk Detection and Short‑Book Alpha: LLMs in Snowflake’s Cortex and CoWork Production Workflows

play_circle_outline Democratization of solutions accelerates research‑to‑production and broad access

play_circle_outline Determinism for LLM outputs via stored procedures, skills, and orchestration

play_circle_outline Semantic Normalization and Open Semantic Interchange: Governance and Role‑Based Access for Enterprise AI Data

Info
Transcript

John Heisler, Snowflake & Liam Hynes, S&P Global Market Intelligence

John Heisler

Head of AI for Financial Services Snowflake

Liam Hynes, PhD

Global Head of New Product Development for Public Markets S&P Global Market Intelligence

In this interview from Snowflake Summit 2026, John Heisler, head of AI and financial services at Snowflake, joins Liam Hynes, global head of new product development for public markets at S&P Global Market Intelligence, to talk with theCUBE's Dave Vellante and Rebecca Knight about how AI is transforming qualitative investment analysis and democratizing financial intelligence. Hynes traces the evolution from Michael Burry's manual approach — reading millions of words in SEC filings to predict the 2008 financial crisis — to S&P Global's AI-powered rebuild of the... Read more

explore Keep Exploring

What does the "Lazy Prices" paper show about using changes in the risk sections of 10-K and 10-Q filings to predict stock returns? add

How has the advent of large language models (and tools like Snowflake Cortex/CoWork) changed the way text-change signals are used to build and improve long–short portfolios compared with the 2020 approach that only flagged any text change? add

Is the current trend the democratization of data (as in the dotcom era) or the democratization of solutions? add

How can large language model agents be made to return deterministic, repeatable answers when querying database-backed data (for example by using Snowflake CoWork's semantic views and stored‑procedure‑style functions)? add

How should organizations govern data and AI — including role-based access control, semantic normalization across the data estate, and open semantic interchange — so that agents and analysts can securely and accurately find and use the right datasets within the enterprise and across the wider ecosystem? add

bolt Powered by CUBE AI

John Heisler, Snowflake & Liam Hynes, S&P Global Market Intelligence

search

Rebecca Knight

>> Hello everyone and welcome back to theCUBE's Live coverage of the Snowflake Summit here at Moscone Center. I'm your host, Rebecca Knight alongside Dave Vellante, co-founder, co-CEO of theCUBE I would like to welcome our next guest to the show. I'd like to welcome back Liam Hynes, global head of new product development for public markets at S&P Global. Welcome back, Liam.

Liam Hynes, PhD

>> Thanks, Rebecca. Thanks, Dave.

Rebecca Knight

>> And John Heisler, industry principle financial services at Snowflake. Thanks so much for coming on the show.

John Heisler

>> Thanks for having me.

Rebecca Knight

>> So Liam, I'm going to start with you because S&P Global is using AI for its qualitative investment analysis, which is a process that's historically been intensely manual, which is a vast understatement. What was the catalyst for using AI for this kind of work?

Liam Hynes, PhD

>> Sure. All right. Well, maybe we can have a little bit of a quick history lesson.

Rebecca Knight

>> Okay. We're game.

Liam Hynes, PhD

>> Let's say, I'm sure everybody remembers Michael Burry who used to be the head of SI and-

John Heisler

>> How can you not remember?

Liam Hynes, PhD

>> Exactly, right? So famously predicted the subprime mortgage, the great financial crisis. So back in 2005 or 2006, Michael Burry would take all of the 10-K documents and 10-Q documents from the SEC filings, he'd lock himself in a room and he would read thousands, tens of thousands and millions of words of financial risk documents. And that's essentially how he predicted the global financial crisis was by going into a room and reading all of these risk documents from the SEC. Now, that takes hours, weeks, months of reading time to be able to do that for a human. And actually subsequently in 2020, there was a paper called Lazy Prices, which was published by a professor from Harvard called Lauren Cohen. And what Lazy Prices looks at is it looks at these 10-K and 10-Q documents. So for people that don't know, a 10-K and a 10-Q is a filing that a company has to do with the regulator to disclose their financial statements, but they also, as a fiduciary duty, they also have to say, what is the risk to our shareholders? So in a 10-K filing and a 10-Q filing, you have a risk section. You also have a management discussion section. So this paper, Lazy Prices, does something very, very simple. It looks at this year's risk section and then it looks at last year's risk section. And if there's any change in the risk section, it flags that as bad. So the Lazy Prices paper builds a very straightforward, long, short portfolio. The long portfolio holds companies that have no change in their risk section and a short portfolio holds companies that do have a change in their risk section. And the long portfolio outperforms the short portfolio in the long run, right? Now, this was published in 2020 before the advent of large language models. So the long portfolio doesn't even look, or the short portfolio and long portfolios, they don't even look at the text that was changed. They just flag that there was a change in the text.

Rebecca Knight

>> Was a change.

Liam Hynes, PhD

>> So agnostic of what the language is actually telling you. So even that kind of, let's say blunt or let's say analog version of that long short portfolio has returns. Cue now, this agentic and large language model era we're in, we actually have tools where we can go in and now identify the risk, identify the changes and actually go in and say, "Okay, well this risk is good or this risk is bad." Or maybe they've removed a risk and that's why it was flagged. And now using a large language model, you can actually go in and augment the short book. So we went and we actually, with the help of Snowflake, went and rebuilt that research. But then with all the Snowflake Cortex and Snowflake CoWork tools, we were able to augment the alpha in that short book by removing the companies that didn't have any risk and just concentrating on the companies that did have the risk.

Dave Vellante

>> So why do you think it worked historically?

Liam Hynes, PhD

>> Good question. Why did it work historically? Well, if you think about the document itself, the language actually is articulating to shareholders what the risk of the company is. So if you go in ... And one of the reasons they called it lazy prices is I think probably two reasons, but one reason is that your investor relations department at a company and your executives, normally what they actually do is they actually go in and copy the risk section from last year and they copy and paste it into the new risk section.

Dave Vellante

>> New risk section.

Rebecca Knight

>> So it's very lazy. It's very ...

Liam Hynes, PhD

>> It's a bit lazy. Yeah. And then another way to say the lazy prices was because that the market is lazy to catch up with it. Because if I go in and I add, let's say two or three sentences to a document that has 10,000 words in it, the prices are very lazy to catch up with that new information. There isn't a press release. There isn't a big song and dance about this has changed. It's kind of hidden in the document. So that's the aspect of the lazy prices. Why has it worked is because if you add something to a risk document, it's probably going to be a new risk. So if there's a change year on year in the risk, it's normally a bad thing. It's normally risk has been added or something like that. So that's why it previously worked.

Dave Vellante

>> And going forward though, it might have detected a font change in the past and now it's going to tighten up the accuracy, I presume, right?

Liam Hynes, PhD

>> Yeah. Yeah, exactly.

Dave Vellante

>> I heard, John, you were talking about you were in neuroscience. What's the relationship between neuroscience and financial services?

John Heisler

>> Man, I wish there was a straight path to it. It would have saved me a lot of time.

Liam Hynes, PhD

>> Well, actually, you know what? Michael Burry started off as a doctor.

Dave Vellante

>> Yes, that's right.

Liam Hynes, PhD

>> As a physician. Yeah.

Dave Vellante

>> And now he's an AI prognosticator.

Liam Hynes, PhD

>> Yeah, exactly. Same with you, John.

John Heisler

>> The best paths are meandering. So there was no direct connection to this, but I'm very lucky that I had that undergraduate and that graduate experience. It was learning and memory. So specifically what are the neurological underpinnings of having an experience and building a memory, which is exactly how these LLMs are built. That's exactly how they're built. So that background really plays a lot into what I do now. I talk about it all the time actually to the point that people roll their eyes at me. I say, "That's exactly how the brain works." So when we think about modernizing the research process within financial services, how do I do that? Well, what does your brain do? Well, your brain has a question. It says, "What memories do I have?" It prunes those memories down and then those memories drive an action. That sound familiar to anybody? It's the same exact thing. And so it's really helped me think about and really boil things down to first principles when I talk to customers about how should we be building these agentic systems? When should we build a skill? When should we be using sub-agents? How should we be thinking about agent to agent communication? I always just go back to my canonical model of the brain. Well, how does the brain do it? That's not how the brain does it. Probably it's not the right way to do it. So it's been fascinating to me. I had no inclination that that's where this would lead, but it applies every day.

Rebecca Knight

>> Wow. So when you hear Liam talk about how they're using this for qualitative research, AI for qualitative research, what does it signal to you about where the rest of financial services is going or could go?

John Heisler

>> So two things with that. One, just a method. And so there's methods, but then there's also the implication of actually this lazy pricing approach. So from a methods perspective, I think you heard us talking about this, think about the implications of scientific research or fundamental research, peer reviewed, institutional research. If I can take something like that, that's the bleeding edge, an academic just came up with this concept, they put it through peer review, typically there's a massive lag between that research and value. If we collapse that by 10%, that's massive. If we collapse it the way that he did, the implications to the financial services industry and just I would say, not overly dramatically, humanity in general are massive to go from research to implementation to value is huge. So just that method. But then the actual approach of lazy pricing, which is, what did you say? A million words he read? That's a big context window.

Liam Hynes, PhD

>> Yeah. Big context window. Yeah. .

Rebecca Knight

>> He had a special brain.

Liam Hynes, PhD

>> Yeah. He was though.

John Heisler

>> Speed reader.

Liam Hynes, PhD

>> Yeah, yeah, yeah.

John Heisler

>> Speed reader. If you take something like that, which is a massive amount of context but really whittle it down to what is signal within what is largely noise to this short versus long, that's massively impactful. So can we take that approach to internal research notes? Can we take that approach to deal sourcing? Can we take that approach within insurance claims and claims notes? So it's this idea of instead of treating a massive corpus of documents as one thing or a document, can we find the thing that drives signal and focus the AI on that? Those are the two things that I think of.

Dave Vellante

>> Is the compression of that ... Because you're right, it can be a decade or more before you can kind of commercialize these papers. Is the compression due to ideation or is it democratization of that knowledge?

Liam Hynes, PhD

>> It's the latter. It's the democratization. So if you think about what happened in the dotcom area, so when the internet came on, what was democratized then was data. Data and knowledge. I can go online and I can go and look up the data that I need to go and solve a problem. So everybody had the library at their fingertips basically. But you still had to take that input and you still had to solve the problem that you're trying to solve. Whereas now, back then it was the democratization of data, now it's the democratization of solutions. So what's going to happen now is that like you said, you're going to have this acceleration in ... There's this niche area where you've got academic research. Now all of a sudden, because we have these tools, very quickly, you can have this solution broadcast to multiple different stakeholders because it's so quick to be able to stand up what happened here and distribute it to the rest of the world. Number one, because there might have been a very technical lift to do this, but the tools that we have now are making that technical lift a lot easier and you democratize that solution now to everybody. So previously, this lazy prices research would have been available and would have been done for let's say quant hedge fund practitioners, but now we can take this and we can productionize it and now a fundamental equity analyst has it in his hand. He can log in the morning time and say, "Scan the S&P 500 companies that have released their earnings this year, take their SEC documents and identify, out of all of those companies, companies who have new incremental risks." And you can have that done in 10 minutes and that just gives that edge and that incremental piece of information to that equity analyst in minutes rather than locking yourself in a room for two or three months and trying to understand the documents.

Rebecca Knight

>> But there'll be no hedges left because the market will get pretty smart, right?

Liam Hynes, PhD

>> Yeah. Well, we could probably talk about how alpha is turning to beta much, much quicker basically. I think that's the trend.

Dave Vellante

>> Well, and the speed to respond to earnings transcripts now. I mean, if people just ingest those, how many times they said AI and how many times they said supply chain and analyzing that. Okay. Where do you guys see your relationship going? What's the direction? What do you want to say a year from now that you're not able to say today?

Liam Hynes, PhD

>> Well, I want to be able to say ... So this time last year when I was here, I would say the past year, the acceleration in the agentic workflows that we've built for clients has just accelerated basically. Now I would say last year, there was some nice new, shiny new tools and we were trying to get used to what the playground was, and I would say probably in the past two or three months that's accelerated. But in the next year, we're going to be able to expose all of these workflows maybe to clients via an MCP or something like that. But the idea is that clients will be utilizing lazy prices frameworks. They'll be utilizing risk analyst frameworks, equity analyst frameworks and plucking these all off the shelf, and hopefully John will help me do all the hard work to get that done in front of our clients, will you?

John Heisler

>> Challenge accepted.

Liam Hynes, PhD

>> Yeah, yeah. Okay, great.

John Heisler

>> Of course echo all of those, but I think what's really interesting within financial services is how closely knit first and third party data are. And that's happened for decades. That wasn't restricted to the digital revolution. What excites me the most is how quickly we can look at these existing workflows and say, "Okay, well, which pieces of this need to be built into an AI layer?" And the role of S&P's data, the role that S&P plays in that is now keenly differentiated. The data that you provide underpins those workflows in a new way. So I'm very excited for that as we start to see people build agents on top of ecosystem data and their first party data.

Liam Hynes, PhD

>> Yeah. And I actually think it's going to be a lot more deterministic as well. So I remember when this first came out a few years ago, people would say the main gripe was they'd be engaging with a large language model and they get an answer and they go, "Okay, great." But they ask the same question again and they get a different answer. So I think what the next layer, and actually funnily enough, Rebecca, we spoke about this last year was that kind of uploading that IP or that skill to the model to be able to interpret things like this. So as an example, anytime you ask an agent something, you want it to be deterministic. You want to get the same answer back. So that's why when semantic views came out a couple of years ago, that's great because it means I can conversationally engage with the data frame and the data that I just set up. But my next step is, well, you know what? I'm asking a question and I know that I definitely wanted to come from this portion of the data and I wanted to use maybe this revenue variable or something like that. Well, that means that I come along and I can use one of the tools that Snowflake CoWork has, which is setting up a function, which is really like a stored procedure. So that means anytime I'm engaging with the database, the large language model knows, "Oh, he's asking about revenue revisions for the energy sector this year. That means I need to go to this skill and this stored procedure to get the correct information to answer that question." So we have all of that ecosystem now within CoWork. That makes it far more predictable, much more consistent, and that you get a deterministic output. Because if I'm an equity analyst, I want to make sure that every time I ping my agent to give me an answer, it's giving me the same answer all the time.

Dave Vellante

>> So you're bringing determinism to the probabilistic and the generative piece of that. How do you ensure that you don't create islands of intelligence and you're able to share that knowledge across the organization? Because determinism, we have determinism today in our SaaS apps, but it's kind of a halfway house because it's within each department you've got determinism and then the human debates across departments. How do you ensure that enterprise wide determinism?

Liam Hynes, PhD

>> Yeah, that's a good question. And actually it's one of those things where it gets to ... What you're describing is that you will have a skill that you can write to carry out this task and multiple skills. And what happens is at the end of the day, you actually might have 100 skills, 200 skills basically. And there might be some duplication in those skills as well. But one we've only just started to play with it is the orchestration aspect in Snowflake is where you can actually have, I guess you're calling it orchestrator because you want to have your conductor of the orchestra, but you can say the orchestration aspect will enable you to be able to go and pick multiple skills to be able to conquer a specific component. So I think that aspect of it is interesting where it'll be deterministic, but you still have to be able to go in and orchestrate which skill to pick and which one not to pick. And also as well, it's one of those things where you might have written 100 skills, but sometimes you actually have to take a step back and say, "Okay, can I actually make this a bit more efficient? Is there two or three skills that I could combine into one skill that will solve these problems and just refine it a little bit better?" So you have to keep auditing, you have to keep the checks and balances going to make sure that you're getting the results that you need. Always working on it for sure.

John Heisler

>> One of the ways that we think a lot about this is governing data is governing AI. So when we think about role-based access control, I'm John. Whether I'm using an agent or I'm querying the data directly, I'm John. I have permissions to certain data sets and I don't have permission to other data sets. So that's a big piece of this, which is mechanical. But let's say you build those skills, you still have a data estate to navigate. And so an intelligence layer interacts with that data estate differently than we do. You ask an analyst a question, they know. They say, "Oh yeah, I need to go to front office data. I'm going to pull this data." They immediately know where to go. If you ask a trade operations question, they immediately know what tables to go to. That's semantics. So if we can normalize semantics across an enterprise ... And within the four walls of an institution, that's excellent. If you then blow that out to the entire ecosystem, that's massive. So semantic commonality, both structural and the actual concept, huge. Open semantic interchange is something that we talk a lot about. That's really driving toward that.

Dave Vellante

>> Yeah. OSI, the emerging standard that's going to be very important in resolving that dissonance.

John Heisler

>> Yeah.

Dave Vellante

>> Cool.

Rebecca Knight

>> Exactly. So last question, looking ahead, for both of you, what is it that is going to separate the financial services organizations that really are able to harness AI in terms of the new skills that financial teams will need to be able to have in this new era in working with AI, what will separate those from those that get left behind? The ones that are able to excel?

Liam Hynes, PhD

>> You know what? I think about this all the time and it's the companies ... Because what the issue now is that you have clients here who are doing their process day in, day out, but it's idiosyncratic. And there's multiple clients, thousands of clients around the world doing that. And then you have vendors like us and like S&P and Snowflake who are trying to help them to get where they need to get to. The clients that handhold and sit down with their ... Sorry, the companies that sit down with their clients and literally say, "Okay, what do you do on a day-to-day process? What is it, when you log in, you sit down, what is your process? What is your brain telling your fingers to type? Et cetera, et cetera." The companies that go in and help those clients understand the workflows that they do, the data that they need, the tools that they're not familiar with to set up these agentic processes, those companies are the ones that are going to succeed, the ones that put their clients first and the clients process first.

John Heisler

>> It's almost like an internal forward deployed engineer.

Liam Hynes, PhD

>> Yeah, yeah.

Rebecca Knight

>> I like that. Yeah. I like it.

John Heisler

>> Internally deployed engineer. IDE.

Liam Hynes, PhD

>> Yeah, absolutely. YEAH.

John Heisler

>> I think somebody already has that acronym.

Liam Hynes, PhD

>> Yeah, you've got the expertise and then you've got this. Yeah. Yeah, that's it.

John Heisler

>> So when I think about it, actually just came from a session and I thought that I was thinking broadly by saying start with data strategy. And the concept was actually, you know what, that's actually too far down already. You need to start with a business strategy. So when people have ... And they nesting doll this. So at the highest level, I have a business strategy. I have a fiduciary responsibility toward my client. That then drives a data strategy because you're going to have no fluff in that data strategy if it's aligned to the business strategy. That data strategy, if you don't have it and start to build an AI strategy is like painting on a bad canvas. None of the things downstream from that are going to work. So having a very clear lineage of your business strategy to data strategy to AI strategy all the way down to a semantic strategy, I think that that's really, really key. And then a more sort of granular piece, we rotated heavy into chat with your data. You mentioned determinism. If I go out and ask the same question every single day, couldn't I get that in a dashboard? So being very realistic with yourself about where and how to navigate determinism versus non-determinism, where AI should be used and AI shouldn't be used. I think about it a lot like willpower. The best willpower, the most powerful willpower is willpower unused. So being very clear about where I should be using AI versus a deterministic process I think is very critical to the success of AI.

Rebecca Knight

>> Excellent. Well, John and Liam, a really fascinating conversation. Thank you both so much.

Dave Vellante

>> Right on.

Liam Hynes, PhD

>> Thanks a million, Rebecca. Thanks, Dave.

John Heisler

>> Absolutely.

Dave Vellante

>> You bet.

Rebecca Knight

>> I'm Rebecca Knight, for Dave Vellante. Stay tuned for more of theCUBE's live coverage of the Snowflake Summit. You're watching theCUBE, the leader in enterprise tech news and analysis.