Jack Kokko – Building the Google of Finance at AlphaSense (EP.461)

0:00A single deep research report that our AI produces now gives them the same 10 pages that they spent three weeks producing with a team of people. So it just gives you now this incredible efficiency and breadth in what you can do. It allows you to do much more diligence and ultimately be more confident in your decision-making and still be a lot faster. So it's going to address the quality and the speed and the confidence all at once.

0:31I'm Ted Seides, and this is Capital Allocators. Our guest on today's show is Jack Coco, co-founder and CEO of AlphaSense, the market intelligence platform often described as Google for finance. The company's 6 ,000 customers canvas 90 % of the top asset management firms, all the world's leading investment banks, and over half of the Fortune 500 companies. Our conversation covers Jack's early frustration as an investment banking analyst that sparked the idea for AlphaSense, the evolution of the business from a simple semantic search tool to an AI-powered research platform, the promise and perils of LLMs in high-stakes decision-making, and Jack's vision of an always-on intelligence machine that will transform how business gets done.

1:23Jack offers a fascinating glimpse at the intersection of technology, data, and investment decision-making. Before we get to Ted's interview, it's football season, which in my house also means it's indoctrination season. Because let's face it, young minds are malleable. And when you've got kids, you've got a once-in-a-lifetime chance to wire them the right way with your favorite football teams. Just ask my four-year-old. Go, Dogs! Sicken! Woof, woof, woof! Now that's an easy one. The Georgia Bulldogs are a college football powerhouse. Three national championships in recent years, tons of glory.

1:59who wouldn't want to be a Dogs fan. But on Sundays... Here we go, Brownies. Here we go. That one's just mean. The Cleveland Browns are famous not for winning, but for testing your character. Year after year, heartbreak after heartbreak. And yes, I made her a Browns fan anyway. Some might call that cruel. I call it parenting. That's the thing about young minds. They believe what you repeat. So, just like forcing your kids to cheer for your favorite football teams, now's the time to plant another seed. Share the Capital Allocators podcast with friends, family, and colleagues in their formative years.

2:33Because if you get to them early enough, they'll be lifelong fans too. Thanks so much for spreading the word. Please enjoy my conversation with Jack Coco. Jack, thanks so much for joining me. Thanks for having me. I'd love you to take me back to your background that led to this entrepreneurial journey. If I go all the way back to where I grew up, that was in Finland. I studied electrical engineering, thought I was going to be building mobile phones. And during that time, I really got enamored by finance. I studied both in parallel. And then one summer, late in those studies, I spent a few months working at a startup in Brussels called Eastac, where we were building.

3:18The company was going to be the European version of NASDAQ, never really took off. But I was talking to US investment banks and trying to raise money for that. business and also made some connections to firms in London where I got an interview at Morgan Stanley and then ended up working as an analyst. That's my first job out of college in London. First during the telecom boom and then convinced them where I wanted to be. My dream land was Silicon Valley and I ended up in San Francisco working on tech deals during the dot-com boom. Spent the first weekend working on a billion dollar deal and things were moving so much faster than in Europe was very exciting at the time, but certainly also a stressful job for someone who wanted to do a good job and the tools just weren't up for it.

4:04And I ultimately planted the seeds in my mind for what I'd one day want to go and build with AlphaSIMs. What were your frustrations at the time doing the work you were trying to do without the tools needed? I had a pretty strong work ethic. I'd work the usual all-nighters and trying to do a really good job doing the analysis, but the pace was really fast and you'd end up not having enough time to do a really good job. I remember some client boardrooms where I'm sweating and barely awake, but also afraid of what did I miss and what am I going to be called on by the CFO or CEO of that company where I just missed something in my analysis.

4:43And that feeling has stuck to me, frankly, still every day when I walk into a boardroom, I have flashes from those situations. That really was because of the lack of technology to help an analyst who needed to consume so much information and trying to catch up on a new industry that you didn't know and new companies you didn't know and sort of cross-sectional information across different industries that you really should have known to be smarter with your analysis, with your viewpoints, and to be able to talk to some really experienced business professionals working on about to bet billions on a deal.

5:16Certainly, we had the data terminals that a refinance professional has still today. That was a big part of the frustration that you could go and manually look for data, but it was very hard to consume it at the scale and speed that was needed. And back in those days, you already started to have technology for consumers where I had Google to search the internet and so forth. And we could get our hands on information very efficiently, but this wasn't available for a professional that desperately needed that every day and every minute in the job. That was a big part of the frustration and has stuck with me.

5:48I know a lot of people still today are doing their jobs with fairly manual tools. What was the vision for what you first wanted to build? In the early days, we even called it, I forget if it was us calling it or clients first calling it Google for finance or Google for analysts. really built a semantic search engine that would understand what an analyst really is looking for when they are reading a financial filing or an earnings call or a research report from Wall Street. We built that system that understood millions of terms and linked them to core concepts, understood that revenue is the same as top line across the whole vocabulary of finance.

6:30So as we built a system that was able to do all that and look at an earnings call happening in Japan or SEC filing in the US, all the information around the world, all companies using different vocabulary and reliably find every single data point about every single topic or theme that an analyst was researching. That was somewhat revolutionary at the time. That just wasn't available. People were still control of searching for individual terms one at a time in PDF reports. Shocking. But that was how work was done. And we had magnified the speed and efficiency and reliability of finding that information.

7:08That was the vision that it started to execute on. And the product we built got pretty quick traction in the hedge fund world where there was this strong thirst for information and efficiency and speed to insight that I had experienced as an analyst. And we were now able to bring to that market. Without going too deep into the technology, what you're describing now, someone would say, oh, yeah, I'll just go into ChatGPT and figure that out. What was it like bringing these pieces together into a technology platform before a couple of years ago? Today, language models are able to be trained and inherently understand a lot of that.

7:45Not all of it, but a lot of it. There's still a lot of value in really specialization and refinement in that industry-level deep dive understanding. But back then, the tools were more basic. We did have an AI first vision from day one, the AI of that day was just a lot simpler. Used AI and trained models to classify information. They would be able to do things like understand the sentiment of a given statement in a paragraph of text in a document and say, is this good or bad for the document? That was something we were really proud about. It took years to build that. We had a team of dozens of people in India tagging very large volumes of those statements to be able to train those models and train it to do that.

8:27one classification task, but it does it so well that even today's LLMs still struggle to be able to do that kind of thing at that deep industry level understanding. And there were many other classification tasks, just taking in feeds of documents and understanding, is this broker research report? Is this an initiation? Is it a change in price target? Recognizing companies in text, being able to do all these things that would help users to slice and dice information efficiently and be able to ask questions that would cut through millions of documents and get to just the right insights, the right documents and the right paragraphs and commentary that you'd be able to very efficiently go through.

9:04So it was much more work to do these individual things that now, for a lot of it, come with large language models as they are trained as really universally capable models when they were very targeted, narrow models back then. When you dive into how you go about implementing this, that first layer of data sources, what's the core set of information that you've wanted to train? And then where are there alternative data sources you've accumulated over the years? We started from the information that was hiding in plain sight, hiding because there was so much of it that it was hard to get to the insights, even though they were available to every professional in the market.

9:47So all the SEC filings, global filings from every country with a stock exchange, earnings call, transcripts, conference presentations at every investment banks. And then, of course, press releases, news, and then broker research, getting Wall Street research on a platform where you could now compare what is the company saying, what is an analyst saying about any topic, any company. And then that was still information that people could get access to on other platforms. One big step for us was acquiring a company called Stream, where they had built an expert transcript library. That's allowed us to start scaling and generating high-value proprietary content that you couldn't get anywhere else.

10:28And we could really point that system to generate information on specific companies, what are their customers saying, what are their suppliers, partners, former employees, executives, saying about things that really matter. Before this, you had to rely on what is the company saying? What are they putting out in the press releases or saying in public forums and filings? But you really had to go talk to management to question that or get alternative points of view. And the expert interviews started to add this invaluable additional perspective. We started to double down on that. That, of course, led to our much bigger acquisition of Tegas, where they were the market leader in that.

11:08They'd built by far the largest expert transcript library and a really incredible operation working with a thousand buy-side firms out there who were doing calls on Tegas and across public equity buy-side as well as private equity venture capital. and you had public company content, private company content, and that started to expand the diversity of what we were able to offer to the market. There wasn't much qualitative research on private companies out there, but these expert interviews started to really pull that in. And today we feel like we've got the richest source of insights on private companies.

11:44We do keep on adding content all the time, but that is where really the core of the focus is because we see so much unique proprietary incremental value that we can add. So what you described, there's some building blocks of quantitative, publicly available information. Then you have company reporting information, and then this huge library of what's called expert opinions. If you were, in that use case, the hedge fund analyst, how would you describe to someone who hasn't used this system what it is they're seeing so that they can pull out whatever information they want to pull out? Today, the easiest comparison point is something like a ChatGPT, where you're asking a language model, a human language question.

12:25The system is now able to precisely understand what you're looking for from your prompt. And now it goes across all the half a billion documents in our system and is able to find the most relevant ones and then dig deep into them and ask those same questions from every single document and see, is the answer here? Is it here? And do that hundreds of times, thousands of times for the most relevant documents for the user to look at and provide a narrative format answer that is granularly cited to those documents. A crucial difference to what people are used to with these chatbots we all use as consumers is that we focus on taking users to those underlying documents.

13:03Our users are serious professionals that care about reading and getting deep into the context that is stated in a SEC filing or research report or expert interview. And we have made our user interface such that it's very easy on the same screen to see all those citations in the narrative format answer, but also dig deep into the underlying document and really understand the context and go deeper and lodge additional queries from there. So you can get a very strong confidence in what you're reading because you know where it's coming from. It's coming from high-quality sources. You can see exactly what company did it come from, what analyst wrote it, or what expert provided an opinion.

13:45And then judge for yourself. Not just trust that, but get the whole 360-degree view of looking at all those different sources and multiple instances of each one on that same topic to gather more of the mosaic of information. And we made that really efficient. So you can very quickly step through all the different breadcrumbs that lead you to that mosaic to then be able to draw conclusions. And of course, the machine is now able to give us those conclusions in that LLM-provided narrative answer. You don't have to draw those connections yourself. You can question what the machine is giving you and ask different questions from different angles.

14:20but at least you get a narrative answer that is very intelligent through chain of thought reasoning that the machine has already done so much work that probably is a lot more than you would have been able to do as an analyst or very few of us have time to spend weeks researching a project and often hear from clients that a single deep research report that our AI produces now gives them same 10 pages that they spent three weeks producing with a team of people. So it just gives you now this incredible efficiency and breadth in what you can do. It allows you to do much more diligence and ultimately be more confident in your decision-making and still be a lot faster.

15:00So it's going to address the quality and the speed and the confidence all at once. As you talk to your hedge fund clients and you're able to get that example of a 10-page report faster, deeper than what would have taken a lot longer, how do they think about the alpha component meaning in the past you would do all that hard work and other people weren't going to do it. Now all they have to do is become a client of yours and they can get that work done. What have you heard from your clients about where they can drive an edge on the market and where does the information that you've created become some type of table stakes?

15:37We are raising the bar for sure as any technology that is introduced into the investment process. now everybody's able to do things much more quickly efficiently and move more in an agile way because the research can be applied to so much more that you in the past you just had to ignore it becomes now a question of who's asking the right questions and how are you asking them and what angles and how do you look at cross industry impacts and read-throughs from this company to that company or this industry to that industry. It's also about technology adopters early and late. And how are you able to adapt to these new solutions and how well can you deploy them?

16:19And people spend a lot of time trying to be great prompters. How do I write a prompt that the system really understands? Well, we see customers asking us, how do we really help them? How do we raise everybody to the same level? So they're very good at asking the machine. And this is even hard for us humans. You ask your colleague a question, and you have to think, did I give a good enough prompt that I can trust the answer? And the same applies to these machine models. But this machine automation really just does the work that nobody wanted to do. The work becomes more interesting, and it's easier ultimately to do the value-adding work when the machine just the heavy lifting and then collects the information for you.

17:00What are some of the initial responses you give to someone who's asking you how to create a better prompt into AlphaSense? I'd go back to how do you ask an analyst that's working for you? How do you make sure that you convey all the information that the analyst needs to know so that you can be sure that your request has been understood? It's the same thing with the machine. If you keep it too vague, it might misunderstand the question or it may make assumptions that you don't like. So if you want to be very clear about what you're looking for, then it pays to be detailed in what you're asking about.

17:38But you can also be iterative. We've built our system so that you have different modes. You can ask questions in fast mode, generative search, where the average answer comes back in six seconds. There's a mode where you let the machine think longer. It takes maybe one or two minutes of chain of thought reasoning. It runs dozens of search sheets, synthesizes an answer, and brings it back. And then there is deep research where you let it work for 10, 15 minutes, and it comes back with a 10-page report, and it's going to have gone much deeper. When you're doing that longer cycle work, you're going to be more careful with your prompts.

18:11You don't want to wait and then realize that you weren't precise enough. But when you're iterating quickly on a six-second cycle, it's cheap to ask lots of questions. So if you don't think the first question got there, then you can ask again. We feel it's our job to make sure we understand the user and understand what they're looking for. And we're constantly refining this and trying to make sure that the system has the, quote unquote, intuition to understand what you didn't say. This kind of is really the value of specialization where we're trying to really understand our financial professional users, our knowledge professional users, and different roles and what they're looking for and allowing the system to take that into account to understand their prompt in a way that their colleague would in that same seat or same industry, same company.

18:56How's in the last couple of years, starting with ChatGPT, I'm curious how that has changed the trajectory of what you've done for a long time before they were around? It's been an incredible breakthrough from a vision perspective. I remember telling our team five, six years ago about how our system is going to be this oracle that you can ask any question in human language and it'll understand your question and go do the research, come back with a great answer. Frankly, I had no idea how and when we're going to get there. When large language models actually started to be able to do that, that was incredible.

19:34It just felt like sort of a Cambrian explosion of opportunity of what we could build with it now that the system can just understand what's on our users' minds so much better than when you just had to express it in keywords. It's just really hard to deduce what someone is looking for when they put in a couple of keywords, but when they put in a full sentence, full prompt, you have a really strong chance of getting to exactly what they were looking for. So it was an incredible opportunity. And of course, it has allowed us to do a lot since then. And it'd be hard to go and cover all of it. Everything really that we've done in the past.

20:08We've almost redone with language models. Although I mentioned the sentiment piece, that is still the incredible sentiment model we built with the prior generation of technology still is running and doing a job that's better than what we see LLMs doing. But mostly LLMs now can just raise the game on everything, every part of what we're doing, and opening up new opportunities. One thing that we built and released just weeks ago was a new AI interviewer. So our Tegas expert transcript library, it's created by buy-side firms doing interviews with experts and analysts at a hedge fund or VC firm, private equity firm has to get on a phone and talk to the expert for half an hour, 45 minutes.

20:47And there's a lot of friction in that process. But what we now have is an AI interviewer that does a pretty good job of that. We're now able to create this new system that is much more scalable where we can point it in any direction and the AI interviewer will hop on the phone anytime the expert is available. And it's also able to help us generate new content sets where maybe buy-side analysts wouldn't be available to do this repeated work where, for example, we're doing channel checks covering dozens of industries and talking to the same expert every month to understand what is going on with pricing and demand, supply dynamics, market share shifts, and so forth in a granular industry and getting a pulse on that industry every single month from the same person and doing that across dozens of industries and getting a pulse on the whole economy that way.

21:36It'd be hard to get by-side analysts to commit to doing that every single month as a repeated process, but AI will do it, whatever we ask it to do. This is a very exciting new product that we were able to launch on the back of AI, being able to do something that just wasn't at all possible even months earlier, and now suddenly it's possible. There were signs of this being feasible in the past, but now able to have the human language conversation with a true expert with high technical and market expertise where they wouldn't talk to an AI unless it was really able to have a conversation, able to respond well and ask smart questions.

22:14I was quite shocked about what I was seeing in the first interviews where it's talking to an expert in the high bandwidth memory market and memory chips able to talk about the various technologies and pricing dynamics and have a very fluid conversation. That is an incredible advancement that we think is game-changing. As soon as we announced that, we started hearing from clients that, can we offload our calls that we like to do ourselves to this AI? There were many people that don't want to be recorded perhaps And I would love to give those calls and say, hey, here's what I'm trying to find out.

22:48Can you recruit the right experts and have your AI do the work and just send me compliance transcripts back? That's another big improvement to the industry's workflows where nobody wants to be spending hours and hours preparing for expert calls. It's a dreaded job. And suddenly, AI can do a lot of it. I think the industry will be very happy about that. and we can still go and get the same insights and people can now focus on what do they do with those insights. Now, of course, the vast majority of this is still happening by people getting on the phone. They do want to do it to the point of the ones that really care about what insights they need to get, they'll do that themselves.

23:24But there's a portion of those calls that get offloaded to AI now and that's pretty exciting. As you worked with the LLMs and as an example of building the AI interviewer, what have you seen as some of the challenges that you had to overcome to make these work the way you wanted them to? There are lots. If we're thinking about technical challenges, it starts from what is the leading edge LLM for this particular task that we're trying to solve? And can we actually get it to do that work? And what is the right combination of can it absorb enough context? In this case, The AI interviewer needs to read a whole deep research board of maybe dozens of pages in some cases to really get expertise on any topic.

24:08Can it hold that in its memory and also take the speech of the expert it's interviewing? And then also when it hears something unexpected, can it do quick research and now adjust what's in its context and pivot and ask a new better question? That took a lot of iterating. You've got to test different models. You've got to test different configurations. we built this our system as kind of an llm agnostic orchestrator that is working with just about every one of the leading edge models in the market and we're deploying them in different ways and depending on the task at hand you'll end up using this model now and maybe another model in a few weeks when a new breakthrough happens so the team has had to build these capabilities of tracking and testing and very effectively staying up to speed on every new breakthrough to understand where is the right combination of these things.

25:00There isn't one perfect model out there. There's models that have been optimized for given tasks better than others. What's the process for figuring that out? Which model is the right one for the right task? It's a very systematic engineering-led process where you're really just testing all of them all the time. We have teams of people evaluating outputs. We have LLMs evaluating outputs. So there's a system that keeps running standard checks. And when you have a new model, you can compare it to the baseline and see, okay, what does this new black box deliver? You can't evaluate it until you've tested it.

25:39You kind of look at the external test results. And there are some of these metrics, some real metrics, some valid metrics, hard to really draw conclusions from what you read. You have to just deploy them and test them and see what is the effective performance. try to test numerical metrics and see how often or what percentage of the time do human analysts agree with the output. But ultimately, there's also a style test. Do I like the way that this LLM speaks? Is it too verbose? Is it speaking specifically in a finance type language where we've ended up even giving financial services, the buy side of a different model and the corporate world, a different model.