How I Invest with David Weisburd - E433: AlphaSense’s Chris Ackerson on AI, the Future of Finance & Finding Alpha

Episode Date: September 23, 2026

What happens to financial analysts when AI can do the work they used to spend all night doing? Chris Ackerson is SVP of Product at AlphaSense, where he focuses on applying information retrieval, natural language processing, deep learning, and recommendation systems to search and discovery. We discuss how AI is changing the analyst role, why vertical AI can outperform general-purpose models in financial research, and why proprietary data may become one of the most important competitive advantages in AI. Chris also explains why LLMs hallucinate, how AlphaSense is using AI to conduct expert interviews, where humans still create investment alpha, and why the future of financial research may look less like software and more like an AI teammate that never sleeps.

In software engineering, there's been this paradox where the more effective AI has become, the more there's been a demand for engineers because now they have, they're kind of all 100x engineers. Do you see the same thing playing out in finance, or do you see the analyst layer being disrupted completely? We do. What we're seeing is that, you know, as our systems get more and more capable, we're able to take a lot of the grunt work, the manual work, that analysts were staying up late at night, executing and now they're able to focus on higher value activities like meeting with clients, corporate action events, et cetera. And what that really means, as companies get more productive, they're able to cover more companies, launch more products, accelerate their roadmaps.

We're seeing that internally at Offsense. You know, you mentioned software engineering. Our roadmap is just accelerating as our engineers, our product managers are getting more efficient with AI. The competitive landscape is heating up, and so we're seeing the same thing with our client base. Are people using tools like the superanalysts to supplement their analysts, or are they disrupting them? This debate around kind of labor replacement is one we're going to have for many years into the future.

I think we're seeing the ability for our customers to do much more, which means they're going to invest more in their people, but the roles are changing and adjusting. I worked with a major investment bank that had divested a portion of its business. And so they wanted to challenge Alphacense to see if they could use us to increase the coverage of all of the rest of their bankers. And so they put Alphacin's through its paces over many months and they were able to prove that Alphacin's could increase average coverage by up to 15%, you know, making their bankers much more efficient. Now, what they do with those savings is up to the bank. They may hire many more bankers because they can cover that many more companies. They can deliver better advisory services to their clients, or they may do different things.

But there's no question that the roles and the responsibilities, the work being executed, is going to evolve as these AI systems get better and better. And you mentioned you see your customers doing more and more business. were some early case studies where you've seen people using alpha-sans and tools like agentic AI in order to increase their business. Let's take the hedge fund space, for example, where the job of an analyst is to track an industry, a sector, they're getting up to speed on new companies, they're generating new investment ideas, they're prepping for earnings and keeping track through earnings season, read-throughs that are impacting all the companies in their portfolio.

All of these are areas where AI is ripe to drive significant efficiencies. And so we're seeing the same across every sector that we serve. A lot of people are looking at using horizontal players like an open AI anthropic to solve their investment needs. And you guys are verticalized. How do you differentiate your process from a horizontal AI? And has that evolved over your history? Our thesis has always remained consistent from the beginning,

which is that by aggregating and controlling high-quality data and building AI purpose-built to understand and make sense of that data, we could deliver higher accuracy at lower cost with more trust into the market. And we're seeing that very much today. One of those leading AI labs just last week published data showing that the error rate, using kind of the leading market data MCPs were between 5 and 9%. That is a non-starter for a serious investment professional. Any error rate is going to be a blocker to real adoption.

And that's why people are still spending so much time and crafting, verifying all the information, because these systems just don't work at the level of a professional. analyst. And so the way we solve that problem is we aggregate all of the data, we index all of the information, we enrich it, understanding with AI, reading every line of every document that flows through the system, exactly what the data is saying, the connections and relationships across disparate sources, and then integrating that into the agentic systems. so that they know exactly what information to pull for the particular task at hand.

So we're able to dramatically reduce the error rates, the hallucination rates, and then the ability for the end user to actually verify that information is dramatically faster and more efficient when the data is actually integrated into the system. And you can see down to the individual line and the individual, you know, filing, for example, where that data was sourced from. I think it's one of the biggest distinctions. One of this governor on AI growth is how critical the information is. On one side, you might have hospital systems where people literally die if you have you in a 5% error rate.

On the other side, you're researching treats for your dogs. People expect all aspects of that to grow at the same pace, but really it comes down to what is the data used for and what is the cost of making it sick. That's exactly right. And we've done the work to test this in absence. We've built evaluation harnesses that put all of our AI systems through their paces, building very detailed and specific rubrics for all of the real-world tasks that our users are executing in AlphaSense and running automated evals and then also human evals. We have financial analysts that we've hired at Alphacens to actually spot check and prove that these evals are accurate. And what we see is that by leveraging frontier AI models inside the Alpha Sense harness with access to the Alphsense data and tools, we can deliver 3x higher quality answers at 3x lower cost.

And that's what the vertical integration really enables. This episode is presented by Juniper Square, the operations partner for private markets. Maybe you could solve this road all. I've been asking people for a couple years now, why do LLMs hallucinate? LLM's hallucinate because by definition they're probabilistic. They're trained on millions of documents, data points, reinforcement loop traces, that essentially encode all of that information into a compressed set of model weights. And that allows these models to generalize extremely well.

It also means that the output from the model is probabilistic. It depends on the data it's seen. These models are trained on the entire internet. The internet, as we all know, is full of information that you can't trust and great information as well. And so at Alphosance, everything that comes out of our AI systems is grounded in data that we curate, which is authoritative, trusted information. I mentioned some of it earlier, what companies are actually saying, interviews with

experts, cell side research analysts. And so what we're trying to do is harness the big. benefits of those probabilistic models. That is the fact they can be creative. They can think outside the box. They can do things that sort of classical machine learning can't do, but ground them in data you can really trust and take into an investment committee or to a CFO and a client meeting. I want to go to that data. It seems to be one of the most valuable aspects for a vertical AI company. How's Alpha Sense using data specifically? People talk a lot about data. I still think it's dramatically underappreciated, how critical it is. Let me take a step back and talk a little

bit about the data sets that we aggregate. So I'll talk about our expert transcript library. So this is a data set where the best investors, the best by side analysts in the world, are doing interviews with experts in industry. So people that used to work at companies, former executives, suppliers, customers of companies, and asking them questions that are going to lead to real investment decisions. All of that is proprietary insight. The vast majority of knowledge in the world isn't written down. It's in your head. It's in my head. And so our expert transcript library is uncovering that insight and transcribing it, making it available to consume by our users and now our AI agents to influence the answers, the outputs that come out of our system. That's an

example of proprietary data that you can't find in any other AI system, any AI system that's pointed at the public internet. So the ability to find proprietary insight is one of the key reasons why sort of control of data is so important in our space. The second critical reason is the ability to actually drive high quality, lower those error rates and hallucination rates. So at Alpha Sense, we're constantly training our search models, our AI models, to improve the quality of the actual context that's fed into these AI models. So we have done a lot of studies showing that the newest frontier models are incredible reasoners. They're amazing tools, but they're actually really bad searchers. They don't know how to find the right context for a given task.

And so we're training our own models to do that extremely well. And you can only do that with sort of control ownership of the data. And then the third reason why data is going to be so important to building agents over the coming years is because the agent output becomes a data set in and of itself. So taking a step back in the question answering age, the chatbot age of AI, you ask a question, you get an answer and it sort of like goes into the chat thread. and maybe you look at it again, maybe it goes away. Today, agents are building investment memos and pitch decks,

like real work that itself has value. And so moving forward, the question any agent company should ask is, like, should we be preempting the questions our clients are going to ask and actually produce the outputs that are going to be high value to them? And an example of this today is, you know, Alphacin's is using RAI to go run channel checks in industry. So we can apply. We'll talk about our AI interviewer for...

Let's talk about that. If you guys are not using human beings to interview, you're using AI to interview. A, what does that mean? And two, how do you utilize it? So it's a combination. So I mentioned part of the expert transcript library. The majority today is, you know,

expert analysts at the top byside firms running interviews themselves. And we've built an AI interviewer to supplement that. And so the AI interviewer is an AI agent trained on Alpha-Side. sense data. So understands the market, all the information available on any company or any market, and uses that to build an interview script to go on a call with a real human expert and interview that person, ask them questions, follow up, drill down into particular topics. And so we started the first application of that AI interviewer was the channel check data set I started talking about, which is we're applying this AI interviewer to go talk to people across every sector of the

economy who understand price and demand signals for these different sectors. And so we're transcribing thousands of calls every quarter and then building an intelligence layer on top that's actually extracting quantitative data ahead of earnings about how a particular company's guidance might ultimately play out. and that AI interviewer is now available to all of our customers as well. And so expert calls have always been a critical part of the diligence process at, you know, private equity firms and hedge funds. And, you know, our users are still very much doing their own expert calls, but now they can

augment those, instead of maybe they're running five calls themselves, but now they can augment that with 20 or 30 or 50 AI calls going deeper on different areas. of investigation, and that's creating this content flywheel that's really accelerating the size of that library and all the insights available to our agents. More than 300,000 interviews in that library today, adding more than 25,000 every quarter and growing fast. And are you also meta-learning on those interviews and learning how to better get information from the interviews? Absolutely. So when we first built the AI interviewer, we applied it to channel checks because that's a sort of contained problem. It's a pretty structured set of interview questions when you're

running a channel check about, you know, how much are you buying? What is the demand for a particular product that you see in market? And so we could take a technology that was early and apply it to a really useful case and it did great. As we improved that system, we evaluated and we now measure the quality of the AI interviewer at or above the quality of our best byside analysts running interviews. So it's extremely high quality. That system improves over time through meta learning and it's pointed at all of the data in Alpha Sense. So imagine you were going to give an interview to me and you could read everything ever written on Alpha Sense, everything ever written on the space we operate and that's what the AI interviewer does before every expert called that it does.

going from the early application of AI agents as you guys are doing to the future of work and the future of finance. Fast word five years from now, what are investment bankers doing? What are investors doing? What does human beings end up to? The core roles of these industries will remain critical. Like investment bankers provide advisory services to their clients. You know, hedge fund analysts figure out, you know, where, you know, Where is the edge in a particular market to try to make an investment return? That core, I think, won't change. One of the things we're really excited about, we should talk a little bit about our new superanalyst product.

But today, that superanalyst is able to do incredible work delegated by analysts, for example. So all of the data in AlphaSense is flowing through and can be monitored by superanalyst so that it can automate work, for example, during earnings, reading all the calls and executing your specific process that you want it to execute on. The sort of deliverables that operate in these industries are very document-based, right?

It's generating slide decks and investment memos and Excel models and things like that, and it's fantastic for those. As we fast forward a few years, we imagine a world where everyone is building kind of their own custom software that's exactly perfect for their specific process and their specific use case. So we're going to push the boundaries beyond just documents and meetings to a world where

everyone is operating on live software that always has access to up-to-date information and is doing exactly sort of built for your specific process. What becomes valuable in this future? Human judgment is absolutely critical. the ability for a firm to learn from all the decisions they make is really important. Our CEO, Jack, always talks about the enterprise value of a business is the sum of all the decisions it makes. And in the agenic era, the most important thing a business can do is ensure that they've captured the learning that is happening from all of the decisions that your human capital or your, your agent capital is executing on. And so investing in systems that help you capture that learning

and your firm can get better and better. And you can apply the judgment on top of what the AI systems are doing is going to remain sort of a critical. What else? Outside of human judges. The world is complex. We see today technology is changing so fast. The business world is evolving. applications are evolving. And so the ability of humans at sort of every level of a business to act as almost architects, you know, system engineers on top of all the AI agents is going to remain critical because of that dynamic change that's always present in human society and in the economy. One thing I've learned from talking to hundreds of investors is that great investment firms aren't built on investment returns alone. The firms that endure are great at the things most

people don't see their operations, their relationship with LPs, and the quality of information they use to make decisions. And here's what AI has changed. Every firm now has access to the very same models so the intelligence isn't the edge anymore. The edge is what you could feed it. A firm with its fund operations and data in one connected record can actually put AI to work. A firm running on disconnected systems simply can't. That's why thousands of GPs run their funds on Juniper Square. Juniper Square puts your fund operations, data, and administration together in one connected record. That means less time managing disconnected systems and more time investing, working with LPs, and building your firm. This episode is brought to you by Juniper Square, the operations partner for

private market GPs. Learn more at Junipersquare.com slash how I invest. That's junipersquare.com slash how I invest. The best conferences do two things well. The content challenges how you think and the people in the seats are the ones whose opinions actually move markets. Alpha Summit is AlphaSense's annual user conference, and it's built around both. Join me at the Glass House in New York City, October 5th through 7th for sessions going deep on where AI, data, and human expertise converge. The room will bring together over a thousand institutional investors, corporate decision makers, and capital markets professionals from firms at Goldman Sachs, JP Morgan,

and the top PE and hedge funds. If you listen to the show, you're already asking, the right questions. Alpha Summit is where you go to stress test your thinking with the people working through the same problems at the highest level. Register at AlphaSummit.AI to secure your seat and as a

how invest listener you will get 50% off. Check the show notes for your exclusive discount code. I look forward to seeing you there. There's a bit of a paradox if every hedge fund in the world is an AlphaSense customer per se and they have all at the same data. How do hedge funds compete against the change? This has been

the question on sort of information systems in investing for decades. We think about the ultimate investment decision is a combination of the investor and Alpha Sense or superanalysts, the technology they're using, the way you prompt the system, the questions you ask, the choices you make when engaging with AI, what you pointed at, the process you ask it to automate, that's all the human differentiation And it's the combination of that plus the agent's ability to look over vast quantities of data and automate these processes. That's what ultimately generates alpha.

I also think financial products are themselves products within customers. You might have pensions, endowments, family offices with very different risk tolerances. And you might look at the same probabilistic model and say, well, I'm going to play it safe and get 8% return. or I'm going to try to have the best net return, even if there's a lot of volatility. So there's a taste and a product side from it on the investment side as well. Absolutely. And we think about that a lot with superanalysts. We're not bringing one superanalyst to market.

We're bringing a platform that can be customized for your process. You can bring your own skills. You can integrate superanalyst into downstream workflows that integrate your own data, your own perspective, to really create that differentiated competitive advantage. Just recently have the president of Peter Thiel's hedge fund. He was there for quite a while. He talked about the systemic risk in markets because of the passive investor.

So before you had these active investors that are coming in and making these buy and sell decisions, and now everybody piles into these names. And it's created this leverage in the system. Do you see AI and the symmetry of data? leading to more volatile markets? We want to help markets allocate capital more efficiently. That's our mission. And so we do think that using tools like AlphSense,

we can contain and reduce risk alongside our users. And so I think the ability to bring really high quality data into decision-making, like one of the key value propositions of Alphsense has always been removing blind spots in the, investing process or decision-making process can get a great answer from AI, but if it's coming from one blog or one analyst, that's much less valuable than a system that's looking across all the cell side analysts, a bunch of experts, what people inside a company and the company themselves are saying about a topic and being able to wrap all of that into an answer, that's going to

give you a lot more confidence in the decision you're making and increase the likelihood that it's the right decision for your business. I want to go to your data infrastructure. You're using both the large LMs as well as open source in order to run your inference. Talk to me about that. Yeah, that's right. So, you know, we've seen a pretty remarkable difference in what different models are good at. Even the frontier AI models have different characteristics. How would you just say now? Well, I think it has to do with, you know, the fact that post-training has become a much, much bigger part of the training process, whereas a few years back, all of these labs were pre-training on the same data sets. The public internet was the majority of it. And so

the models coming out of that pre-training were very similar in what they did. Now, the majority of the GPU spend for these firms is not pre-training, it's post-training. And post-training, you can apply a lot more differentiation in terms of what tasks you're post-training on, what data sets you're acquiring to post-train-on. So these models have actually differentiated, which is kind of contrary to a lot of predictions that thought they would all sort of commodify and converge. And so we're seeing, for example, Opus 5 is really, really great at generating high-quality slides, but it's not very good as a context retrieval system, kind of a search model in Alpha Sense. And so we apply different models based on our evaluation of their strengths and weaknesses. And so

you mentioned open source models. We partner with Cerebris, which is a fantastic, really low latency inference provider where we can leverage the best open source models, but at 10 to 20x lower latency than what could be found elsewhere. And so that's really important in trying to build systems that feel conversational and real time. And so all of our tools, whether it's superanalyst or our generative grid are leveraging multiple models in their pipelines that are optimized for these different tasks. And we'll probably talk about this in a minute, but we're starting to train our own frontier models as well where we think we can add a lot of value in certain tasks.

I want to underline something that you mentioned, which is the LMs themselves have different data sets. This confuses a lot of people. I've been helping one of our portfolio companies go through an M&A or capital raising event right now, and I've been talking to a lot of the heads of the Bustin Banks, and they have all said the same thing, which is a large majority of their work right now, at least in the tech sector, is helping the large LM companies buy companies for data sets. Most people don't know that this is going on, but this is a big part of transactions today is helping these large LMs gain more and more

data and more and more proprietary data so that they could start to differentiate against themselves as well. That's right. All AI companies will become data companies, And we're seeing that with the proliferation of companies that are like a mercure that are helping people build reinforcement learning environments with experts helping to build data sets. Recently bought the data remains of Spirit Airlines was an interesting acquisition. And so we're seeing that more and more. And it just underscores the criticality of differentiated data and why we feel so good about the data that we've aggregated over more than a decade in business. And I'll get back to what you were just saying, which is you're training your own data and you're doing your own post-training data as well. Tell me about that.

Alphacin's has trained our own AI models for almost a decade now. Prior to kind of the generative AI explosion, the post-chat GPT era, all these models needed to be post-trained to make them effective. And so we have lots of kind of search-specific models throughout the system that are helping to find the right content, to understand a user's question, to rank sources. And those models have learned over many years, like what are the most relevant sources of information, the most authoritative content for any given task. And then as we move forward, we're training frontier models to actually power superanalyst, for example, end-to-end. When you look at superanalysts, for example, it's a little different than maybe some of the experiences people have with AI agents where you're really kind of engaging with the LLM pretty directly.

In a context like automating financial research, actually the vast majority of the work is the research loop. It's like 80% of the cost and something like 70% of the kind of influence on quality is how effectively you can research across all the tools and data sets and Alphsense and find the right context for. or an earnings preview or a pitch deck or a strategy memo or whatever you're building. And so that research loop, we think, you know, we've got incredible data and tools to build frontier systems that can run that research loop much more effectively than can a frontier AI model that really doesn't understand all of that. Another way, the LMs are weighted, so they have some weight to all the different sources. And through running your own data, you're able to weight them slightly different.

in order to get an information edge on even the LLMs themselves. The frontier LLMs, the way that they're trained to search, is very much sort of web search frame where you just kind of spray and pray. You just search a ton of stuff. You're looking to try to find a diamond in the rough. I think that works for web search generally. It doesn't work effectively for financial research, where having a ton of low value or noisy content is expensive.

It costs a lot of money to run these research. loops. Two, it lowers quality. It's sort of context pollution as you're pulling in irrelevant information. It lowers latency. The tasks take longer to execute. So we can exactly, as you said, train a model that knows exactly what source to go for, what tool to use to execute a given task. And so we think we're going to be able to reduce cost by up to 40x while delivering much higher quality than even frontier models can. And to be clear, that doesn't mean we won't use frontier models in Alphsense. Like we think they're fantastic for the things they're really good at, like writing code, doing deep data analysis, producing high quality artifacts like slide decks or Excel

memos. And so those two systems or models will work together like two separate agents, one to coordinate the actual code generation or synthesis of data and one that is an ultra effective researcher, those systems will work together. I wanted to actually go there. Mark Benioff has explained the future of software as headless. So you're not necessarily going to Salesforce and making prompts. Salesforce essentially does an ADI through what's called an MCP. Is Alpha Sense looking in that direction? And is that where you see the future of applications?

There's no question that the market is going to move in a world where agents work with other agents. They're going to do that through different protocols. MCP is one. And so when we build a world class agent like Super. analyst, we expect it to work inside the Alpha Sense desktop where we've got an agent workspace that's purpose built to make working with superanalysts really great and a delight. All the history and the work that you produce with superanalyst is there. All the skills to tune it are sort of in that desktop workspace experience, but also that

superanalyst is your teammate. It should work wherever you work. And so today it already works in PowerPoint and Excel, but it'll also be available through API and MCP to integrate into other. systems, applications, and workflows that you know, you're engaging with as an individual user or more even on a broader sense, working with, you know, IT teams, technology teams inside these firms to use AlphaSense as an agentic kind of automation infrastructure for market intelligence to build many different applications and automate many different workflows on top

of. In many ways, it's meeting where the customer is today. They're used to prompting and they're used to being an application in the future, at least in theory, the next generation of knowledge workers will be trained on AI native processes. There's no question. And as you agents allow you to go more from sort of question answering to truly delegating work, superanalyst will be working 24-7 like the analyst on your team that never sleeps. It's going to deliver work where you work to your phone, to your email, to your Slack and teams, just like a remote teammate would do in a much more. human interactive way.

At Alpha sense, you guys are growing at a crazy rate. You've grown over half a billion in revenue. But certainly with any business, there are always things that keep you up at night, especially given where the world is today. What keeps you up at night? We're incredibly focused on delivering decision-grade answers and outputs. And so where we focus is ensuring that, you know,

when someone gets an output from Alphsense late on a Friday night, they can close their laptop and go home and spend their weekend without worrying if that data came from a source they can't trust, if there's a hallucinated metric or piece of financial data. That's what we think about day and night. More tactically, in the sort of chatbot, like generative search era, verification was a pretty simple problem for us to solve because we had, you know, all the data integrated in Alpha Sense. And so everything was attributed to the right source, cited, and you

could click in and read the original sources. As these systems get more and more sophisticated, they're not just generating a report where it's easy to click through the citations. They're generating like much more visualized, abstracted recommendations and suggestions and charts and tables. And so thinking a lot about how to reduce the friction on verification and maintain that trust is, you know, what we think about every day. Going back to the beginning of the career, as I mentioned, you're at IBM Watson. A lot of your peers are now working at companies like Open AI, Anthropic, and really building the future. If you give yourself one piece of timeless advice back then, that would have helped accelerate your career,

helped you avoid costly mistakes. What would that be? Everyone working on AI has a lot of scar tissue from underestimating the pace at which these models would improve. And all of us are, you know, grappling every day with how much do we build the scaffolding around the models based on how they operate today to get them to sort of a quality that makes a product usable today versus where the model is going to be in six months or year. And we are now incredibly unsentimental about deleting scaffolding and moving towards the future and going back five years. You know, the advice I would give myself is, you know, the advice that Jack and Raj, our founders have given all of us, which is, you know, just look to the future, move as quickly

as possible. Don't be sentimental, continue innovating. We're incredibly excited about where that's going to lead us. More than ever, it's embracing this Buddhist philosophy of a beginner's mind. You always have to wake up as a beginner every single day because everything is changing so much beneath you. You mentioned the scaffolding. What's an example of scaffolding that has changed within AlphaSense? With every generation of these models to make sure the products that we were launching met the demands of professional users, that they could trust them, that we weren't putting something out that would deliver a hallucination that would impact an investment decision or something like that they would make.

And so like we started our first generative AI features back in, excuse me, we focused on just document summarization. Could we summarize an earnings call effectively? Things like that. And then we've marched towards a generative search solution that could find and synthesize answers from across all the content in AlphSense to workflow agents that allow you to automate complex workflows

to now superanalysts where you can delegate work. And so in each of those evolutions, have essentially deleted a lot of the scaffold required to in 2023 get an LLM at that time to be able to summarize an earnings call near 100% accuracy. There was a lot of work involved in making that happen. Today, that's not necessary.

These models can do that much more effectively. But now, similarly, if you're delegating your entire earnings process to superanalyst, there's an enormous amount of work to make that effective, reduce, hallucination risks, increase accuracy, manage tokens efficiently. So that's sort of where we're focused today. And this is almost a cultural competitive advantage.

You mentioned Jack and Raja talks about what I would call the beginner's mind, always having to reinvent yourself. A lot of people think that's a technological issue, but that's actually a cultural issue. I think that's right. There are investments that we've made that will be durable through the life of the company, the content data sets that we've aggregates.

The search data that we've built millions of search queries that allow us to train models on how to find the right context for the right job. These are durable data sets that will continue to power the company. Looking forward to the ways in which people will engage with a product like AlphSense, you always have to be looking forward and willing to innovate and disrupt yourself.

Well, Chris, I've been a happy AlphaSense customer. once in a while my companies do actually make it to an IPO. I use it to see if I should buy or so. And thanks so much for sitting down and thanks for making a ton. Thanks, David. It was fun.