Data

CHRO Survey: Hiring is back on the front burner — selectively

Data Analysis

AI Stack

Seizing the generative AI moment

Roundup

Trends

IonQ’s Peter Chapman on quantum computing’s coming of age

Quantum computing’s coming of age

Inside the SoftBank Vision Fund portfolio ecosystem

Video

Remaking the last mile of <br> e-commerce

COVID-19’s impact on technology

CEOs brace for a new wave of challenges in 2023

Strategy

Leading a high-growth company through an economic downturn

Turning IoT’s promise into reality

Building high-performing teams in the age of AI

How CFOs are navigating a changing market

CHRO Survey: Grappling with changing employee expectations

CEO Insights: Thriving in the world of hybrid work

Apollo: Seeding a revolution in Kenyan farming

Seeding a revolution in Kenyan farming

Feature

Paper’s fast learning curve

Fetch: From startup to category leader in a crowded market

Brand new: How a new identity became a unifying force for LTK

Neuron23's CEO Nancy Stagliano: Bringing AI to Life

The strategy behind the SoftBank Vision Fund

Targeting AI, CTOs pursue different approaches

Making responsible and ethical AI a priority

Doubling down on marketing in an uncertain environment

Clearing the path for AI adoption

Venkat Sathyamurthy: Tractable’s mantra for AI success

Venkat Sathyamurthy: A mantra for AI success

CFO Survey: Capitalizing on a brighter economic outlook

Starting at discovery

Jack Kokko: How AlphaSense is creating a copilot for market intelligence teams

Be humble and keep learning

One-on-one with WorkBoard’s CEO Deidre Paknad

Why Ventus Therapeutics always puts patients first

How TravelPerk harnessed the power of human connection

Why a bet on the power of human connection was key to TravelPerk's success

A new C-suite mandate: Maximizing the AI opportunity

How the C-suite is tackling a new mandate: Making the most of AI opportunities

CHRO Survey: Despite layoffs, talent is still king

Despite layoffs, talent is still king

Efficient marketing that moves the needle

Data and ML are creating a golden age of drug discovery

Job van der Voort: The future of work is remote

Job&Talent: Hacking global expansion through M&A

2022 Milken MEA Summit: Tech Investing Panel

CFO Survey: Navigating toward growth

Scaling across borders

SambaNova: Dancing with AI

Cloudbeds: The art of <br> calculated risk

The art of <br> calculated risk

CFO Survey: AI takes center stage as economic woes recede

Generative AI takes center stage as economic woes begin to recede

Thirteen countries, one culture

CEO Survey: Prepping for a rebound

Prepping for a rebound

How Carro leveraged AI to drive used car sales in Asia Pacific

Ensuring human values are central to AI-driven progress

Leveling the financial playing field

CHROs: Hybrid work statistics say it’s here to stay, and employees are in the driver’s seat

Finding opportunity in a cooling market

Mapbox: The fastest route to success

Lenskart: A vision for <br> e-commerce success

A vision for e-commerce success

IPO Readiness: Lessons from the field

IPO readiness: Lessons from the field

<p>As a young investment banking analyst at Morgan Stanley, Jack Kokko always worried that he might have missed something. As part of his job, he would have to present detailed financial analyses about potential M&amp;A deals to executives and boards of directors. Despite spending untold hours doing research and due diligence before each presentation, there was always a “what if?” in the back of his mind, especially when the stakes were so high.</p><p>“It was so tedious, so cumbersome, and manual that there was no way you could have done justice to the importance of the task,” Kokko says. “There was a constant fear of missing something critical. The cost of uncertainty was huge.” </p><p>Kokko started AlphaSense in 2011 to solve that challenge for financial analysts like him, but also for decision makers in businesses who are evaluating impactful investment decisions — be it an acquisition, the launch of a new product or service, or something else. </p><p>Today, <a target="_blank" href="/portfolio/alpha-sense">AlphaSense</a> is used by over 5,000 enterprise customers, including giants like Google, J.P. Morgan, and Pfizer and more than half of the S&amp;P 500 companies. Its path to success was built on its early embrace of AI — and more recently its decision to double down on generative AI. In this conversation, Kokko explains how AI became core to the product, how the company has built guardrails by having different large language models oversee each other for accuracy, and what he’s learned in the process.</p>

<h2><strong>What did the early version of AlphaSense look like?</strong></h2><p>The first incarnation was really to build a search capability across millions of documents from thousands of sources of high-value, paywalled content. Typically, analysts would look for information one source at a time, one document at a time, one keyword at a time. Our system searched everything at once through a semantic search platform, which was able to understand business language and could connect the dots between what the user was looking for and what was being discussed in some document out there in a mountain of data. For example, if you typed in “sales,” it would find “revenue” or “top line” and so forth across lots of synonyms, while ignoring “cost of sales.” Over the years, we expanded it to cover anything a business person or investor would care about, so they’d no longer have to do all that manual work to find the insights they need to support decision making.</p>

<h2><strong>What types of questions does the system help answer?</strong></h2><p>Use cases and questions like: Am I using the right multiples in my valuation analysis? What is the industry and competitive landscape like? How big is the addressable market? What are the right peers to use for this business? What threats do competitive solutions pose? Who is likely to win the market?</p><p>Any search you run will instantly surface an answer from a massive universe of content, including SEC filings and their equivalents overseas, earnings calls, conference presentations, sell-side research, expert call transcripts, trade journals, government disclosures, and more.</p>

<h2><strong>What role did AI play? </strong></h2><p>We had an AI-first mindset from the beginning, even at a time when AI was still in this era of lots of promise but not much delivery. To achieve semantic search, we had to be able to map out all the language that businesspeople use and leverage machine learning to find synonyms for terms automatically. Machine learning was also critical for categorizing information — so we could look at a document and say, this is an initiation report, this talks about a specific company or industry, this is positive or negative for the company. These categorization tasks were quite feasible in the early days, but it still took years to get them truly right. Of course, the LLMs of today have really revolutionized this again and allowed us to go and build something even more powerful.</p>

<h2><strong>How did you incorporate generative AI?</strong></h2><p>For us, generative AI was an opportunity to automate the workflows that users have been going through every day with AlphaSense. We had already provided a really powerful platform for people to find information, but the user would then have to take those details, connect the dots, and summarize everything by writing a memo. Generative AI can do all that better and faster. </p>

<h2><strong>Was that just about having clean data?</strong></h2><p>Cleanliness of data has always been critical. What really had to be built is a system of many LLMs working together where one is optimized to take a question from a user, understands what it’s about, and then creates a research plan that includes the sub questions needed to find relevant information so they can reliably and thoroughly answer the user’s question.</p><p>After pulling and summarizing the answer to those questions, another validation layer and process ensures clear, crisp connections between where the information is coming from so the user can be confident that it’s credible. We have different systems pulling results and underlying data points out of the AlphaSense searches, synthesizing and summarizing them. Basically we’ve built guardrails to make sure that one model checks on the work of other models, sort of like how a manager checks the work of an analyst.</p>

<h2><strong>What LLMs are you using?</strong></h2><p>Most of the leading LLMs, depending on the task at hand. We built a system that is LLM-agnostic. When a new innovation comes out, we test it, we validate whether it is improving the performance of one of the key tasks within the system, and if it does, we can easily swap it in. Which LLM is best isn’t a binary thing. For a given task, it might be the one that has the largest context window that it can deploy reliably, or the one that can be run the fastest, or it can be used in a secure environment so clients’ sensitive data isn’t ever accessible to third-party LLM providers. That’s why we end up with a menu of models for the range of tasks.</p>

<h2><strong>And not being beholden to one or two models was a strategic decision because the technology is changing so quickly?</strong></h2><p>Actually we’d be happy to use a single best LLM if that existed, but the model is only part of the system, and the key question is what is the system that helps to best solve a problem for users. Today, it’s a combination of models in different systems that depend on the desired workflow. </p><p>Given that there isn’t a single best model, if you want to operate at the cutting edge, you have to have a team that is able to constantly track new developments and validate whether a leap that somebody’s talking about is applicable in practice, and how well it can work for your use case. A lot of time goes into discussing, debating, analyzing, and testing new breakthroughs that aren’t always exactly as advertised.</p>

<h2><strong>You talked about this idea of models checking each other. What does it look like practically?</strong></h2><p>We have a model that very rapidly reads everything that is about to go on a user’s screen, compares it to the citations it was linked to, and validates its accuracy. Similarly, if you are capturing quantitative information, another model can check that against historical data points and make sure they are consistent. Frankly, it is very similar to what a human analyst would do.</p>

<h2><strong>You recently acquired a company called Tegus. Was this about having more proprietary data to feed into your models?</strong></h2><p>First and foremost, it was about providing our customers with the best possible intelligence on the market, especially on private companies. A couple of years ago, we acquired a company called Stream, the No. 2 player in what you might think of as a library of interview transcripts between investors and experts on a company — their former executives, competitors, customers, partners. It allowed us to produce information that you can’t access anywhere else, and our customers loved it.</p><p>Tegus is the market leader in that sector and was the natural next step for us. They have the world’s best library of private company content, and we immediately wrapped our search capabilities, our generative AI, our monitoring capabilities around their content and made it much more valuable. We now have the broadest coverage of insights on both public and private companies, allowing our AI to provide even better answers to our customers.</p>

<h2><strong>What lessons have you learned that are applicable to other startups?</strong></h2><p>Generative AI has created a vast opportunity for entrepreneurs to do new things. That’s both great and also can be a risk. It’s very quick and easy for a new startup to build a demo of something, but you have to be careful and ask the hard questions: What am I trying to do? Is that going to deliver something that customers really want and need? If I’m selling to large businesses, what does an enterprise-grade solution look like? It’s what building a startup is all about: going back to first principles and asking yourself, what problem am I really solving?</p><p>And then speed matters more than it’s ever mattered before. It’s likely that somebody else is thinking about the same idea you are thinking about, and they can probably execute it very quickly as well. So you really have to get to market quickly, address the right problem, and build something truly unique.</p>

Amid growing optimism, CFOs are stepping up investment and eyeing growth

CFOs - Q2 2024

Terms

Privacy

Regulatory

Work With Us

Portfolio

Team

Insights

Portfolio Careers

Sustainability

Presentations

Company

Contact

<p>We use necessary cookies to make our site work. We would also like to set optional analytics cookies to help us improve our site by anonymously collecting and reporting information on how you use it. For more information on how these cookies work, please see our <a href="https://visionfund.com/terms#1&amp;cookie-policy">Cookie policy</a>.</p>

Jack Kokko: How AlphaSense is creating a copilot for market intelligence teams

Layering LLMs, the company is helping investors and businesses do research more thoroughly and faster than humanly possible

What’s been the impact of the generative AI work on AlphaSense’s core service?

Job&Talent: Hacking global expansion through M&A

Amid growing optimism, CFOs are stepping up investment and eyeing growth

The strength of 450+ leading tech companies—direct to your inbox

Jack Kokko: How AlphaSense is creating a copilot for market intelligence teams

Layering LLMs, the company is helping investors and businesses do research more thoroughly and faster than humanly possible

What’s been the impact of the generative AI work on AlphaSense’s core service?

Job&Talent: Hacking global expansion through M&A

Amid growing optimism, CFOs are stepping up investment and eyeing growth

The strength of 450+ leading tech companies—direct to your inbox

Sign up for Sōzō newsletter