Identifying Winning AI Startups
Amidst all the hype around AI, how do you find the next billion dollar startup?
Since the first demo of ChatGPT was released in November 2022, the pace and capability of AI-first companies have been astonishing. This rapid advancement has led to a surge in the number of companies seeking funding. As venture investors, this period reminds us of the mid-1990s, when numerous internet startups emerged, making it challenging to identify the winners. With the excitement surrounding GenAI, it is important to develop a robust framework for identifying the right startups.
In this article, I aim to outline a framework for identifying promising AI startups, drawing on lessons learned from previous technological waves.
I discuss:
Learnings from post-war technological revolutions
Three exciting areas to invest in AI
A framework to identify winning startups
Plan to sharpen thesis and source startups
Part 1: Learnings from post-war technological revolutions
Since WWII, there have been four major technological revolutions that have changed the world. These are the inventions of computers, the internet, the cloud and mobile. And now there is a fifth - generative AI.
In a recently released report by Goldman Sachs on GenAI, Kash Rangan, GS’s Senior Research Analyst covering software, talks about an interesting progression called IPA - Infrastructure first, Platform next, and Application last - that every major technological innovation goes through. He argues that the AI cycle is still very much in the infrastructure buildout phase, so finding the killer application will take more time, but believes we will soon get there.
History has shown that usually, there is a crucial breakthrough in the infrastructure or platform stage that unlocks mass-market innovation at a later point. Let’s look at some of the historical triggers and the mass-market innovations they enabled.
Source: Menlo Ventures
The above table shows that it typically takes 3-6 years after a major technological breakthrough for a killer application to surface. However, in the immediate 3-6 year period and even afterward, there is often a lot of noise and false positives. For every Amazon, there are many like Pets.com, Kozmo.com, Boo.com, and eToys—companies that grew quickly and shut down even faster.
I believe this generative AI wave is fundamentally different from other technological waves, hence we might see a killer application emerging even before the 3-6 year mark. My reasons to believe are as follows -
Established Infrastructure: We already have the railroads; now we need to build the amusement parks, shopping centers, and other developments to induce demand. GenAI services can be seamlessly integrated with existing infrastructure without requiring end-users to shift to a different platform (e.g., from computer to mobile or from on-premise to cloud).
Led by Incumbents: Unlike previous waves driven by upstarts, this technological wave is led by established giants (Mag 7). These incumbents have deep pools of capital and extensive distribution networks, which can significantly accelerate the pace of adoption.
In summary, while GenAI might still be in the infrastructure building phase, the launch of ChatGPT in 2022 has already revolutionized the industry and accelerated the pace of innovation. Historical precedents suggest that we might see a killer application emerge between 2025 and 2028. However, given the rapid pace of adoption, I wouldn’t be surprised if we see a killer application as early as 2024.
Part 2: Three exciting areas to invest in AI
I like to think about the utility of AI to humans on two dimensions:
Task complexity, i.e. whether the task is a) Simple (routine, low cost of failure tasks that require limited cognitive effort) or b) Complex (difficult problems with high cost of failure requiring high cognitive effort)
Impact on work, i.e. whether genAI a) increases efficiency or b) creates something entirely new
Using the above two dimensions, I arrive at the following 2X2
My research and conversations with industry colleagues indicate that generative AI excels at simple tasks but often proves counterproductive for complex tasks, wasting human time. A few industry excerpts:
“We are redefining the way architecture firms fill RFPs for construction projects, cutting down a process that once took days to mere minutes using AI. The efficiency gains are unreal. Our AI model, trained on proprietary client data, fills the RFPs to near perfection.” - Director, Strategy @ Agnetic.ai (backed by Sequoia, General Catalyst, NEA and General Catalyst)
“While we have been successful at business process automation using AI mainly in the finance function, it’s hard for us or any other company to develop the AI CFO - an agent that handles complex tax questions, forecasts revenue and costs, designs capital raising strategy etc. The tech is not there yet.” - Founding team member, Kognitos (backed by Khosla Ventures)
“AI needs to earn our confidence by delivering on the simple tasks first, before we even think about handing over the complex tasks” - General Manager, IT at a large CPG company
Given the technological readiness and the industry's need to build trust in generative AI, I would focus on companies that are addressing simple tasks.
Having said that, the following three themes seem most exciting to me for a venture investment:
Vertical Enterprise AI: My research suggests that two industries are leading the way for AI adoption - IT and Professional Services. For e.g. law firms have heavily adopted a legal AI tool called Harvey for document compilation, litigation preparation, compliance etc. Over time, I expect more conservative industries like construction and transportation to embrace AI. I’d be on lookout for exciting companies building vertical specific enterprise solutions.
Source: Goldman Sachs
Full-stack AI agents: I believe the biggest unlock in productivity will happen when an AI tool can think and act independently to complete a series of simple tasks end to end. Such a delivery will result in a full-stack experience for the end user. Cognition Labs AI agent Devin is an example in the software development space.
The AI Middleware: While we have foundational models and companies building applications on top of them, my research indicates a growing need for specific services like data infrastructure, data security, model customization, and prompt management. These services are crucial for accelerating AI adoption by enterprises. Many exciting startups are already building in these niches, such as Gable (data processing), OctoML (model training), and Baseten (deployment)
Part 3: Framework to identify the winning startups
During my two-year stint at Matrix Partners, India’s leading venture capital fund with over $1B in assets under management, I learned and internalized a proven framework for investing in startups. This framework, which I believe still holds true, is known as FPBM, focusing on four key dimensions: Founder (Team), Product, Business Health, and Market.
In the context of analyzing AI startups, here’s what I'll look for within each of these dimensions:
Note: Bolded points indicate higher priority within each dimension
Founder (Team):
● Balanced team that combines AI expertise with deep domain experience or a strong interest in the industry they are targeting
● Qualities I’d like to see in the founders: ambition, missionary, ability to sell, humility, integrity, grit
● If building for enterprises, the team should have sharp POV on integrations required with existing workflows
Product:
● Early signals of product market fit (e.g. high customer retention, strong referral rate, etc.)
● Ability to get embedded fast in current workflows
● For enterprise products, matches the quality of incumbent SaaS products, if not better
● Structural distribution advantages, e.g. network effects
● Seamless user experience
● Easily scalable as the demand grows
Business:
● Differentiates on data rather than model quality (my view is that models will get commoditized and true differentiation will be result of differentiated data)
● Established revenue model - subscription, usage based, or hybrid
● POV on reducing inference costs over time
● Strong operating metrics - high customer growth, low turnover, manageable overheads costs
Market:
● Incumbent competitors are slow moving, don’t have a clear vision of AI use cases
● Existing labor costs are high, and hence are ripe to be disrupted by AI
● White-collar service functions (e.g. customer support, business development, advertising, etc.)
● Types of tasks are mostly homogenous across different firms within the market (less product customization is required)
● Large TAM
Part 4: Plan to sharpen thesis and source startups:
The strategy laid out above is the starting point. I believe it is crucial for us to remain deeply engaged in the startup and the broader AI ecosystem, constantly absorbing market signals to refine and sharpen our investment thesis. Here is a detailed plan to ensure we stay grounded and position ourselves to connect with innovative founders building in our space
● Attend quarterly earnings calls of public market companies, especially Mag 7, semiconductor and SaaS companies to better understand broader developments in the AI space
● Attend developer conferences like The AI Conference 2024, Ai4, etc. to stay current with latest on-ground technological developments
● Do 10-15 networking calls in a week to get a better handle on industry specific AI developments. This pool of people will include - other VCs, established and funded founders building in AI, and executives at big tech and leading SaaS companies
● Build a strong sourcing top of funnel by focusing on outbound sourcing strategies - targeted cold reach-outs, founder references, professional database/social media scanning (e.g. LinkedIn, AngelList, Twitter, Pitchbook etc.)
● Use content-led strategy to increase inbound interest - publish thought pieces or start a podcast to share hot takes and insights on AI
● Target speaking with 10-15 startups per person in a week for evaluation purposes
If you have any questions or feedback about the thesis or the framework I have discussed, please feel free to reach out.