No one small can hope to win the infra layer in AI (NVIDIA) or the model layer (OpenAI, Anthropic, Meta). The opportunity is to compete at the application layer. There will be three levels of business opportunities in app-level AI based on the amount of work a company can take on:
- The most modest opportunity lies in developing agents that can replace or augment existing worker profiles — software engineers, accountants, marketers, etc. However, this type of intelligence will become commoditized and diminished through competition. It will be trivial to stand up companies here in the long run. We will see excitement and investment but little long-term viability - with only a few durable businesses likely to emerge. I predict market capitalizations ranging from a few billion to tens of billions of dollars.
- The medium opportunity lies in building comprehensive agentic workflows with domain and application-specific reasoning tailored for a fragmented and competitive market. To deeply understand the space and its opportunities and create a specialized vertical operating system. This is analogous to the existing vertical SaaS plays but with more significant growth potential. The unique intelligence and problem-solving capabilities necessary to compete, capture, and retain enterprise customers will support the development of durable businesses worth tens to hundreds of billions of dollars.
- The most significant opportunity lies in developing a vertically integrated, full-stack solution to compete effectively and capture substantial market share. This approach can substantially increase profit margins in the right market, allowing these companies to compete on price and establish market dominance. These markets often feature only a few significant players or have deep regulatory moats. Companies with this business model have the potential to redefine the makeup of global markets.
I believe the latter two opportunities are the durable ones worth focusing on.
The recent essay "Generative AI’s Act o1” by Sonya Huang and Pat Grady had the correct arguments but the wrong conclusions. The three points that resonated with me are:
"[W]e still need application or domain-specific reasoning to deliver useful AI agents. The messy real world requires significant domain and application-specific reasoning that cannot efficiently be encoded in a general model.”
“Application layer AI companies are not just UIs on top of a foundation model. Far from it. They have sophisticated cognitive architectures that typically include multiple foundation models with some sort of routing mechanism on top, vector and/or graph databases for RAG, guardrails to ensure compliance, and application logic that mimics the way a human might think about reasoning through a workflow.”
“Cloud companies sold software ($ / seat). AI companies sell work ($ / outcome)”
However, the essay doesn’t go far enough. It makes a critical assumption: building individual AI agent types for others is the best way to capture value. It presumes the horizontal or functional SaaS business model—whether they rename it Service-as-a-Software or not.
They are wrong. To take an example, McKinsey, Accenture, the Big 4, Tata, EPAM, etc., are just derivatives of global economic activity, not the bulk of that activity itself. Those firms all do well for themselves but suffer from always being one step away from the actual work.
Horizontal (functional) agents and intelligence will be commoditized
The future of agentic applications is vertical. Verticalization has been the trend over the last few years and will only accelerate with AI.
By ‘horizontal’, I primarily mean functional. In many ways, this is the original era of SaaS — every function or department inside a company would be replaced by software. We’ll have a SaaS tool for the legal department and one (or more) for sales, marketing, finance, operations, engineering, product, etc. The idea is now to reimagine that “software” as “service,” rethinking making tools for a department as creating agents that replace workers — we have built the lawyer agent, or the SDR, marketer, accountant, analyst, or software engineer.
Applying that functional/horizontal software playbook misses the opportunity. There will be some billion-dollar companies that way, but that isn’t the big prize. If you buy individual agents like software, you are recreating the functional divisions that slow down work. You're recreating the friction that exists in organizations. We don't need an AI lawyer; we must reimagine legal work. To use that example, the fundamental problem in the real world is that legal analysis isn’t embedded in workflows and processes.
Horizontal or functional software will also be trivial to create internally or through competing startups trying to sell to enterprises. Horizontal SaaS margins will disappear. Intelligence alone will be a race to the bottom.[1] Individual agents will be trivial to spin up.
Choosing between being a vertical OS or a full-stack vertical competitor
To truly solve a problem, one has to understand it completely end-to-end. Solutions are where the ultimate margin expansion will come from. That opportunity comes from doing the much more challenging job of understanding a complete problem, integrating and mixing AI, software, automation, and workflows to do the actual work.
That means two business models will drive the next wave of innovation: being a vertically integrated “operating system” for customers (currently labeled “vertical SaaS”) or vertically integrating a solution directly.
Choosing between these two models depends on the magnitude of economic transformation, the inefficiency of selling software, the market composition (lots of players / or a few, is there power law in market share), and the equity efficiency of the build path.
In some ways, this can be summarized by the famous Alex Rampell quote — “The battle between every startup and incumbent comes down to whether the startup gets distribution before the incumbent gets innovation.” Incumbents are vertically integrated — the battle in every sector will be whether they will adopt vertical AI tooling before a startup integrates an end-to-end AI solution, reimagining the cost structures in the process.
This can be summarized in two questions:
- How competitive and fragmented is the space? The more competitive a space is, the more likely individual companies inside it will adopt solutions to outperform one another. The more monopolistic, oligopolistic, or regulatory capture in a space, the better it will be to integrate a solution vertically.
- How much of the work in a space is automatable? The more work there is in a space that must continue to be done by humans, the better it is to be a vertical OS vs vertically integrated directly. That way, you’re building high-margin tooling and leaving the human-intensive tasks and management to your customers.
Our bet: vertically integrated life insurance
It is informative to explain the industry we’ve taken on and why we choose to be vertically integrated.
Life insurance companies are traditionally slow to adopt new technologies and unlikely to embrace full-scale, end-to-end automation, even when they purchase horizontal agentic point solutions. The industry has a deep regulatory moat, and even though the market is fragmented (in the United States, no single company holds more than 10%), it operates much like an oligopoly.
Despite their large workforces, life insurers are essentially data and technology companies. The majority of their activities are centered on white-collar, knowledge-based work.
Recognizing this, we saw the right opportunity to vertically integrate a solution, leveraging agentic AI, to transform the industry (and also a generational opportunity for digital money to enter the market).
Our vision is to build the world's largest life insurer as measured by customer count, annual premiums sold, and total assets under management. We are a tech-first, vertically integrated insurance and reinsurance stack. Our secondary goal is to achieve the lowest combined ratio globally (a measure of cost structure) while also operating with a workforce three orders of magnitude smaller than the largest incumbents (hundreds of employees instead of hundreds of thousands).
Nevertheless, we plan to collaborate with life insurance agents and other embedded distribution partners because life insurance agencies are both highly competitive and fragmented spaces that are very human and relationship-driven. We plan to develop tools that empower these partners.
Ultimately, we have launched a fully operational life insurance company. We are regulated and licensed in Bermuda — the insurance capital of the world, a jurisdiction known for its stringent regulatory standards for life insurers. We have the exact requirements as any other life insurer, including actuarial modeling and reserving, capital calculations, underwriting, investments, customer service, claims, know-your-customer and anti-money laundering compliance, risk management, (internal and external) audit, and much more.
We manage all this with just eight people through homegrown software and automation. We have also developed customized AI agents for four key roles: (1) reserving actuary, (2) sales/customer support, (3) underwriter, and (4) risk specialist (in the future, we plan to add a claims agent too) - however, more importantly, we have integrated these agents into automated workflows to handle the full spectrum of insurance operations.
Size of the prize
There is another, not particularly intuitive, seemly small advantage that I think makes a huge difference in building a company vertically: In any of the cases where you're selling something, software, agents, or "outcomes," you still need to build the software, agents, or outcomes to sufficient completeness and robustness that you can actually sell them. You still need a sales team and sales cycle to do B2B.
But if you are just doing the work yourself, the software can be used internally, and that is SO much faster and more flexible when building with a good team.
Suppose you're founding a startup or a new business unit today. In that case, you must ask yourself: Will fragmented incumbents dominate the space I’m attacking, or is this a once-in-a-lifetime decade to redefine it?
If incumbents are likely to maintain their dominance, focus on building the operating system that powers the industry. This can be a highly profitable business with significant margins to capture.
However, if there is an opportunity to create a vertically integrated solution, seize it. The work may be more challenging — you will have to slog through rugged terrain; that’s why the space is defensible. In our case, life insurance and annuity premiums account for 3% of the world’s GDP. The market capitalizations of the largest public life insurers are over $100 billion. Our aim is significantly higher. Yours can be, too.
[1] Marc Andreessen recently said, “Are all those {AI} companies actually in a race to the bottom in which it turns out that selling intelligence is like selling rice?” If intelligence is rice, you have to make paella or risotto — somewhat challenging to cook, with a lot of other ingredients. Andressen was talking about the big model providers like OpenAI, Anthropic, and Google, but I think it applies equally well to the next level of horizontal intelligence — the accountant agent or the SDR agent.