In the Era of AI, What Is Your Moat?
Why Extreme Domain Expertise is the Last Defensible Edge
Most AI startups today are building on top of third-party or open-source models. Meanwhile, general-purpose LLMs are rapidly mastering specialized tasks, blurring the lines of differentiation. In this landscape, where technical differentiation is eroding quickly, the very concept of “defensibility” is under siege. As an investor who’s diligenced many AI ventures, I’ve seen this tension firsthand. It’s pushing us to dig deeper into what truly sets a startup apart in a world of abundant capability.
The Anatomy of Modern AI Startups
Most AI startups today follow a tripartite architecture: a foundation of third-party or open-source LLMs (think OpenAI, Anthropic, or DeepSeek), a layer of proprietary data for fine-tuning, and an application layer tailored to specific use cases. This structure gives rise to four distinct categories, each with its own degree of resilience OR vulnerability:
#1 API Wrappers: These startups slap a thin application layer atop third-party models like OpenAI’s or Anthropic’s. Think chatbots or basic automation tools. Quick to build, but fragile.
#2 Fine-Tuned Models: These take open-source or third-party LLMs and tweak them with proprietary datasets for niche applications, like a customer service bot trained on a company’s support logs.
#3 Vertical AI Startups: These go deeper, crafting AI products for specific industries (e.g., legal, healthcare) and integrating them into end-to-end workflows. Think enterprise-grade products.
#4 Model Builders: Relatively few startups choose to build their own foundational models.
The Erosion of Defensibility
Here’s the hard truth: as general-purpose LLMs get better and cheaper, the moats for categories #1 and #2 are shrinking fast. API wrappers face an existential threat as large enterprises, armed with budgets and talent, integrate AI directly into their operations, bypassing middlemen. Fine-tuned models aren’t much safer; as foundation models improve, the marginal edge of proprietary fine-tuning diminishes. commoditization is relentless, and the pace of innovation in open-source models only accelerates this trend.
The Data Moat Myth
The idea of a "data moat" has long been championed as a way to defend AI-driven businesses. In fact, many argue that access to proprietary data is one of the most valuable assets an AI company can have. In sectors such as healthcare and finance, where datasets are both rare and valuable, having the right data can enable a company to build models that are nearly impossible to replicate.
But here’s the catch: very few companies possess truly proprietary datasets.
Take Miele, for instance, the German high-end consumer appliance manufacturer. Miele likely holds a treasure trove of data about how customers use their products inside their homes, which is intimate and valuable. However, competitors like Electrolux, Samsung, and LG likely have similar data sets, diminishing its uniqueness.
Most companies overestimate the value of their data moat. Simply owning data is not enough. It must be exclusive, costly to acquire, and hard to replicate.
The Real Moat: Domain Expertise
So where does an enduring moat reside in this AI era? After diligencing dozens of startups and investing in a select few, I’ve come to a conviction: domain expertise is the irreproducible advantage. In a world where anyone can spin up an LLM-powered app in a weekend, technology is no longer the bottleneck—understanding and servicing the problem the right away is.
Consider Harvey AI, a standout in the legal tech space. Their team isn’t just building AI. They’re solving law firms’ thorniest problems with a precision that only insiders can muster. Harvey AI co-founder Winston Weinberg was a securities and antitrust litigator at O'Melveny & Myers. Harvey AI’s team includes lawyers who’ve endured law school, passed the grueling bar exams, and toggled between grunt work and high-stakes cases. They know the bottlenecks: endless document reviews, clunky research processes, and the inertia of manual workflows. Harvey’s CEO and head of legal research are lawyers who “speak the lingo” and empathize with their customers. Their product and AI teams embed legal experts alongside engineers, ensuring the tech doesn’t just work—it wows lawyers, a notoriously tough crowd to impress.
A quote from their team captures it perfectly:
What Harvey does really well is embedding legal expertise across all functions… Our CEO is a lawyer, our head of legal research is a lawyer… They go to a firm, speak the language, and are super empathetic. On the product and AI side, we have lawyers working hand-in-hand with engineers. - Aatish Nayak, Head of Product at Harvey AI
This isn’t just about data or algorithms. It’s about lived experience. A generic AI tool couldn’t crack the legal market without understanding how lawyers think, work, and resist change. Harvey’s moat isn’t their tech stack. It’s their ability to meet customers where they are and deliver value that sticks.
Consider Forest Flager, a construction industry veteran with deep roots in sustainable design, who saw firsthand the clunky inefficiencies plaguing construction material sourcing. The process was a mess, with emails ping-ponging between distributors and buyers, one-off spreadsheets, and critical data and context vanishing in the shuffle, all while eating up time with bespoke, disjointed workflows. Flager’s first-hand experience laid bare the vast potential for improvement, leading him to launch Parspec. Leveraging AI, Parspec digitizes and streamlines this tangled process, empowering construction buyers to source cost-effective, sustainable materials with ease.
Conclusion: Domain Expertise Wins
AI-driven software doesn’t just compete with other tools—it battles (entrenched) human workflows. In enterprise software, legacy vendors offer a clear benchmark to beat. But when AI replaces a manual process—like a lawyer’s research or a doctor’s chart review—the challenge is stickier. Success demands four things:
Minimize Friction: Adoption hinges on seamless integration into existing habits. No one rewrites their day for a shiny new tool.
Provide Proof: AI must prove it’s reliable and worth the leap, delivering a wow every time.
Enable Scalability: Success relies on the product’s ability to grow with its users, handling increased complexity and volume without breaking.
Cultivate Stickiness: Beyond initial adoption, long-term success requires users coming back and continuously deriving value.
This is brutally hard. Founders who’ve lived the problem, who’ve been the customer, have an unfair edge. They see the pain points others miss, anticipate resistance, and design products that stick. That’s why I over-index on founder-market fit at the early stage.
The founders I back don’t have a hammer looking for a nail. They’re building companies in domains where they have extreme expertise. They are doing their life’s work.