Assessing AI Business Models

Hello! Welcome to the first lesson in our module on conducting due diligence for AI startups.

In the previous module, you mastered the general framework for due diligence and learned how to synthesize your findings into a compelling investment memo. Now, we'll sharpen that framework for the unique challenges and opportunities presented by AI-native companies. Your goal to launch an AI-focused incubator makes this a critical area of expertise.

This lesson addresses the learning outcome: Assess the technical feasibility and defensibility of AI-native business models.

We'll move beyond traditional software metrics to answer two fundamental questions:

1. Feasibility: Can the team actually build and scale their claimed AI solution reliably and economically?
2. Defensibility: If they succeed, what stops a competitor—especially a tech giant—from immediately replicating their product and capturing the market?

This analysis is the first step in separating genuine, durable AI businesses from the plethora of superficial "AI wrappers" that have emerged.

1. Technical Feasibility: Beyond the "Wow" Demo

In the age of powerful foundation models, creating an impressive-looking demo is easier than ever. As an investor, your job is to look past the initial "wow" factor and assess the practical, real-world viability of the technology. This involves understanding the journey from a prototype to a production-grade system that customers can rely on.

Jake Heller, the founder of the AI legal tech company Casetext (acquired by Thomson Reuters for $650M), provides an exceptional founder's perspective on what it truly takes to build a reliable AI product.

As Heller explains, true technical feasibility isn't about having access to the latest model; it's about the painstaking work required to make that model perform a specific task reliably. Here are the key diligence questions that emerge from his insights:

• Deep Domain Expertise: Does the team have an almost obsessive understanding of the workflow they are automating? Can they break it down into micro-tasks? An AI team that can't map out the nuanced, real-world steps of a professional's job will struggle to build a product that's more than a toy. This is where your experience in business model validation is a huge asset.
• The Path to Reliability (Evals): Does the team have a rigorous evaluation framework? This is a critical diligence artifact to ask for.
  • A team that says, "We're 90% accurate," should be able to show you the test suite (evals) that proves it.
  • How did they create their "gold standard" answers for testing?
  • How many test cases do they have? (Heller suggests at least 100 to get to beta).
  • A team that can't produce a robust evaluation framework is likely still at the demo stage.

Economic Feasibility: The Unit Cost of Intelligence

A unique aspect of AI business models is that, unlike traditional SaaS, there's often a meaningful variable cost associated with every action a user takes. This is the inference cost—the computational price of running the model to generate a result.

A product can be technically brilliant but economically unviable if its inference costs are too high, especially as it scales. An investor, Ben Pouladian, created a diligence framework that highlights this.

Your background in computer science will help you appreciate the importance of these questions. When evaluating a startup, you need to assess:

• Are they just making API calls to an expensive model like GPT-4, or have they engineered a more efficient solution (e.g., using smaller, specialized models, fine-tuning, caching)?
• How do their pricing model and unit economics account for these costs? A startup that hasn't modeled this is at high risk of seeing its margins collapse as it gains traction.

2. Defensibility: Building Moats in the AI Era

Once you've established that a startup can build its product, the next question is whether they can defend it. The competitive landscape for AI is fierce, with large, well-funded labs and incumbents able to move quickly. Traditional software moats are still relevant, but AI introduces new dynamics.

The following video from Y Combinator provides a great overview of the "seven powers" framework applied to modern AI startups.

Let's synthesize the most critical AI moats discussed in the video and other resources.

Key AI Moats

1. Process Power & Speed: As the YC video emphasizes, the most important initial moat is speed of execution. In the early days, the team that builds, learns, and iterates fastest wins. This evolves into "Process Power," which is the defensibility that comes from building a complex, deeply integrated system that solves the "last 10%" of the problem. It’s the result of the relentless evaluation and prompting work Jake Heller described. It's hard to replicate because it's the sum of thousands of small, hard-won engineering and product decisions.

2. Data & Data Network Effects: This is the most-cited AI moat. A data network effect occurs when the product gets smarter and more valuable as more people use it, because their usage generates data that improves the model for everyone.

   • Diligence Question: Is the data truly proprietary and generated by the core use of the product? Data scraped from the public web is a weak moat. Data from exclusive partnerships or, even better, unique data generated from user interactions within the product, is a strong moat.
   • NFX, a seed-stage VC firm, provides simple tests to evaluate this. When assessing a product, ask:
     • The Switching Cost Test: If I stop using this product, what do I lose? A weak answer is "I'll use a similar tool." A strong answer is "I'll lose months of accumulated context and workflows."
     • The Collaborative Value Test: Is this product more valuable when others on my team use it?

3. Counter-positioning: This is a strategic moat where a startup adopts a business model that incumbents cannot or will not copy because it would cannibalize their existing business.

   • Example: An AI customer support agent priced per resolved ticket is counter-positioned against an incumbent like Zendesk that charges per human agent seat. If the AI is successful, it reduces the number of human agents, directly threatening the incumbent's revenue model. This gives the startup air cover to grow.

4. Embedding & Switching Costs: This is particularly powerful for B2B AI companies. By deeply integrating into a customer's critical workflow (e.g., becoming the core engine for a bank's loan origination process), the AI product becomes incredibly sticky. The operational cost and risk of ripping it out and replacing it become prohibitively high, creating a powerful lock-in effect.

3. A Unified Framework for Assessment

Different investors use different frameworks, but they all point to the same core principles. The "5D" framework from Sapphire Ventures provides another useful lens.

As you build your own diligence process, you can use these frameworks as a starting point for your checklist. The goal is to develop a systematic way to probe the feasibility and defensibility of every AI startup you evaluate.

Test your understanding!

You are evaluating an AI startup that has built an AI-powered code generation tool specifically for large enterprise COBOL systems (an old but critical mainframe language). Their tool helps big banks and insurance companies migrate their legacy systems to modern languages.

• They have exclusive pilot programs with two of the world's top five banks.
• The product gets better at generating correct code as it processes more of a specific bank's proprietary COBOL codebase.
• They are priced based on the number of lines of code successfully migrated, not on a per-developer-seat basis.
• The founders are a mix of veteran COBOL engineers and young AI researchers.

Based on the moats discussed in this lesson, what are the startup's two strongest potential sources of defensibility? Explain your reasoning for each.

Show answer

Here are the two strongest potential moats:

1. Data Network Effect & Cornered Resource: The strongest moat is the proprietary data loop. By working with large banks, they gain access to massive, non-public COBOL codebases—a "cornered resource." As their tool processes this code, it learns the unique patterns, dependencies, and styles of each enterprise system. This improves the model's accuracy for that customer and likely for future customers with similar systems. A competitor without access to this specific, high-value data cannot replicate the performance, even with a better general-purpose model.

2. Counter-positioning: The startup's business model is a powerful strategic moat. Incumbent developer tool companies (like Microsoft or IBM) often sell on a per-seat subscription basis. This startup is priced on outcomes (migrated code). An incumbent would find it difficult to adopt this model without cannibalizing their existing, predictable subscription revenue from their developer tools division. This gives the startup a unique value proposition and protects it from direct competition from established players.

The deep Domain Expertise of the team (veteran COBOL engineers) also contributes significantly to their "Process Power," allowing them to solve the complex edge cases of COBOL migration that a generic AI team would miss.

Conclusion

Assessing AI-native business models requires a new lens. You've learned that you must look beyond the demo to verify true technical and economic feasibility, and you must understand the new and evolving sources of defensibility that separate enduring companies from fleeting features.

Key Takeaways:

• Feasibility is Reliability at Scale: A feasible AI product is not just one that works sometimes; it's one that works reliably in the messy real world, backed by deep domain expertise and a rigorous evaluation framework. Its unit economics must also hold up as usage grows.
• AI Moats are Multi-faceted: Defensibility isn't just about one thing. It's a combination of proprietary data loops (network effects), deep workflow integration (embedding), superior execution (process power), and clever business models (counter-positioning).
• Ask the Hard Questions: Your job as an investor is to be the disciplined, skeptical voice that asks: "How do you know it works?" and "What happens when Google launches a similar feature?"

Preview of the Next Lesson

In this lesson, we focused on the characteristics of a strong AI business model. In the next lesson, we'll focus on identifying the weak ones. We will tackle the learning outcome: Distinguish between genuine AI innovation and superficial 'AI-washing'. You'll learn the red flags and tell-tale signs that a company is simply wrapping a standard API call in a nice user interface and calling it revolutionary AI.