1. If OpenAI Ships a Competing Feature, What's Left of Your Business?
This is the existential question. A startup's value can be reduced to a single API parameter overnight. A good answer involves moats that OpenAI can't or won't build, such as deep, industry-specific workflow integrations or proprietary data loops that the foundation
model can't access. A bad answer is, 'We'll move faster.'
2. What is Your Unique Data Flywheel?
Many 'AI' companies are just renting intelligence. Truly defensible businesses have a data flywheel: the act of using the product generates unique data that, in turn, improves the model or the user experience. Ask for proof. Does customer A's usage actually make the product better for customer B? If not, there is no data moat.
3. How Deep is Your Workflow Integration?
A Chrome extension is a feature; a system of record is a business. The harder a product is to rip out, the stickier it becomes. Look for startups that don't just augment a workflow but become the authoring layer where work begins. If the product is the place where essential data is created and managed, it's far more defensible than a simple add-on.
4. What's Your Gross Margin, and How Will It Scale?
Compute cost is the silent killer of AI wrappers. If a startup's gross margin is 40% because they're paying for every single API call, they don't have a business; they have a resale operation with venture-funded losses. A strong answer will show a clear path to 70%+ gross margins, either by shifting to open-source models for certain tasks, negotiating better rates, or having a pricing model not solely tied to usage.
5. Why Can't a Competitor Replicate This with a Different Model?
With a growing number of powerful open-source and proprietary models available, being 'model-agnostic' isn't a strategy by itself. The defensibility must come from something other than the specific model being used. This could be a unique user experience, a distribution channel, or a brand built on trust within a specific niche.
6. How Do You Solve for the 'Last Mile' Problem?
Foundation models are great at getting things 80-90% right. The real value—and the defensible business—is often in solving the final 10-20%. This requires domain-specific logic, human-in-the-loop processes, or complex integrations that a generic model can't handle. Ask how they ensure 100% accuracy where it matters.
7. What Is Your Go-To-Market Advantage?
In a world where technology is increasingly accessible, an effective go-to-market strategy becomes a moat. A startup that has exclusive access to a specific customer base, a viral adoption loop, or a deep partnership with a major distributor has a significant advantage that is difficult for a large company like OpenAI to replicate quickly.
8. How Do You Handle Hallucinations and Liability?
For any application in a regulated or mission-critical field like healthcare, finance, or law, this is non-negotiable. What is the protocol when the model produces inaccurate or harmful output? Who is liable? A strong founding team will have a clear, auditable process for validation, error-checking, and transparency with customers.
9. Does the Founding Team Have Unique Domain Expertise?
When the tech is a commodity, the team's insight is the differentiator. A team of former accountants building a tool for financial analysis will understand nuance, build trust, and navigate regulatory hurdles better than a team of generalist engineers. This 'founder-market fit' is a powerful, often overlooked, moat.
10. What's Your Plan for When API Access or Pricing Changes?
Platform risk is not a hypothetical. OpenAI and other providers can and will change the rules of the game, from pricing structures to API access. A resilient startup has a plan B. This might involve abstracting the LLM layer to easily swap providers or strategically using self-hosted, open-source models for core functions.
11. Is This a Feature or a Company?
Many AI startups are building features that should be part of a larger product. A tool that summarizes PDFs is a feature; a platform that manages the entire lifecycle of legal contracts is a company. VCs must be disciplined in distinguishing between a cool demo and a sustainable business that solves a large, painful problem.
12. What Happens to Your Customer's Data?
Data privacy and security are paramount. Are customer prompts or data being used to train the foundation model? Even if the current policy is 'no,' those policies can change. A startup that handles sensitive information must demonstrate a robust architecture for data isolation and prove they aren't inadvertently leaking their customers' proprietary information to a third-party model.
