What is the Defensible Moat for an AI Startup

As an occasional angel investor I’ve had the privilege of listening to a few startups recently who are looking to raise money, many of whom appear to have vibe coded their MVP.

Nothing necessarily wrong with that, but 'how defensible is the IP' is the first question I ask after the pitch

This was always the question I used to get asked when fundraising for my own startup back in the day, but with where we find ourselves now with AI, its even more pertinent.

I’ve written before about how easy it is to replicate an entire app just by observing the outputs . You don’t need to decompile anything. You don’t need to steal source. In many cases, you just need:

➡️ The same foundation models everyone else has
➡️ Some prompt hacking
➡️ A half-decent engineer with time on their hands

So IP defensibility is not one of those obtuse investor questions, the investor is trying to understand:

➡️ What do you have that can’t be copied

➡️ If an AI-native competitor shows up tomorrow, what do you still have that they don’t?

➡️ Is your “product” just leveraging the coding capabilities of OpenAI/Anthropic/Gemini

So what does defensible IP or a startup moat look like ?

✅ Proprietary data: Unique, hard-to-access, or hard-to-clean data that meaningfully improves your models or workflows.

✅ Deep integration: Being wired into customers’ systems and processes in a way that’s painful to rip out.

✅ Compound workflows: Not just a chat box, but end-to-end automation that spans multiple tools, teams, and edge cases.

✅ Domain expertise baked into the product: Years of domain knowledge encoded as guardrails, playbooks, and system prompts that is hard to replicate.

'How Defensible is your IP' should not be a 'rabbit in headlights' moment, its one of those questions founders should be prep'd and comfortable with answering post their pitch.