Is an AI Startup Defensible Without Owning a Model?

A startup can build a valuable AI product on a model it doesn't own. For an investor, the question is whether customers would still choose that product after a rival gets access to the same model. The answer sits in the work the product completes, the context it can lawfully use, and the evidence that customers stay.

Treat model access as a dependency

Ask which model providers the product uses, which tasks depend on each provider, and what happens to quality and gross margin when prices or terms change. Run the same representative tasks through a credible alternative model. A startup doesn't need every model to perform equally well, but it should know where switching breaks a workflow and how much repair that takes.

In its 2025 interviews with enterprise technology leaders, Andreessen Horowitz reported that multi-step AI workflows can make model changes harder because downstream steps depend on earlier outputs. That is a reason to inspect the system around the model, not to treat provider dependence as a moat by itself.

Inspect the whole workflow

Follow a real customer task from intake to final action. For a contract review product, that means receiving a document, finding the right policy, flagging an exception, getting approval, and recording the decision. A polished answer in a demo covers only part of that job.

Ask the founder to show permissions, integrations, exception handling, audit records, and the work a person still does. Which steps would a customer have to rebuild if they switched suppliers? An interface that sits inside daily work may earn loyalty; the claim needs evidence from customer behaviour. Andreessen Horowitz's enterprise AI builder interviews describe workflow adoption as a possible source of switching costs.

Test the data advantage

“We have proprietary data” is a starting claim. Ask who owns the records, whether contracts permit their use, whether customers can take them elsewhere, and how those records improve a named task. Then compare the product's results with and without that context on the same cases.

Look at the feedback loop too. Does the team capture expert corrections, label failure types, and retest after a product change? OpenAI's guide to business evaluations explains how task-specific tests and reviewed production outcomes can build a context-specific dataset. The investment case is stronger when the company can show the quality improvement in its own workflow, with a baseline and a repeatable test.

Look for distribution and retention

Distribution can be as hard to copy as software. Find out how the company reaches its buyer, who signs the contract, how long setup takes, and what customers do after the pilot. Review usage by account and by team, renewals, expansion, and the reasons customers stop using the product. Separate paid adoption from free trials and founder-led demonstrations.

Ask a customer what they would use if this product disappeared tomorrow. A spreadsheet, an incumbent tool, a general-purpose assistant, and a competing startup each imply a different level of dependence. The customer's answer is more informative than a slide about “lock-in.”

Run the 90-day copy test

Give a capable rival the same public models and 90 days. List what they could reproduce: the visible interface, core prompts, public integrations, and marketing claims. Then list what would take longer: permission to use customer data, tested handling of rare cases, embedded approvals, trusted relationships, and a sales route into the buyer. Verify each item with a product demonstration, customer reference, contract, or operating metric.

Invest if the startup solves a costly job better than the alternatives and has evidence that its advantage grows with use. Be cautious when the only proof is a strong model response, a large document store, or customers who haven't moved beyond a pilot. Owning a model is neither a requirement nor a substitute for those checks.