The wrapper question: what seed investors ask an AI startup about defensibility in 2026

Sooner or later in a seed meeting for an AI product, someone asks it. Sometimes politely ("what stops OpenAI from building this?"), sometimes not ("isn't this just a wrapper?"). Founders tend to hear it as an insult and answer with a defence of their prompts. That is the answer that loses the room.

The question is fair, and it has a good answer for most products that deserve funding. It just isn't about the model. This piece covers where the question came from, what investors who have written about it publicly actually think, the four answers that land, and the evidence a founder with a few dozen customers can put on one slide.

Where the question came from

In January 2023, Andreessen Horowitz published a long essay on who would make money in generative AI. Two lines from it set the tone for every pitch meeting since: "applications lack strong product differentiation because they use similar models", and "There don't appear, today, to be any systemic moats in generative AI." If everyone calls the same API, the reasoning goes, anyone can build what you built.

Then the market supplied an example. Jasper, an AI writing tool built on OpenAI's models, raised at a $1.5 billion valuation in October 2022, weeks before ChatGPT launched. By mid-2023 it had cut its 2023 revenue forecast by at least 30%, laid off staff and cut its internal valuation by about 20%. Its customers had found the same capability in a free chat window. Every investor who watched that now asks the question early.

Sequoia's "Generative AI's Act Two" (September 2023) added the number that still sits behind it: the median generative AI app had 14% of its monthly users active on a given day, against 60 to 65% for the best consumer apps. People tried AI products and drifted away. The same essay also gave the first useful answer: "The moats are in the customers, not the data," and "workflows and user networks seem to be creating more durable sources of competitive advantage."

Why it is still asked in 2026, and why it is no longer a verdict

Three things keep the question alive. Models improve every few months, so a gap you built with clever prompting closes when the next model ships. The model vendors themselves keep moving up into applications: agents, coding tools, document features, browsers. And building software is now cheap, which cuts both ways: you built your first version in weeks, and so can a competitor. Andrew Chen, a general partner at a16z, put that last risk plainly in February 2025: AI could make cloning a product trivial, so the head start a new product enjoys may shrink.

But the same essay, titled "Revenge of the GPT Wrappers", argues the other side. If the models are commoditising, the advantage moves to the things software companies have always defended themselves with: distribution, network effects, and in business software, "workflow, compliance, security", brand and proprietary data. Cursor is the example everyone reaches for. It started as an editor built on other companies' models, and in November 2025 raised $2.3 billion at a $29.3 billion valuation with more than $1 billion in annualised revenue. Nobody calls it a wrapper now; they call it a product people open every morning.

Y Combinator's partners have made the same point from the other direction. In their Lightcone episode "Vertical AI Agents Could Be 10X Bigger Than SaaS" (November 2024) the argument is that the winners take over a whole job in one industry, not one step of it, because that is something a general chatbot cannot do from a chat window.

The question is not whether you call someone else's model. It is what your customer would lose if they left.

The four answers that land

Each of these is a claim. What makes it land is a number or a fact behind it, which is the next section. In rough order of how convincing they are at seed:

- Workflow. The product does a whole job inside the place the work already happens: it reads the shared inbox, files the result in the customer's system, routes approvals, keeps the audit trail. A chat window can produce the same paragraph; it cannot run the process around it. The stronger version: you replace a task someone was paid to do, so you are compared with a salary line, not a software line.
- Switching cost. After three months, the customer's configuration, history, templates, integrations and team habits live in your product. Leaving means rebuilding them. This is the most boring moat and the one acquirers pay for.
- Distribution. You reach your buyers more cheaply than a well-funded rival can: an owned audience in a niche, a partner that resells you, a marketplace listing that ranks, a community you run. The model vendor has a better model; it does not have your list of 4,000 veterinary practice managers.
- Data, the honest kind. Not "we have proprietary data" meaning a scraped corpus. The data that counts is what only your product produces as it is used: which drafts customers edited and how, which suggestions they rejected, which outcomes followed. Sequoia's warning stands: the next model may wipe out a static data advantage. A feedback loop that makes your product better for this customer each week is different.

Two more that help, though rarely on their own: speed (you ship a fix for a customer this week; a platform ships for millions next quarter) and cost control (you can change models, cache and route without the customer noticing, which keeps your margin when prices move).

Answers that lose the room

- "Our prompts are our secret sauce." Prompts leak, can be rewritten in an afternoon, and break when the model changes.
- "We fine-tuned our own model." Sometimes useful, rarely a moat at seed; the next general model often matches it.
- "We are first." Jasper was first.
- "We will train our own foundation model once we raise." This tells the investor you plan to spend their money competing with companies that have spent billions.
- "OpenAI would never bother with our niche." They may not. Someone with your same API key and a free weekend might.

What a small founder can show

A seed investor does not expect a moat to be finished. They want to see that it is forming. With a few dozen paying customers, these numbers tell that story better than any adjective:

The last row is the one founders skip and investors increasingly ask about. If your product only works on one model from one vendor, you have a supplier risk as well as a moat question. Showing that you test the job on more than one model, as we described in fifty examples and a spreadsheet, answers both. Showing what the model bill does to your gross margin, as in when the AI feature eats the margin, answers the follow-up before it is asked.

Then answer the question head on, in one sentence, before they ask it. Something like: "If OpenAI shipped our feature tomorrow, our customers would still have their approval rules, their eighteen months of case history and their integration with the practice software in us, and they would have to rebuild all of that to switch." If you cannot write that sentence yet, the meeting is not the problem.

A test you can run before the meeting

Take your five best customers and list, for each, what they would lose if they cancelled today. Not features: their stuff. Settings, history, connected systems, the people in their team who are trained on it, the outputs they rely on every week. If the honest list for most of them is "a few saved prompts", you have a wrapper, and the next three months of roadmap should go into the workflow around the model rather than the model call itself. If the list is long, put it on the slide.

The investor mindset behind all this is the same one we summarised in how investors see internet startups in 2026: fewer, more efficient companies, judged on retention and revenue per head. The wrapper question is that mindset applied to AI. Answer it with customers, not with the model.

Sources

a16z, Who Owns the Generative AI Platform? (Bornstein, Appenzeller, Casado, 19 January 2023; quotes on differentiation and systemic moats); Sequoia Capital, Generative AI's Act Two (Huang and Grady, September 2023; 14% median DAU/MAU, "the moats are in the customers, not the data"); Andrew Chen, Revenge of the GPT Wrappers: Defensibility in a world of commoditized AI models (4 February 2025); Maginative, Jasper appoints new CEO and cuts internal valuation as AI growth slows (2023; $1.5 billion round, forecast cut, valuation cut, layoffs); Wikipedia, Anysphere (November 2025 round, $29.3 billion valuation, revenue above $1 billion annualised); Virtualization Review, The Future of AI: Follow the Money (on Y Combinator's Lightcone episode "Vertical AI Agents Could Be 10X Bigger Than SaaS", November 2024). Pages read on 17 October 2026. The evidence table and the example sentence are our own suggestions, not a list any investor has published.