What is an AI wrapper, and when is it a real product?

What an AI wrapper actually is

An AI wrapper is a product built on top of a model somebody else trained. Instead of building its own system, the company sends prompts to a foundation model API from OpenAI, Anthropic or Google, adds an interface and some logic, and ships the result as an app. The name is a metaphor: the startup wraps a thin layer of its own code around intelligence it does not own. By that definition, most of the AI tools you use are wrappers of some kind, because training a frontier model costs hundreds of millions of dollars and almost no application company does it.

The word is rarely neutral. Packed inside "just a wrapper" is a claim that the layer is trivial. If the intelligence comes from GPT or Claude, and anyone can call the same API with the same few lines of code, then what exactly is the company? On that reading a wrapper is a prompt, a login screen and a payment button, and it can be cloned in a weekend, including by the model provider itself.

The "just a wrapper" critique, and why it stings

The reason the insult lands is that it points at a real risk: no moat. In an influential 2023 essay mapping the generative AI stack, the venture firm Andreessen Horowitz concluded that "there don't appear, today, to be any systemic moats in generative AI," and noted that application companies were growing revenue quickly but struggling with retention and gross margins. If your only differentiator is a prompt and a clean interface, a competitor can match it fast, and the model provider can fold the same feature into its own app the next time it ships.

There is a second version of the fear, which is platform risk. A wrapper's core capability is rented, so a price change, a rate limit or a new first-party feature from the provider can reshape the business overnight. Writing in Forbes in September 2025, the consultant Sajal Singh argued that low API costs and minimal technical barriers make most thin wrappers unsustainable, and predicted that the large majority will fail as platforms absorb their best features. That is the bear case, and for genuinely thin products it is largely correct.

When a wrapper is actually a real product

Here is where the dismissal turns lazy. Being a wrapper describes how a product is built, not whether it is valuable. Plenty of durable software wraps something it did not invent; a bank's app wraps a payments network it does not own. The real question is what a company adds around the model, and whether that layer is hard to copy.

Four things tend to turn a wrapper into a defensible product. The first is proprietary data: usage that generates a dataset the provider does not have, which can feed evaluation, fine-tuning or a feedback loop that improves the product the more it is used. The second is workflow: when the tool becomes the place where work actually happens and the system of record for it, switching costs appear. The third is distribution: an existing audience, sales motion or integration that puts the product in front of users the model provider cannot easily reach. The fourth is user experience and trust in a specific domain, such as law or medicine, where getting the details reliably right matters more than raw model horsepower.

The clean example is GitHub Copilot. It is, in one sense, just GPT in an editor, but the integration into the developer's workflow creates real switching costs. The model is borrowed; the product around it is not.

The defensibility debate

So the honest debate is not wrapper versus not-wrapper. It is thin versus thick. As one product-strategy analysis from the software firm HatchWorks put it, "Models are a commodity. Strategy is the only moat." Foundation models are converging and getting steadily cheaper, a trend the economics of AI pricing make plain, so the value that once sat in the model is migrating up into the application, into the data it collects, the workflows it owns, the distribution it holds and the trust it earns.

None of this is automatic. A wrapper that never accumulates any of those assets really is exposed, and the churn and margin problems Andreessen Horowitz flagged are real. But a company that uses cheap, commodity intelligence to own a workflow and build a dataset is doing exactly what good software companies have always done. The word wrapper describes its plumbing, not its ceiling.

How to tell a thin wrapper from a thick one

There is a single test a founder should be able to pass: if the model provider shipped your exact feature tomorrow, why would customers stay? A thin wrapper has no answer. A thick one can point to something the provider cannot copy by pressing a button, the accumulated data, the integrations wired into a customer's systems, the distribution, the workflow a team has standardized on. If the only answer is "our prompts are better," that edge tends to evaporate as the underlying models improve, because the models improve for everyone at once.

Our take

"Just a wrapper" is a useful challenge and a bad conclusion. It is useful because it forces the right question: what do you own that the model provider does not? It is a bad conclusion when it stops there, because almost every application company is a wrapper in the literal sense, and some of them will be very large. The model is becoming a commodity input, like bandwidth or cloud compute, and the value is accruing to the companies that build a genuine product around it. Judge a wrapper by its data, its workflow and its distribution, not by the fact that it calls an API. Thin wrappers fail. Thick wrappers can be enormous companies, and the difference is a strategy question, not a technical one.