AI Wrapper App
An AI wrapper app is a program that has no AI of its own, but instead uses the AI of a large provider over the internet and builds its own interface around it. The debate about such apps revolves around the question of whether a pretty shell around someone else's technology is a viable business.
Many programs sold today as AI products don’t actually have any AI built in. They send the user’s input over the internet to a large provider such as OpenAI, Google, or Anthropic. There, a huge language model does the computing — that is, a program that has learned from texts and generates fitting answers. The answer comes back and is displayed in the app’s own interface. Such programs are called wrapper apps, from the English “wrapper,” meaning shell or packaging. The actual engine belongs to someone else; the packaging belongs to the app’s provider.
The dispute over the thin shell
The term is often used disparagingly. When someone says “that’s just a wrapper,” they mean there’s no real achievement of its own behind it. Indeed, a simple wrapper app can be built in a weekend. That also means anyone else can rebuild it in a weekend. In economics, this is called a lack of a barrier to entry.
There’s also a second risk. The large provider can offer what the wrapper app does at any time itself. When OpenAI added a feature for analyzing uploaded PDF files, several small providers lost their purpose overnight. The model can also become more expensive, change its rules, or be shut down entirely. The app thus hangs on a leash held by someone else.
Still, the contempt isn’t always justified. An online shop, too, doesn’t have its own payment system or its own postal service. What matters is whether something emerges around the foreign model that is hard to copy. That can be proprietary data, a workflow very precisely tailored to a profession, or simply many satisfied customers.
What happens between input and model
Technically, the core runs through an interface, in English an API. This is a fixed address on the internet to which a program sends requests and from which it receives answers back. The wrapper app needs an access key for this and pays by volume. Billing is usually per token, meaning per word fragment that the model reads and writes.
Before sending, the app restructures the request. It prepends a fixed instruction, the so-called system prompt. This sets role, tone, and boundaries, for example: “You are an assistant for tenancy law and answer in simple sentences.” The app often also attaches text excerpts from its own documents so the model works with the appropriate knowledge. The user sees none of this.
After the answer, the work continues. The app checks the result, formats it, saves it, and connects it to other programs. This is exactly where the difference between a weak and a strong wrapper lies. A weak one merely passes the question through. A strong one manages documents, knows the previous conversation history, and hands results directly over to accounting or the ordering system.
Wrappers in everyday life and in quarterly earnings
A large share of the AI tools found online are wrappers. These include writing aids, translators, learning programs, application generators, and chatbots on company websites. Many features in well-known software work this way too. Anyone who has a text summarized in their browser is often indirectly talking to a foreign provider’s model.
In stock market and startup news, the term regularly appears as a warning sign. Investors ask young AI companies how much proprietary value exists on top of the foreign model. One indicator of this is gross margin: whoever passes on a large share of every euro of revenue to the model provider earns little. Wrappers should not be confused with companies that train their own models. Their costs are on an entirely different scale, but in return they own the technology itself.