A wrapper, or AI wrapper, is the product built on top of an AI model. The model is the engine. The wrapper is the truck built around it, and the truck is the part you drive. Most of the AI tools on the web right now are wrappers sitting on a handful of existing models. Almost everything you touch when you use one belongs to the wrapper, and almost none of it is the model.
Is ChatGPT a wrapper?
Yes, and it is the best known one there is. ChatGPT is OpenAI’s wrapper product, built on top of their GPT-5 model.
Look at what is on screen. The interface, the buttons, the layout, the input field, the chats, the text as it lands. All wrapper. Underneath that sits data and integrations with other platforms. And in front of your very first message sits a system prompt, a block of instruction OpenAI inserts before anything you type, so the model answers the way they want it to answer.
AI wrapper examples: Cursor, Perplexity and Jasper
Cursor is a wrapper for writing code, and it is model agnostic. It runs on GPT-5. It runs on Claude’s models, Haiku, Opus and Sonnet, and switches between them depending on the use case. Perplexity is a wrapper for search, an LLM that goes out and looks things up on the web. That sounds basic today. Two years ago when they launched, it was new. Jasper is a wrapper too.
A wrapper exists to give the model better direction and get a better output on one specific task. Cursor could have kept it simple: take your codebase, hand it to a model, say go fix this bug. They added depth instead. Examples of how similar issues were fixed, extra layers of instruction, so the model gets a precise task rather than a generic prompt with a generic amount of information in it.
Thin wrapper vs thick wrapper
Thickness is what separates one AI company from the next.
| Wrapper | Thin | Thick |
|---|---|---|
| What it looks like | A chatbot with one or two features | Often not a chatbot at all |
| Where the work happens | In the prompt you type | In the background, before the model is called |
| What it knows | Whatever you paste into it | Your data and your systems |
| What it does | Answers | Runs skills and workflows |
Four things sit in that background layer, and you never see any of them: the data flowing through it, the integrations into the platforms you already run on, your own IP, and the skills and workflows set on top so the model can do differentiated work instead of answering questions.
Claude Code is one of the better examples of a thick wrapper on a big lab’s model. It lives in the terminal. It has access to your entire file system and codebase, and a lot of functionality packed in behind it that surfaces the more you use the tool.
Should you buy an AI wrapper or build one
It comes down to your requirements. Standard workflow, and someone has already built the wrapper for it. If your workflow is unusual, tailored to how your business actually runs, you will not find one that fits.
That is why a lot of owners end up building their own internal AI tools. It is most of the consulting and implementation work we do, and almost none of them want a thin one. A chatbot that answers questions does little when the real work is sitting in the ERP and the CRM. They want it wired into those systems and doing things in the middle of the night: the after-hours intake call for an HVAC company, the lease renewal a property manager would otherwise chase on Monday morning.
An AI agent, by the way, is a wrapper in itself.
So take the question in that order. Is there a wrapper on the market that meets my requirements? If there is not, build. Building does not have to mean big, either. Pointed at one job you actually do, even a thin wrapper gets it done.
Implementation is its own job, and it is what comes next in the track: what an agent is, what a tool call does, and how both get wired into the systems you already run.
