There is a specific kind of company that emerged in 2023 and has been multiplying ever since. It has a polished landing page, a waitlist, a founder thread on Twitter with 47 slides, and exactly one API call to OpenAI underneath all of it. Welcome to the age of the AI wrapper.

The term gets thrown around loosely, but the core accusation is precise: a thin wrapper takes a user’s input, passes it to a foundation model with a prompt template, and returns the output. The company adds a login screen, a pricing tier, and maybe a Slack integration. That is it. The product is a UI around someone else’s intelligence.

Most of these companies will die. That is not a controversial prediction. But the interesting question is not whether thin wrappers are doomed, it is where the line sits between a wrapper that adds nothing and a product that happens to use AI. Because that line matters, and most people drawing it are getting it wrong in both directions.

The case against wrappers is real

When GPT-4o or Claude can be accessed directly, a wrapper that adds a textarea and a submit button is offering negative value. It introduces latency, another account to manage, and a markup on tokens you could buy yourself. The moat is zero. The switching cost is zero. The moment the underlying model provider ships a comparable feature natively, the wrapper is finished.

This has already happened repeatedly. A dozen startups built “AI writing assistants” in early 2023. Google shipped Gemini in Docs. Microsoft shipped Copilot in Word. The startups evaporated. The same pattern hit code completion, image generation, summarisation, and meeting notes. If your entire value proposition is “we call an API and show you the result,” you are not building a company, you are building a feature request for someone who already has distribution.

There is also a trust problem. Wrappers obscure which model is running, what the prompt says, and how data is handled. Users hand over sensitive information to a product that adds a logo between them and the actual intelligence. When the wrapper company is three people in a flat in Berlin with no SOC 2 report, that is a risk most buyers have not fully reckoned with.

But some wrappers are legitimate products

Here is where the lazy criticism breaks down. Not every product that wraps an API is a thin wrapper. The test is not whether AI is involved, it is whether the company would still be useful if the AI layer were removed or replaced.

Consider a legal contract review tool. It uses an LLM to flag risky clauses, but it also maintains a clause library built by actual lawyers, integrates with a document management system, tracks redline history across versions, and routes flagged items to the right reviewer based on contract type. Remove the AI and the workflow still exists, it is just slower. The AI makes the product faster and more consistent, but the product’s value is in the workflow, the data model, and the domain-specific logic surrounding the AI call.

That is the distinction. A thin wrapper is an API call with a face. A real product is a system where AI is one component among several, and the system would still function, poorly, without it.

The three types

After watching this space for two years, I see three categories:

Type 1: The prompt reseller. Takes your input, adds a system prompt, calls GPT, returns the output. Sometimes adds a file upload. The entire technical stack is an HTTP request. These are feature demos masquerading as companies. They will be absorbed into the model providers or die.

Type 2: The workflow integrator. Uses AI as a component inside a larger process. Stores results, connects to other tools, builds domain-specific data over time. The AI is important but not sufficient. These companies have a chance if the rest of their product is strong enough to survive model commoditisation.

Type 3: The model modifier. Fine-tunes, retrains, or builds proprietary models on domain-specific data. The AI itself is the product, but it is a different AI than what you get off the shelf. These are hard to build, expensive to maintain, and defensible when done well.

A framework for telling them apart

When evaluating whether a product is a thin wrapper or something real, ask these five questions:

1. What breaks if the API goes down? If the answer is “everything,” you are looking at a wrapper. If the answer is “a feature degrades but the product still works,” you are looking at a product.

2. Does the product get better with use? Thin wrappers are stateless. Real products accumulate data, learn patterns, and improve over time, not because the model improves, but because the system builds context the raw model cannot access.

3. Can you replace the model? If swapping GPT-4o for Claude produces identical results, there is no proprietary intelligence. The product has no moat in the AI layer.

4. Is there domain expertise encoded somewhere? Prompts, fine-tuning data, evaluation sets, custom tooling, training pipelines, or hand-built knowledge bases that took real effort to create. If there is nothing domain-specific, there is nothing defensible.

5. Would an expert pay for this over using the raw model? If a lawyer, doctor, or engineer would rather type into ChatGPT than use your product, your product is not solving a problem the model already solves. You are adding friction, not value.

The AI wrapper era will produce a graveyard of landing pages and a handful of real companies. The difference will not be in the AI, it will be in everything around it. Build the system. Not the shell.