3
There is a class of AI company that looks like a real business from every angle — revenue, users, growth, press coverage — right up until the moment it isn’t one anymore. The founders are smart. The product is polished. The investors are reputable. And the entire thing is built on top of someone else’s API, which means the entire thing can be made irrelevant by a single platform update.
These are wrapper companies. And the uncomfortable truth about most of them is not that they’re bad. It’s that they’re temporary.
You can always add another newsletter that you don’t read. Besides this one.
- * *
A wrapper, in the simplest sense, is a product that takes an AI model’s API, adds a user interface and some prompt engineering on top, and sells access to the result. The term sounds dismissive, and sometimes it is used that way, but that’s not the point here. Wrappers were genuinely useful. When frontier models first became available through APIs, the raw interface was unusable for most people. Someone had to build the layer that made the technology accessible. That layer was the wrapper, and for a window of time, it was a legitimate business.
The problem is that windows close.
What made wrappers viable was a gap between what the model could do and what a normal person could get the model to do. The wrapper bridged that gap. It handled the prompt engineering, the context management, the output formatting, the workflow integration. The value was real.
But that gap is shrinking. It’s shrinking because the models are getting better at understanding what people want without elaborate prompting. It’s shrinking because the platform providers — OpenAI, Anthropic, Google — are building their own interfaces that handle the things wrappers used to handle. And it’s shrinking because every improvement to the base model makes the wrapper’s proprietary layer thinner.
A business built on a gap has to worry about what happens when the gap closes.
- * *
The standard defense of a wrapper company goes something like this: “Our moat is UX.” Or: “Our moat is our prompt engineering.” Or: “Our moat is that we understand the workflow better than anyone.”
Each of these sounds reasonable in a pitch meeting. None of them survive contact with the actual competitive landscape.
UX is not a moat. UX is table stakes. A better interface can be copied in weeks by any competent team. If the only thing separating your product from a competitor is that your buttons are in better places, you are in a race you will eventually lose to someone with more resources and fewer scruples about borrowing your design decisions. UX is what you compete on when you have nothing else to compete on.
Prompt engineering is a depreciating asset. The value of a carefully crafted prompt is inversely proportional to the capability of the model it’s talking to. As models get better at interpreting intent, the elaborate prompt chains that wrapper companies spent months building become unnecessary. What was once proprietary knowledge becomes a default behavior of the next model version. You can’t build a durable business on an asset whose value decreases every time your supplier ships an update.
Workflow knowledge is real but fragile. Understanding a specific industry’s workflow is the closest thing a wrapper company has to genuine differentiation. A legal AI tool that knows how lawyers actually review contracts has something. But “knowing the workflow” is not the same as “owning the workflow.” The moment a platform provider decides that legal document review is a large enough market, they can hire domain experts and build the same understanding into their own product. The wrapper company’s head start is measured in months, not years.
The pattern across all three defenses is the same: each one describes an advantage that is real today and shrinking tomorrow. A moat that gets narrower every quarter is not a moat. It’s a countdown.
- * *
Here’s what makes the wrapper problem structural rather than situational: the platform providers are not standing still. They are actively expanding what the platform does natively, and every expansion kills or wounds a category of wrappers.
This has already played out in public, repeatedly.
When OpenAI added Code Interpreter, an entire category of “AI data analysis” wrappers became redundant overnight. When function calling became a standard API feature, companies that had built proprietary tool-use systems on top of the models lost their primary differentiator. When vision capabilities shipped natively, products built around “upload an image and ask questions about it” went from novel to commoditized in a single release cycle.
Each of these was somebody’s startup. Each had users, revenue, maybe investors. And each was made obsolete not by a competitor, but by a feature addition from the company they were paying for API access.
The trajectory is clear. The platforms are moving upward through the stack, absorbing functionality that currently lives in the wrapper layer. File handling, memory, multi-turn conversation management, structured output, web browsing — features that were once the domain of third-party products are becoming native capabilities. The question for any wrapper company is not whether the platform will build what you’ve built, but when.
And the honest answer, for most of them, is soon.
- * *
There is a version of this argument that says none of this matters because the AI market is so large that there’s room for everyone. The pie is growing, so even a small slice is a big business.
This is the argument people make right before a market consolidates.
The size of the market is irrelevant if your position in it is indefensible. A growing market with no barriers to entry is just a larger arena in which to be commoditized. If anything, a bigger market attracts more competition and more attention from platform providers, which accelerates the exact dynamic that kills wrappers.
The companies that survive in a large, consolidating market are the ones with something the platforms can’t easily replicate. Everything else gets compressed to zero margin or absorbed entirely.
- * *
So what does real differentiation look like? If wrappers are temporary, what’s permanent — or at least durable enough to build a business on?
Proprietary data. If your product generates or accumulates data that makes it meaningfully better over time, and that data can’t be replicated by a competitor starting from scratch, you have something the platform can’t just build. This is why vertical AI companies that ingest industry-specific data — medical records, legal filings, financial transactions — have a fundamentally different risk profile than horizontal wrappers. The model is commodity. The data is not.
High switching costs. If your product is woven deeply enough into a customer’s workflow that ripping it out would be painful and expensive, you have durability that a wrapper never will. This isn’t about lock-in through dark patterns. It’s about building something so integrated and so relied upon that the customer’s operations would degrade without it. That kind of dependency takes years to build, which is exactly why it’s defensible.
Network effects. If your product gets better for each user as more users join — because of shared data, collaborative features, marketplace dynamics — you have a compounding advantage that scales with your user base. The platform provider can replicate your features. They can’t replicate your network.
Domain expertise that compounds. Not just “we know the workflow” but “we’ve been operating in this domain long enough that our product reflects thousands of edge cases, regulatory requirements, and user behaviors that a new entrant would take years to discover.” The difference between knowing a workflow and having lived through every way it breaks is the difference between a wrapper and a product.
The common thread is time. Each of these advantages takes time to build, which means they can’t be replicated quickly, which means they constitute an actual moat. A wrapper can be built in a weekend. A data asset, a deeply integrated product, a network effect — these take years. That’s the point.
- * *
If you’re building on top of an AI model’s API right now, none of this means you should stop. The wrapper phase is a valid starting point. Plenty of real companies started as wrappers and built something underneath before the window closed.
The danger isn’t being a wrapper. The danger is staying a wrapper — mistaking the initial traction for a durable business and never investing in the layer beneath the API call that would make the company defensible.
The honest self-assessment is simple: remove the model. Remove the API. What has your company built that is yours — that no platform update can take away and no competitor can replicate in a month?
If the answer is a proprietary dataset, a deeply integrated workflow, a network of users who make the product better for each other — you’re building a real business that happens to use an AI API. The API is an ingredient, not the product.
If the answer is a UI and some prompts — you’re renting. And the landlord is building the same apartment complex next door, with more rooms and a lower price.
The lease doesn’t last forever. Build something you own before it expires.
- * *