Is My AI Startup Just a Wrapper?

Your AI startup is just a wrapper only if the whole product is a thin layer of prompt and interface around a model anyone can call. Calling something you do not own is not the problem. Most valuable software businesses run on infrastructure and APIs they do not own, and Stripe still processes payments on cloud and card networks it did not build. The real test is whether anything sits underneath your model call that a competitor cannot copy after they call the same model, whether that is proprietary data that compounds, a workflow a team runs its day inside, or distribution a clone cannot rent. If the honest answer today is nothing, you are a wrapper for now. That is a normal place to start, not a verdict, as long as you know which moat you are building next.

Is my AI startup just a wrapper?

Ask it plainly. You are a wrapper if a well-funded competitor can copy your interface, call the same foundation model, and reproduce your product in a weekend. You are past being a wrapper the moment something you have built makes that copy fail. A wrapper is defined by what is missing underneath the model call, not by the fact that you call a model at all.

The confusion comes from mistaking capability for defensibility. Access to a frontier model is a capability, and today anyone with a credit card and an API key has it. Defensibility is what remains once that access is universal. So the wrapper question is really a moat question, and the honest version of it is uncomfortable. After a competitor calls the same model you call, what exactly is left that they still cannot copy?

What "just a wrapper" really means

A wrapper is a product whose entire value is the prompt and the interface. The critique stings because it is often true at launch, and because the ground keeps moving underneath these products. A model capability that is a differentiator this quarter becomes a default feature the next, and the price of raw intelligence keeps falling.

The numbers here are stark. According to the Stanford HAI AI Index, the cost of running a model at the level of GPT-3.5 fell more than 280-fold between November 2022 and October 2024. When the core capability gets that much cheaper that fast, any advantage that lives purely in the model layer erodes on the same curve. David Cahn of Sequoia Capital made the macro version of this point, arguing that the capital pouring into the AI model and infrastructure layer has raced far ahead of the revenue it has produced. Compute and capital are not the scarce inputs. Defensibility is.

None of this makes wrappers worthless. It makes a wrapper a starting position rather than a finish line. Depending on infrastructure you do not own is normal, and it is not what the critique is about. Stripe runs on cloud and card networks it did not build and is one of the most valuable software companies in the world, because it layered settlement, risk, and developer trust on top that rivals cannot clone by calling the same rails. The problem is never that you call an API. The problem is a thin layer with nothing underneath it.

GPT wrapper vs defensible company: what actually separates them

A GPT wrapper and a defensible company can look identical in a demo. The difference is invisible in the interface and lives in what compounds underneath it. Warren Buffett borrowed the image of a castle protected by a moat to describe businesses that stay profitable because rivals cannot easily reach them. Greylock partner Jerry Chen updated the idea for software in his widely read essay "The New New Moats," arguing that durable software companies win not by owning a model but by becoming a system of intelligence, software that sits inside a customer's workflow, unifies data scattered across other systems, and gets smarter with use. That splits into three moats you can actually build, and none of them is the model.

Proprietary data that compounds

Public data trains everyone's model equally, so it defends no one. A real data moat comes from a loop. Your product captures data no competitor has, that data measurably improves the result, the better result drives more usage, and more usage produces more data. Fine-tuning on data anyone can scrape is not a moat. A claims tool that absorbs how one insurer handles its edge cases has a loop. A generic chat box does not.

Workflow depth

This is often the strongest moat for a young company. When your product becomes the place where work actually gets done, rather than a tool people open and close, you stop being optional. Depth looks like integrations into the systems a team already runs, multi-step processes that live inside your product, permissions and audit trails, and outputs other people in the organization depend on. A prompt box is shallow. A system a finance team runs its monthly close through is deep, and deep is expensive to rebuild even with a better model.

Distribution

Founders underrate this one. The history of technology is full of better products that lost to worse products with better distribution. For an AI startup, distribution can be a design partner who becomes a reference for a whole vertical, a platform you plug into, a community you built before you had a product, or a brand that makes you the default answer when buyers search. A cheaper clone still has to find and earn the customer you already reached. In Brazil and Latin America, where local trust, language, and presence weigh more heavily than in the United States, distribution built on the ground is especially hard to copy.

Where switching cost fits: it is the readout, not a fourth moat

Founders often add switching cost to a list like this and try to score it on its own. It does not belong there, because switching cost is not independent of the three moats above. It is what they add up to. The data a customer has accumulated inside your product, the processes their team has standardized on, and the integrations wired into their stack are the same data, workflow, and distribution planks seen from the customer's side. Counting switching cost as a separate fourth moat double counts the first three.

So use it as the readout instead. Once you have scored the three real moats, switching cost is the single question that tells you whether they are working. If a competitor launched tomorrow with a better model, how much would it actually cost your best customer, in time, risk, and lost history, to leave? Healthy switching cost is earned by delivering value, not by trapping anyone. It is the bill a customer pays to walk away, and the three moats are what put a number on it.

Score yourself: wrapper or defensible company?

Give yourself one point for each of the three moats that is real and growing today, not planned for someday.

Then read the total through switching cost, the bill a competitor makes your best customer pay to leave.

• **Proprietary data.** Does using your product create data no competitor has, and does that data visibly improve the next result?
• **Workflow depth.** Is your product where a team's work actually happens, holding the integrations and records they depend on, or a tool they visit and leave?
• **Distribution.** Can you reach and win your next hundred customers through a channel a clone cannot simply rent?
• **Zero points.** You are a wrapper today. That is fine if you are early, as long as you can name the one plank you are building next.
• **One point.** You have a head start turning into a moat. Deepen it before the model layer commoditizes your edge.
• **Two points.** Defensibility is forming. You are past the wrapper question and into the harder work of compounding what you have.
• **Three points.** You have a real moat. A better model rarely wins on its own against data, workflow, and distribution a competitor cannot reproduce.

Wrapper today does not mean wrapper forever

Almost every category-defining software company looked thin at the start. What separated the winners was a deliberate plan to convert early usage into data, workflow, and distribution before competitors caught up. In our experience co-founding companies from day zero, most begin at zero or one, a thin wrapper with a real idea underneath. That is not the problem. Standing still is.

This is also where the earliest choices matter most. Deciding which moat to prioritize, instrumenting the product to capture proprietary data from the first week, and picking a wedge that deepens into a workflow are architectural calls that are cheap to make early and expensive to retrofit. They are also hard to make well while you are still finding customers and staying alive. A studio that builds alongside founders from day zero, which is how Avante co-founds AI-native companies for Brazil and LATAM, exists to make those calls early rather than after the wrapper critique has already cost a funding round. If you want the full framework and a worked example, read does your AI startup have a moat, and if you are weighing the model itself, whether a venture studio is right for your AI startup is the next question to ask.

Preguntas frecuentes

Is my AI startup just a wrapper?

You are a wrapper if a competitor can copy your interface, call the same foundation model, and reproduce your product, with nothing left that they cannot match. You are past that point if using your product builds proprietary data that compounds, becomes the workflow a team runs its day inside, or reaches customers through distribution a clone cannot rent. Calling a model you do not own is not the issue. Having nothing underneath the call is.

Is my startup just a GPT wrapper?

A GPT wrapper is a specific case of the same question, a product whose whole value is a prompt and an interface around a general model. It is a fine place to start and a bad place to stay. The test is whether the product accumulates something the model does not have, such as exclusive data, deep workflow integration, or hard-won distribution. If it does, you are a company that happens to use GPT. If it does not, you are a feature that a model provider or a faster competitor can absorb.

What is the difference between an AI wrapper and a defensible company?

They can look the same in a demo, because the difference is not in the interface. A wrapper's entire advantage lives in the model layer, which keeps getting cheaper and more widely available. A defensible company has built something underneath the model that a competitor still cannot copy after calling the same model, usually proprietary data that compounds, workflow depth, or distribution. Defensibility is what survives once model access is universal.

Is being an AI wrapper always bad?

No. Many durable companies began as thin layers over a platform. A wrapper only becomes a bad business when it never converts early traction into a moat and the underlying model commoditizes its single advantage. The danger is standing still, not starting thin. The useful move is to name the one asset you are compounding and organize the company around feeding it.

What is the strongest moat for an early AI startup?

For most young companies it is workflow depth, what Greylock's Jerry Chen calls a system of intelligence. Becoming the place where a team's work actually happens is harder for a better-funded rival to copy than any prompt or model choice, and it naturally accumulates proprietary data and switching cost over time.