PraxonOpen the network tab on ten "AI powered" products launched this month. A striking share reduce to an identical request pattern: user input goes into a text box, gets forwarded with a static system prompt to a foundation model's completions endpoint, and the response gets rendered back with some CSS on top. Strip the landing page and the pricing tiers and what remains is a single API key with a UI wrapped around it.
That's not cynicism, it's the direct consequence of where the cost structure moved. Building an AI product used to require training a model or at minimum fine-tuning one, plus the serving infrastructure to run inference at scale. That entire layer is now a metered API and a dozen lines of code. The capital and expertise barrier that used to gate this market collapsed, and a lot of companies shipped straight through the gap it left behind.
The tell is consistent. Ask what happens when the upstream provider deprecates the model version you built against, rate-limits you during a traffic spike, or silently changes output behavior after a model update, and there's no fallback, because nothing proprietary was ever built underneath the call. Ask what the product does that a well-written system prompt pasted into a general-purpose assistant couldn't replicate, and the honest answer is usually "the prompt is tuned well and the UI is clean." Those are real, but they're not defensible IP. They're a starting position, not a moat.
If the entire product is a system prompt and a form, the competitive set is anyone with API access and ten minutes to read your prompt out of the browser's dev tools.
None of this makes wrapper products worthless. Distribution, workflow design, and a well-scoped interface in front of a rented model are legitimate value, even with zero proprietary model weights involved. The failure mode is pricing that thin layer like proprietary technology and being caught off guard when a competitor reproduces it in a weekend with a marginally better prompt.
What actually creates defensibility
The products that hold up add something a raw completions call structurally cannot provide on its own:
- Retrieval over non-public data. A retrieval-augmented generation layer indexing something genuinely proprietary, not a vector store built on public documentation the model could already answer from.
- Orchestration, not a single call. Multi-step state management, validation, and error handling that turn one inference call into a pipeline that reliably completes an entire task.
- Domain-specific evaluation. A harness that actually catches when output is wrong for your use case, instead of shipping whatever the model returned unchecked.
- Switching cost independent of the prompt. Value embedded in accumulated data, integrations, or workflow history, durable enough that the product still matters if the underlying model provider changes.
Strip those away and what's left is a system prompt behind a subscription, which is an expensive way to proxy an API you could call directly at cost. Fine for a weekend build. A weak foundation for a company, because the only thing separating you from a free alternative is text anyone with the same model access can copy.
The AI companies worth watching right now aren't the ones with the sharpest prompt. They're the ones where handing a competitor the entire prompt, verbatim, wouldn't close the gap, because the defensible value was never sitting in the text box to begin with.