There’s a category of AI product that looked, for a while, like a business. A clean interface, a focused use case — AI business advisor, AI legal assistant, AI sales coach, AI [your industry] copilot — and behind it, a handful of carefully tuned prompts feeding a commercial LLM. Stripe set up, Series A raised, dashboard live.
That category is finishing.
Not because the underlying idea is wrong. Domain-specific AI assistance is genuinely valuable, and many of these products do solve real problems. The issue is structural: a wrapper around someone else’s model has no defensible position once the model provider — or the next four competitors — decides to occupy the same space. And they are deciding.
The question that sorts the survivors from the writedowns is uncomfortably simple.
The moat test
Can the LLM provider replicate your product overnight with a system prompt?
If yes, you don’t have a product. You have an arbitrage — a price gap between the raw model and your packaged version of it — and arbitrages close.
It’s a brutal test, but it’s the right one. Every “AI X for Y” startup should run it before the next board meeting:
- Can a user achieve 80% of what we offer by pasting our category name into ChatGPT or Claude with a decent system prompt?
- Can OpenAI or Anthropic ship a “Projects” template, a Custom GPT, or an agent skill that covers our use case in an afternoon?
- If a competitor copied our prompts verbatim — which they can, by simply using our product and watching the outputs — what would they be missing?
If the honest answer to any of these is “not much,” the clock has already started.
Why this is happening now
The wrapper era worked, briefly, because the model providers were focused on the foundation layer. That window has closed. The platforms are deliberately moving up the stack:
- First-party agent runtimes that let users compose long-running workflows without third-party software.
- Skills, projects, and custom assistants that let any user package the exact same “system prompt + a few tools” recipe a wrapper startup was charging for.
- Direct integrations to the data sources wrappers used to broker — calendars, CRMs, file stores, the works.
- Pricing pressure as inference costs fall and consumer subscriptions absorb more capability.
The natural endpoint: anything that can be expressed as “the base model, plus a prompt, plus a couple of tool calls” gets absorbed into the platform’s default surface. The wrappers built on that exact recipe become features of someone else’s product.
This is not a hostile move. It’s gravity. Platforms expand into adjacent value until something stops them. Wrappers, by construction, are nothing but adjacent value.
What “no moat” actually looks like
The tells are recognizable once you know what to look for:
- The product is mostly a prompt. Take the prompts away and there’s a thin CRUD app underneath.
- No proprietary data flowing in. Whatever the user types is the only context the model sees — and the user could type it into a base model just as easily.
- No fine-tuning, no custom evaluation. The model is exactly the one the provider ships, with no measurable specialization.
- Switching cost is a logo. Customers leave the moment a 30%-cheaper or platform-bundled equivalent appears, because there’s nothing else holding them.
- The roadmap is “more prompts.” Roadmap items read like a list of features the LLM will support natively within a release or two.
If three or more of those apply, the product is downstream of a commodity and priced as if it isn’t. That gap is the arbitrage. Arbitrages compress.
What survives
The AI products that will still be around in three years share a common property: something about them is hard to copy with a prompt. The hard-to-copy thing is the moat. The candidates are limited and well-understood:
Proprietary data. A model that has been trained, fine-tuned, or grounded on a dataset competitors don’t have access to. Internal operational data, decades of domain transcripts, a labeled corpus that took years and SMEs to build. The model provider cannot replicate this with a system prompt because they don’t have the data.
Custom or specialized models. Fine-tuned weights, distilled models, smaller models tuned for specific tasks, retrieval pipelines whose architecture has been earned through trial and error. The base model is a starting point, not the product.
Subject-matter depth. Workflows designed by people who actually do the work, encoding judgement that doesn’t exist in any generic LLM’s training data. The product isn’t “AI for X” — it’s “the way X is done, made faster by AI.” Those are different.
Deep workflow integration. The product lives inside the operational reality of the business — connected to internal systems, sitting in the daily-use surface, integrated into how decisions are made and audited. Ripping it out is genuinely expensive, not because of contracts, but because of how the work has reorganized around it.
Compliance, evaluation, and trust infrastructure. Especially in regulated industries: documented evaluation pipelines, human-in-the-loop review where it matters, audit trails, model governance. This is unglamorous and slow to build, which is exactly why it’s a moat.
Notice the pattern: the moat is always something that took real time, real domain expertise, or real data to build. None of it is a prompt.
The connection to the bespoke thesis
We argued recently that generic SaaS is dying because the economics of bespoke have flipped — modern tooling makes custom systems faster and cheaper than ever, and the compromise tax on off-the-shelf platforms has finally become legible.
The wrapper-product collapse is the same argument from the other direction. Generic AI products — packaged prompts sold as a service — are the SaaS-era playbook applied to a layer that doesn’t support it. The model underneath is the commodity; the package on top is the compromise. Both ends of that arrangement are getting squeezed.
What remains, in both worlds, is the same answer: systems built for a specific business, using that business’s data, encoding that business’s expertise, integrated into how that business actually operates. Sometimes that means a bespoke ERP. Sometimes it means a fine-tuned model trained on your operational corpus. Often it means both, in the same system, because they’re solving versions of the same problem.
How to apply the moat test
If you’re building or buying an AI product, the diligence is short:
- Write down the prompt. If the value of the product is meaningfully captured by a paragraph of instructions, that’s your ceiling.
- List the data the product uses that nobody else has. If the list is empty, the product is renting from a commodity layer.
- Identify the smallest change the model provider could make to absorb your category. If it’s a Custom GPT template, you have months, not years.
- Look for the part of the product that took years to build. If you can’t find one, that’s the moat-shaped hole in the business.
A useful follow-up: does any of this require us to commission custom AI work — fine-tuning, a domain-specific model, a proprietary data pipeline, an evaluation framework no one else has? If the answer is no, you’re in the wrapper category. If the answer is yes, that work is the business.
If you’re trying to figure out what part of your AI strategy is a product versus an arbitrage, we can help. The work that builds a moat — fine-tuned models, domain-specific systems, the data infrastructure underneath — is what we do.