Three kinds of AI product: GPT wrappers, specialists, and agents
Not all AI apps are built — or defended — the same way. A look at the three archetypes: thin wrappers, domain specialists, and agentic products, and where the moat actually is.
From the outside, most AI products look alike: a box you type into, an answer that comes back. Underneath, they fall into three archetypes with very different economics and defensibility. Knowing which one you're building — and which you should be — is a strategy question, not a technical one.
1. GPT-wrapper apps
A thin interface and a clever prompt over a model API. These are fast to ship and perfect for validating demand — but they have little moat. Anyone can rebuild them in a weekend, and the model provider can absorb the feature into their own product overnight. A wrapper is a great way to start; it's a dangerous place to stay.
2. Specialised / domain-expertise apps
These go deep on a vertical: proprietary data, integration into a real workflow, domain-tuned prompts and evals, and the compliance a serious buyer demands. Much harder to build than a wrapper — and much harder to copy. The AI is one ingredient; the domain depth around it is the product.
3. Agentic apps
These don't just answer — they take actions across systems to complete work end to end. The potential value is the highest of the three, and so is the difficulty: they need guardrails, evals, orchestration, and real reliability engineering before you can trust them in front of customers.
Where the moat actually is
- Not the model — everyone can call the same APIs; the model is a commodity input.
- Proprietary data and workflow lock-in — what you have and how deeply you're embedded.
- Domain trust and reliability — being demonstrably right, and safe, in a specific field.
- Execution — for agents, getting them to actually finish the job dependably.
The model is the same for you and every competitor. Your moat is everything you wrap around it — data, workflow, trust, and reliability.