Here's an uncomfortable industry secret: many "AI products" on the market are a prompt, an API key and a nice interface. They demo beautifully. Then they meet real users, real data and a real invoice — and fall over. Knowing the difference before you sign is worth more than any discount.
What a wrapper actually is
A wrapper sends your user's question to a model API with some instructions bolted on, and displays whatever comes back. That's it. No knowledge of your business, no memory of your customers, no quality measurement, no cost control. It's not fraud — it's a fine validation tool. It's just not a product, and it shouldn't be priced like one.
The six layers that make AI a product
1. Grounding. The model answers from your documents, policies and data — retrieved on demand, cited in the answer — instead of improvising from internet memory. This is what kills hallucinations, and it's the difference between "sounds right" and "is right."
2. Evaluation. A test suite for answers. Before anything ships, hundreds of real questions run against the system and get scored. Change the prompt, rerun, compare. No evals means the vendor is shipping vibes.
3. Guardrails. What happens when a user asks something off-limits, hostile or simply weird? A product has an answer engineered for that. A wrapper has a public relations incident waiting.
4. Cost engineering. Caching, model routing, right-sizing — the unglamorous work that decides whether your success bankrupts you. A feature that costs $0.40 per use at 100 users is a rounding error; at 100,000 users it's a crisis.
5. Observability. When an answer goes wrong in week six, can anyone see what happened and why? Logged, traceable, debuggable AI is what "production-grade" actually means.
6. The feedback loop. Real questions users ask become test cases, knowledge updates and product decisions. The system gets better every week it runs — that's the moat a wrapper never builds.
Five questions that expose the difference
- "What does the AI answer from — and what happens when the answer isn't in there?"
- "Show me your evaluation suite for a past project."
- "What's your cost per conversation at 10× my current volume?"
- "What did your last AI product do in production when a user tried to break it?"
- "Who owns the code, the prompts and the data when we part ways?"
A serious builder answers these in specifics, happily. A wrapper shop changes the subject to the demo.
When a wrapper is the right call
Honesty cuts both ways: if you're validating whether users even want the feature, a two-week wrapper is smart spending — as long as everyone knows that's what it is, and the budget for the real build is planned, not a surprise.
We're biased, but we've earned the bias: we run our own AI products — grounded chat, realtime voice, AI personas — and every client build reuses architecture we've already battle-tested on ourselves. Test the difference yourself: our homepage is a live AI we built. Try to break it, then ask it what it costs.
