There are roughly 10,000 "AI companies" today. Most of them are a login page, a prompt template, and an OpenAI API key. They're wrappers, and the enterprise is paying a wrapper tax.
None of this is inherently wrong. Wrappers can be useful. The problem is when a wrapper charges enterprise licensing fees, claims proprietary technology, and locks you into a relationship that adds cost without adding value.
Here's how to spot one.
1. They can't explain what happens between your input and the model's output.
A real AI product has orchestration: retrieval, pre-processing, context management, post-processing, validation, guardrails. A wrapper sends your text to an API and returns the response. Ask what their system does besides call the model. If the answer is vague, you have a wrapper.
2. Their "proprietary AI" launched within weeks of ChatGPT.
Building genuine enterprise AI capability takes months of engineering. If a company pivoted to "AI-powered" in early 2023 and had a product by March, they didn't build AI. They wrapped it. Check their engineering blog, their team page, their hiring history. The evidence (or lack of it) is usually obvious.
3. They can't operate without a single model provider.
Ask: "What happens if OpenAI doubles their API pricing tomorrow?" A real AI product has abstraction layers, model-agnostic architecture, the ability to swap providers. A wrapper collapses. If your vendor's entire capability is downstream of one model provider's API, their pricing, availability, and roadmap are not in your vendor's control. Or yours.
4. They charge per-seat for something you could build in a weekend.
The test: could a competent developer replicate the core functionality with the model provider's API, a basic prompt, and a simple frontend? If yes, you're paying for the wrapper, not the capability. Enterprise-grade AI adds value through integration, governance, domain-specific fine-tuning, and workflow orchestration, not through a nicer chat interface.
5. They get nervous when you ask about data sovereignty.
Where does your data go when it enters their system? Which model provider processes it? In which jurisdiction? Under what terms? A wrapper often can't answer these questions because they don't control the answer. Their upstream provider does. For NZ enterprises with sovereignty obligations, this is a disqualifying gap.
The Wrapper Tax
The wrapper tax isn't just the margin between the API cost and what you're paying. It's the opportunity cost of building on a dependency chain you don't control.
When OpenAI changes their API (and they will), your wrapper vendor scrambles. When a better model launches (and it will), you're locked into whatever your wrapper supports. When regulation requires audit trails for AI decisions (and in NZ, it's coming), your wrapper may not have the infrastructure to provide them.
When a Wrapper Is Fine
For low-stakes, internal productivity tools (summarising meeting notes, drafting emails, brainstorming), a wrapper is fine. The risk is low, the switching cost is low, and the convenience is real.
For enterprise capabilities that touch customers, make decisions, or process sensitive data: you need more than a wrapper. You need architecture that compounds.
Ask the five questions. The answers will tell you what you're actually buying.