How to tell if an AI vendor is selling you a wrapper

Most AI vendors are reselling GPT-4 with a logo on it. That is the honest summary of the current market. Before you spend five figures on something that could be replicated with a £20 ChatGPT subscription and an afternoon, here are the questions worth asking.

Why does the wrapper problem exist?

Building a genuine AI system is hard. You need to understand the client's data, their workflows, their edge cases, and what failure looks like in their specific context. That takes time and expertise most agencies do not have.

A wrapper, by contrast, takes about a week. You point a prompt at the OpenAI API, build a thin interface around it, add some branding, and sell it as an AI platform. The margin is excellent. The value to the buyer is close to zero.

S&P Global found that the share of businesses abandoning most of their AI initiatives rose 17 percentage points to 42% in a single year. A lot of that abandonment comes down to buyers purchasing wrappers, getting wrapper-level results, and concluding the whole category was hype.

The vendor wins either way. You are the one left explaining to your board why the AI budget produced nothing.

What questions expose a thin wrapper fast?

Ask these in your first vendor call. The answers will tell you almost everything.

1. Where does my data live and how is it used to improve the output?

A wrapper has no good answer here. Your data passes through to OpenAI or Anthropic, the vendor has no control over it, and nothing about your specific business improves the model over time. A real system will describe a retrieval layer, a fine-tuning approach, a knowledge base, or at minimum a structured prompt architecture that is genuinely tailored to your context.

2. What happens if OpenAI changes its pricing or deprecates the model you are using?

If the vendor goes pale, you have your answer. A system with real engineering underneath can swap models. A wrapper is glued to one API and any change upstream breaks it or doubles your cost overnight.

3. Can you show me the system running on data that looks like mine?

Not a demo on their curated test data. Yours, or something structurally similar. Wrappers fall apart the moment they meet real-world messiness: inconsistent formatting, missing fields, edge cases the vendor never anticipated. A real system has been stress-tested against that kind of input.

4. What does the output quality look like when the input is bad?

Every AI system gets bad input eventually. A wrapper hallucinates confidently or produces garbage it presents as fact. A well-engineered system has guardrails, validation logic, and failure modes that are predictable and recoverable.

5. Who built this and what did they build before?

Ask for names. Ask for previous work. A team that has built real systems will have a portfolio of specific, describable outcomes. A wrapper shop will have testimonials, case study PDFs with no numbers in them, and a lot of talk about their methodology.

What does a real system look like in practice?

To make this concrete: Bidwell ingests 32,858 UK contract-award records, runs programmatic indexing across 2,100+ pages, and drafts complete tender responses grounded in that data. The AI is doing specific retrieval work against a structured dataset, rather than forwarding your prompt to GPT and hoping.

The model is one component. The architecture around it, the data pipeline, the validation layer, the output structure: that is the product.

A wrapper has none of that. It is a text box connected to an API.

How do you evaluate the commercial terms?

The contract often reveals more than the demo.

• Usage-based pricing tied to API calls. If you are paying per token and the vendor takes a margin on top of OpenAI's rate, you are paying for a middleman with no added value.
• No SLA on output quality. A vendor who cannot define what good output looks like and commit to it has not thought hard enough about what they are building.
• Lock-in with no data portability. If you cannot export your data and your configurations, you do not own what you have built with them.
• Vague IP clauses. Who owns the prompts, the fine-tuned layers, the custom logic? If the contract is silent on this, assume the vendor does.

A vendor selling a real system will be comfortable with specificity on all of these. A wrapper vendor will use words like "flexible" and "scalable" and change the subject.

What should a UK SMB do before signing anything?

Half of UK SMEs now use AI in some form, which means the vendor market has expanded rapidly to meet that demand. A lot of what has entered the market in the last 18 months is thin.

Before you sign:

1. Run the five questions above in a live call, not over email. Watch how they answer, not just what they say.
2. Ask for a technical contact, not just a sales contact. If there is no technical person available to talk to you before the sale, there probably is not one after it either.
3. Request a scoped proof of concept on your actual data, with defined success criteria agreed in advance. Any vendor worth working with will accept this. A wrapper vendor will stall.
4. Check what the system does when you push it. Give it a genuinely hard input. Ask it to handle an exception. See whether it fails gracefully or confidently produces nonsense.
5. Ask what the system cannot do. A vendor who answers this question clearly and without defensiveness understands their product. One who says it can do everything does not.

The single most useful question you can ask any AI vendor is: what does failure look like in this system, and how do you handle it? The answer separates engineers from salespeople in about thirty seconds.

What are the red flags in vendor marketing?

Some patterns in vendor positioning should make you cautious:

• "Powered by GPT-4" listed as a feature rather than an implementation detail.
• Screenshots of chat interfaces as the primary product demonstration.
• Case studies measured in time saved with no description of what the system actually does.
• "No-code AI" positioned as a selling point for a business-critical process.
• Pricing tiers named after animals or planets with no technical differentiation between them.

None of these are disqualifying on their own. Combined, they suggest a vendor who has prioritised the pitch over the product.

MIT research found that 95% of enterprise GenAI pilots see no measurable return. Most of those pilots were not failed experiments with genuinely novel systems. They were wrappers that looked like products until someone tried to use them for something real.

You do not have to be in that 95%. You just have to ask harder questions before you spend the money.

---

If you want a second opinion on a vendor proposal before you commit, or you want to understand what a system built around your actual data and workflows would look like, book a consultation and we will give you a straight answer.