Skip to content

Security and privacy20:06

A quote 10 times higher than the API price list. 12 questions for the meeting

How to verify whether a vendor sells custom AI implementation or just a wrapper around someone else's model. 12 vetting questions, margin thresholds, and when a wrapper makes sense.

AI-generated image.

Listen to the article summary

Synthetic voice.
0:00
0:00

A practical guide written from the perspective of a service provider. It describes common solutions and does not assess individual cases. The first test is simple: ask for the exact model name, version, and endpoint. If a provider talks about a “proprietary model” but cannot provide documentation, it is usually a wrapper around someone else’s API. Price is another signal: offers for the same task can vary by 3 to 10 times because some companies simply resell access to a model. The median integration margin for a genuine integration is 18–35% of the model cost. A margin above 50–100% without proprietary engineering components is a sign of a wrapper. Since 2 August 2025, general-purpose AI models have had their own obligations, and from 2 August 2026, most provisions of the AI Act apply.

AI-generated image.

Where this comes from

Language model APIs are cheap and easy to wrap. An interface, prompt, and RAG layer on top of GPT-4o can be built in a week, while OpenAI’s official pricing lists USD 2.50 per million input tokens for the GPT-4o mini model. A provider that does not control the model cannot take responsibility for the provider’s errors or guarantee the processing region. Article 28 of the GDPR requires a data processing agreement and a list of subprocessors. A wrapper usually has neither. This is why some offers cost 10 times more than the API price, with no engineering justification.

What you can do about it

When meeting with a provider, it is worth going through 12 due diligence questions. First: what exactly are the base model, version, and endpoint? Second: is the agreement directly with the model provider or through Microsoft Azure, Google Cloud, or AWS, and what is the provider’s partner status? Third: is BYOK and direct billing available? Fourth: who is the GDPR data processor, and what is the list of subprocessors? Fifth: in which region is data processed, and is an EU-only option available? Sixth: does the solution use fine-tuning, RAG, and vectorization, or just a UI and prompt? Seventh: what is the net price per 1 million input and output tokens? Eighth: what SLA is guaranteed, and how is it measured? Ninth: can the base model be changed without amending the agreement, and is data export guaranteed? Tenth: have prompt injection and hallucination tests been conducted, and how are results reported? Eleventh: how does the provider classify the system under the AI Act, and which obligations does it accept as a provider? Twelfth: who is responsible for model errors, and what are the liability caps?

The difference between an integration and a wrapper comes down to control and responsibility. An integration configures the model within the client’s architecture, with a security layer, monitoring, an SLA, a data processing agreement, and responsibility for the entire solution. A wrapper is an interface or prompt on top of someone else’s API, with no control over the model and no responsibility for the provider’s errors. In practice, we check this by comparing the price per token with the provider’s official pricing. The median integration margin for a genuine integration is 18–35% of the model cost. An offer priced 10 times higher than the API price, with no engineering justification, should be rejected.

AI-generated image.

When this will NOT work

A wrapper may be acceptable if it meets four conditions: clear disclosure that it is an interface built on someone else’s model, pricing close to the API cost plus a 5–15% margin, a BYOK option, and a short pilot with a low level of risk. In regulated industries such as finance, healthcare and e-commerce, a wrapper without control over the data processing region usually fails the GDPR test. The Act on Combating Unfair Competition also covers misleading claims about owning a model, so a provider that hides the fact that it resells someone else’s API risks legal disputes. In practice, a 50% margin without proprietary components is the point at which it is worth requesting a technical audit.

HEXART helps you work through these 12 questions in practice: we review offers, compare prices with vendors’ price lists and prepare a list of evidence for a supplier audit. Learn more about our implementations at hexart.pl.

Source materials

Sources consulted during research. The text above is our own; these third parties are not responsible for its content and have not authorised it.

No-obligation call

Schedule a workshop with Paul

Thirty minutes or a full workshop, choose what suits you. Book a slot directly in the calendar and get instant confirmation.

Paul Lazniak

HEXART Founder