Wiring a feature to a model provider is a few lines of integration code. Deciding what that provider is now allowed to see is a decision that deserves more attention than the code takes to write.
The first question is not technical
Before anything else, we ask whether the feature needs to send the data at all, or whether it needs to send a derived, minimal version of it. A support-summary feature does not need a customer's full order history in the prompt if the last three orders answer the question; sending less is not just a privacy improvement, it also produces more focused output.
What we check on the provider side
- Data retention terms — how long the provider keeps the input, and whether it is used for further model training by default or only by explicit opt-in.
- Regional processing — whether data can be constrained to a specific region, relevant for clients under data residency obligations.
- Sub-processor list — who else the provider's infrastructure passes data through, since the answer is rarely just the one company whose logo is on the API.
What we check on our side
- Field-level minimisation — a defined, reviewed list of exactly which fields a given feature is allowed to send, enforced in code rather than left to whoever writes the prompt that week.
- Logging discipline — request and response logs used for debugging do not become an unmanaged second copy of customer data with none of the same controls as the primary database.
- A named owner for the decision, so "can we send this field" has an answer that does not depend on who is in the room when the question comes up.
Sending less data to a model is not only the safer choice. It is usually also the choice that produces a more reliable answer, because it removes the volume the model would otherwise have to sift through.
Where clients get this wrong before we arrive
The most common issue we find auditing existing AI features is not a bad provider choice, it is an unreviewed prompt template that grew to include entire customer records over several months of feature additions, with no single change that looks alarming on its own.
What this checklist actually costs
An extra half-day at the start of an AI feature, spent naming exactly what data the feature is allowed to touch. Compared to the cost of a retroactive data audit, or a difficult client conversation about what was already sent, that is not a close call.