The most common follow-up question in projects where AI models are to be used for development, editorial work or support rarely concerns the quality of the results. It concerns the data: what happens to what the company feeds into the model? Who can see it, how long is it retained, and under what circumstances is it evaluated?

Against this backdrop, a recent report is interesting. OpenAI wants to make its most advanced models available to business customers without storing their data — while at the same time being able to detect misuse. In doing so, the provider addresses precisely the point at which many approval processes in mid-sized companies currently stall.

The conflicting objectives behind the announcement

Misuse detection and data minimisation technically pull in opposite directions. Anyone who wants to detect whether a service is being used for harmful purposes needs access to inputs and outputs in the classic implementation, often over a longer period, in order to recognise patterns across many requests. Anyone who does not store customer data does not have this basis.

That is precisely why the approach is relevant for companies, regardless of how it is solved technically in detail. It marks a requirement that is now being raised by many customers: the protection of one's own content must not depend on trusting the provider, but should be safeguarded by the architecture of the service.

No reliable details on the specific technical mechanisms are available so far. For assessment within your own project, this means: the announcement does not replace an examination, it merely provides an occasion for one.

Questions that should be clarified before selection

Regardless of the provider, a fixed catalogue of questions is worthwhile before an AI service is integrated into productive processes. The following points in particular have proven their worth:

  • Retention period: Are inputs and responses stored at all, and if so, for how long? Is there a contractually assured option without storage?
  • Training: Is customer data used to improve the models? Is exclusion the default, or must it be actively set?
  • Access by people: Under what conditions can the provider's staff view content, for example in the context of misuse checks or support cases?
  • Processing locations: In which regions does processing take place, and which subcontractors are involved?
  • Contractual basis: Does a data processing agreement exist that actually reflects the technical commitments — and does not merely repeat the marketing statement?

The order matters: clarify the data question first, then select the model. In practice it often happens the other way round, and approval then fails on points that could have been clarified before the prototype.

What you have in your own hands

Part of the risk can be reduced independently of the provider, namely through your own architecture. Anyone who builds AI functions into a website, a content management system or an internal tool decides which data leave their own area of responsibility in the first place.

An intermediate layer has proven its worth: instead of individual applications speaking directly to the provider's programming interface, traffic runs through your own service that acts as an intermediary. There, you can centrally define which fields are transmitted at all, which content is pseudonymised or removed beforehand, and what is logged.

This has three practical advantages. First, the provider remains interchangeable, because only one place needs to be adapted. Second, you can concretely demonstrate to data protection officers which data flows. Third, a place is created where costs and usage become visible, which experience shows quickly becomes relevant in AI projects.

Data classes instead of a blanket decision

Instead of granting or refusing approval for "AI" as a whole, a distinction by data class makes sense. Publicly intended content such as product texts, meta descriptions or blog drafts is uncritical. Internal documents, customer data, contract drafts and source code with business logic are at a different level.

For the first class, you can start quickly and gain experience. For the second, stricter requirements apply to storage, access and processing location — and that is precisely where commitments such as the one announced by OpenAI are relevant. This separation saves fundamental debates and makes decisions traceable.

Classification for web and software projects

For concrete projects, this means: the data question belongs in the requirements analysis, not in the acceptance phase. Anyone who routes AI integrations through their own intermediary layer from the outset and distinguishes by data class remains able to act when offerings, prices or commitments change. And they will — announcements like this are an interim state, not an end state.

Sources