Define zero before comparing vendors
Ask whether zero retention means no prompt and response storage after processing, no human access, no training use, or all three. These are related but distinct commitments, and marketing language can blur them.
Document the data classes in scope, including attachments, embeddings, tool results, safety signals, account metadata and logs. A review that considers only the text sent to the model will miss important copies.
Map the complete data path
Trace data from the user's device through identity systems, API gateways, orchestration services, model providers, tools and output storage. List the operator, region, purpose and retention rule at every step.
This map often reveals that the model provider is not the only retention point. A debugging platform may capture full prompts, a browser tool may store screenshots and a downstream application may persist the generated answer indefinitely.
Confirm feature-level eligibility
Request a matrix of models, endpoints and features covered by the retention control. Asynchronous processing, files, fine-tuning, assistants and abuse monitoring may have requirements different from a basic synchronous request.
Test the configuration in a non-production environment and preserve evidence of what was enabled. If eligibility requires vendor approval, record the approval and ensure new projects cannot bypass it through default settings.
Align the contract and architecture
Contract language should name retention periods, training use, subprocessors, breach notification, deletion, audit rights and what happens when a feature changes. A general privacy policy is not a substitute for terms attached to the purchased service.
Technical enforcement should match the promise. Use data-loss controls, access restrictions and configuration monitoring so a developer cannot silently route regulated information through an ineligible endpoint.
Operate the control after purchase
Review logs that confirm requests used the approved route without recording the sensitive payload. Monitor configuration drift, new product features and changes to subprocessors. Reassess when the use case or data class changes.
Finally, minimize data even when retention is zero. A system can expose information during processing, return it to the wrong user or send it to an unsafe tool. Strong procurement reduces risk, but accountable design and least-privilege operation finish the job.
Sources & further reading
Social-media activity is treated as a signal of attention, not proof. Product claims are attributed to the linked publisher or announcement.