Before you tell customers your LLM provider keeps nothing, ask which endpoints, which features and which log types the promise covers. Zero data retention programs usually need approval, apply to some endpoints and not others, and can still leave exceptions. Get the scope in writing and match your own claims to it.
What the default usually looks like
Defaults differ by provider and change over time, so read the current terms. As one example, one major provider's documentation states that data sent to its API is not used to train its models unless you opt in, and that abuse monitoring logs are kept for up to 30 days by default, unless longer retention is required by law or reasonably necessary to protect the services or third parties.
So even a provider that does not train on your data may hold your prompts for a period. That matters if you have promised customers that prompts are never stored.
What the zero retention option may look like
The same documentation describes a zero data retention option that requires prior approval and additional requirements, and that excludes customer content from abuse monitoring logs. It also describes a modified abuse monitoring option that excludes customer content from those logs across endpoints, with a rare exception for image and file inputs.
The detail to notice is that eligibility differs by endpoint. In that documentation, chat and response style endpoints, embeddings and moderation are listed as eligible with some limitations, while features that store state, such as files, vector stores, batches, fine-tuning jobs and conversation threads, are listed as not eligible. If your product uses stored files or batch processing, the promise may not cover them.
Questions to ask
- Which endpoints and models does the program cover?
- Which features that store state are excluded, such as file uploads, stored conversations, batch jobs and fine-tuning?
- What logs remain, such as billing metadata, request IDs and token counts?
- Are there exceptions for image or file inputs?
- Is there any retention for legal or safety reasons, and how is it handled?
- How is the setting applied: per account, per project, per request? Can a developer switch it off by accident?
- What is the approval process, and can it be revoked?
- Does it apply in every region you use?
- How will you be told if the terms change?
Match your claims to the answers
A customer-facing sentence such as "we never store your prompts" is risky unless you control the whole path. Your own logs, error trackers, analytics and vector databases may all hold prompt text. Say what is true: "our provider does not retain prompt content for abuse monitoring under our agreement, and we keep application logs for 14 days with personal data removed", if that matches reality.
For the inside half of the story, see logging LLM interactions without leaking sensitive data.
Do not skip your own retention
Zero retention at the vendor does nothing for the copies in your own systems. Inventory where prompts land: logs, traces, support tickets, evaluation sets and backups.
Quick check
Does zero retention mean the provider never sees the data?
No. The provider processes the request to answer it. Zero retention is about what is stored afterward, within the scope of the program.
Is it required for GDPR?
Not by itself. GDPR asks for appropriate safeguards and a proper processor contract. Zero retention can help you meet data minimization goals, but it is a choice, not a rule.
Sources
Last reviewed: 19 October 2026.
