Skip to main content
SecureAI Guard

Zero data retention

Zero data retention means a provider does not persist your prompts or completions. The term means little until you pin down scope, duration and exceptions.

Published ·2 min read·SecureAI Guard
glossary
procurement
data handling

Zero data retention (ZDR) is a commitment that a service does not store the content you send it — prompts, completions, and usually the documents attached to them — beyond the moment it takes to serve the request.

It is the single most useful clause to have in writing when an LLM feature goes through a data-protection review, and it is also one of the most loosely used terms in the category. The phrase on its own commits a vendor to very little.

The questions that make it mean something

Ask all of them, and get the answers in the contract or the DPA rather than from a documentation page that can be edited:

  1. Scope. Does it cover prompts, completions, uploaded files, tool call arguments, system prompts and embeddings — or only the chat turns?
  2. Duration. "Not retained" often means "retained for up to 30 days for abuse monitoring". That is a legitimate design, but it is not zero, and a 30-day window is a 30-day breach exposure.
  3. Exceptions. Trust-and-safety review, incident investigation, legal hold and support tickets are the usual carve-outs. Which apply, who can read the data, and is access logged?
  4. Training. Retention and training use are separate promises. A provider can retain nothing and still have trained on your traffic before the contract started, and can retain data without training on it.
  5. Sub-processors. Does the commitment flow down to whoever the provider runs on?
  6. Which endpoints. ZDR is frequently available on some models, regions or tiers and not others, and enabling it may require an account-level setting that is off by default.
  7. Evidence. Is it in the DPA, and is there an audit or a report you can read that covers it?

Why it is worth the effort

Because it converts an open-ended question into a bounded one. A team that can state "prompt content is not persisted by the provider, per clause X of our DPA, and our own retention is 30 days in eu-west-1" has answered the hardest part of a third-party risk assessment in one sentence. A team that cannot has an unbounded finding sitting in front of the deal.

See our own answer on the trust page.