All Glossaries

/

Zero Data Retention (ZDR)

Zero Data Retention (ZDR)

Zero data retention (ZDR) means a provider keeps no copy of your content after processing. What it covers for LLM and voice AI vendors, and what to ask.

Zero data retention (ZDR) is an arrangement in which a service provider processes your content, such as a prompt, a transcript or audio, and keeps no copy of it once the response is returned. The term comes from large language model (LLM) APIs, where it usually means request and response content is excluded from the provider's logs. It is a contract or account setting, not a law or a certification.

What does ZDR mean for an LLM provider?

ZDR means the model provider does not store your inputs and outputs after the request, but every provider defines the scope differently. Two examples from the providers' own documentation:

  • OpenAI keeps abuse monitoring logs for API traffic for up to 30 days by default. ZDR removes customer content from those logs, requires OpenAI's prior approval, and does not cover endpoints that store application state, such as assistants, threads and vector stores.
  • Anthropic applies its ZDR agreement to eligible commercial API use, not to consumer plans. It still keeps safety classifier results to enforce its usage policy and can retain data where needed to comply with law or combat misuse.

So "we have ZDR" is never the full answer. The useful question is: ZDR with whom, for which endpoints, and with what exceptions?

Why is ZDR harder to pin down for a voice agent?

A phone call creates data at several layers, and a ZDR agreement with the LLM covers only one of them.

Data from one callWhere it is created or heldCovered by an LLM ZDR agreement?
Live audioCarrier, speech-to-text provider, voice platformNo
TranscriptSpeech-to-text provider, voice platformNo
Prompt plus transcript sent for each replyLLM providerYes, if the endpoint is in scope
Synthesized speechText-to-speech providerNo
Call recording, logs, tool call dataVoice platformNo
Post-call summariesVoice platform, plus another model callDepends on that call
Timestamps, duration, phone numbers, costVoice platform, carrierUsually kept

The carrier layer sits partly outside the platform's control. Retell's privacy policy notes that telecommunications providers may process call data as independent controllers for their own purposes, such as billing and fraud prevention.

ZDR also differs from two controls it is often confused with. A retention period stores content and deletes it after a set number of days. PII redaction stores content with identifiers removed. ZDR stores no content at all.

How do Retell's storage settings map to ZDR?

Retell's documentation describes per-agent storage and retention settings rather than a mode called zero data retention. The closest controls:

  • Basic Attributes Only. Retell stores metadata only, with no transcripts, recordings or logs (data storage settings). Your webhook still receives the transcript and recording, and the recording link expires 10 minutes after the webhook arrives, so content lands in your systems instead.
  • Everything except PII. Retell deletes the raw originals and keeps scrubbed copies. Call IDs, timestamps, call status and cost fields are never scrubbed.
  • Retention period. Set automatic deletion per agent from 1 day to 730 days; the default is to keep data forever. Expired recordings, transcripts, logs, knowledge base retrieval logs, dynamic variables and metadata are deleted daily and cannot be recovered, though some basic metadata is retained internally (data retention policy).
  • Custom LLM. If prompts and transcripts must be processed in your own infrastructure or by a model you host, Retell's docs point to a custom LLM. Retell still handles telephony, transcription and speech synthesis.

Storing less has costs. With Basic Attributes Only, the Get Call API no longer returns transcripts or recordings, and PCAP files for debugging SIP failures exist only when storage is set to Everything. For a healthcare setup, see our guide to HIPAA compliance for AI phone agents.

What should you ask a voice AI vendor about ZDR?

Ask for answers per layer, in writing:

  1. Which layers does ZDR cover: LLM, speech-to-text, text-to-speech, telephony, and your own platform?
  2. Which subprocessors have ZDR in place, and which run on default terms? Ask for the subprocessor list.
  3. What do those providers keep for abuse monitoring or safety, and for how long?
  4. What metadata survives deletion, and is it linked to a phone number?
  5. Is call data used to train or fine-tune any model, yours or a provider's? Get the answer in the DPA.
  6. If content is not stored, where does it go: webhooks, exports, expiring links?
  7. Does deletion reach backups, and on what schedule?
  8. Which features stop working when nothing is stored: call history, QA, post-call analysis, debugging?

ZDR answers how long data is kept. Data residency answers where it is kept, and a SOC 2 Type II report shows whether the controls held up over an audit period. Ask about all three together; our enterprise voice AI compliance guide covers the DPA review.

If your security review needs per-agent control over what a voice agent stores and for how long, see how the Retell enterprise platform handles it, and test your storage settings on real calls during a pilot.

Add Retell AI as a preferred source on Google
Recommendation

Related AI Voice Agent Terms

Revolutionize your call operation with Retell