Zero data retention for AI note takers: what it covers, what it leaves out, and how to verify it
On this page
"Zero data retention" (ZDR) is an arrangement between an AI note-taker vendor and its AI model provider under which the provider does not keep the prompts and outputs it processes for that vendor once it has returned a response. OpenAI and Anthropic both document ZDR for their APIs, and Microsoft's Azure equivalent is called modified abuse monitoring; all three require approval or a contract. ZDR does not cover the vendor's own storage of your recordings, transcripts and summaries, its speech-to-text provider, or the copies you make. Facts below are from each company's own documentation as of October 2026.
This page explains what each provider's version actually changes, what it leaves out, and what to ask for as evidence in a security review. It is general information, not legal advice.
What does "zero data retention" mean, provider by provider?
The phrase is not a legal or industry standard. Each provider defines it in its own terms, and the definitions differ in scope.
| Provider | Default without ZDR | What ZDR (or the equivalent) changes | What still stays outside it | Source |
|---|---|---|---|---|
| OpenAI API | Abuse monitoring logs, which may contain prompts and responses, are kept for up to 30 days. API data is not used for training unless the customer opts in. | Customer content is excluded from abuse monitoring logs, and the "store" setting is forced off on the Responses and Chat Completions endpoints. Requires prior approval by OpenAI. | Endpoints listed as not eligible, including Assistants, Threads, Files, Vector stores, Batches, Fine-tuning and Conversations, may still store application state. | OpenAI data controls |
| Anthropic API | Inputs and outputs are deleted within 30 days, with exceptions for longer-retention features, usage policy enforcement and the law. | Anthropic does not store prompts or responses at rest after the API response is returned. Enabled per organization through Anthropic's sales team. | Batch processing (29-day retention), the Files API, the Console, Claude's consumer and team apps, some newer models that require 30-day retention, and content flagged by trust and safety systems, which may be kept for up to two years. | API and data retention; commercial retention |
| Azure (Microsoft Foundry, including Azure OpenAI models) | Prompts and completions are not used to train models. Abuse monitoring can store flagged prompts and completions for review by authorized Microsoft employees. | "Modified abuse monitoring", for customers who meet Limited Access eligibility criteria, removes that storage and human review; automated review may still run. | Features that store data in the customer's resource by design, such as the Responses API, Assistants threads, stored completions, batch and uploaded files. | Data, privacy and security; abuse monitoring |
Two details matter for interview tools. First, ZDR is set per account: Anthropic says each new organization needs it enabled separately, and OpenAI lets an organization set it at organization level and override it per project. A vendor can have ZDR on one account and not on the one your data runs through. Second, all three providers keep the right to retain data where the law requires it or where abuse is flagged.
What does ZDR leave out?
A note taker's ZDR claim is usually true and usually narrow. It describes one hop in the data flow, the call to the AI model that writes the summary. Everything else needs its own answer.
The vendor's own storage
Most note takers exist to keep a record. BrightHire, for example, says it protects candidate data "through Zero Data Retention programs with our AI sub-processors", and on the same page describes interview records preserved as complete references (BrightHire, AI in hiring). Fireflies says it has zero data retention agreements with its AI vendors, naming OpenAI and Anthropic, while its admins manage the retention of the meetings Fireflies itself keeps (Fireflies security policy). Neither statement is a contradiction: the AI provider keeps nothing, the vendor keeps your interviews. Your retention question goes to the vendor; which tools keep a recording at all is compared in AI note takers that don't store audio.
The vendor's own model training
ZDR binds the model provider, not the note taker. Granola, for instance, says it does not let third parties such as OpenAI or Anthropic train on your data, does not store meeting audio, stores transcripts and notes in a US-hosted AWS environment, and "trains on your anonymized data" unless you opt out in Settings; Enterprise users have training off by default (Granola security). Ask about the vendor's own training separately.
The speech-to-text provider
Audio usually goes to a transcription service before any language model sees text, and those services have their own terms. AssemblyAI, a common one, documents that by default uploaded audio files enter deletion from 24 hours, finished asynchronous transcripts are kept for 30 days, a shorter time-to-live can be set, and certain submitted files may be used for model training unless the customer has opted out, has a business associate addendum or uses its European servers (AssemblyAI data retention). A vendor that says "ZDR with all LLM providers" may not be talking about this hop at all.
Stateful features and other routes
Features that need memory, such as batch jobs, file storage and agent threads, sit outside ZDR at OpenAI and Anthropic. And if the vendor reaches Claude through Amazon Bedrock or Google-managed Vertex, Anthropic's page says the cloud provider is the data processor and its terms apply, not Anthropic's ZDR.
Your copies
Summaries pasted into the ATS, submittals emailed to clients and transcripts downloaded to a laptop are outside every vendor's controls. The data map for AI note takers shows where those copies end up.
The vendor says it has ZDR. What evidence should you ask for?
Ask for documents, not adjectives. This list slots into section 4 of the AI note taker security questionnaire.
- Name each AI provider and the platform. "OpenAI via its API", "OpenAI models via Azure" and "Claude via Bedrock" carry different terms. Check the names against the vendor's dated subprocessor list.
- Confirm the arrangement covers the account that processes your data. A written statement that ZDR (or modified abuse monitoring) is enabled for the organization or project your interviews run through, with the date it took effect.
- For Azure, ask for the setting. Microsoft documents that once storage for abuse monitoring is off, the resource's JSON view or the Azure CLI shows a "ContentLogging" capability set to false; the value does not appear otherwise. A screenshot is reasonable evidence.
- List the features and models outside it. Does the product use batch processing, stored files, assistants or threads, or a model that requires retention? Each is a gap.
- Cover speech-to-text. Name the transcription provider, its retention for audio and transcripts, and whether training is switched off for your data.
- State the vendor's own retention per data type. Audio, transcript, summary, scores, logs, backups: default period, whether you can change it, and how one candidate's data is deleted.
- Put it in the contract. The commitment belongs in the data processing agreement, with notice before a subprocessor or retention arrangement changes; see DPAs for recruiting software.
- Check the assurance scope. If the vendor offers a SOC 2 report, confirm the AI features are in scope; SOC 2 for recruiting software explains how to read one.
Which ZDR claims should make you ask more questions?
| What the vendor says | What it may mean | Ask |
|---|---|---|
| "We have zero data retention." | With which provider? It rarely means the vendor keeps nothing. | Which providers, which data, and what do you store? |
| "We use enterprise APIs, so your data is never stored." | Paid API terms usually rule out training, but default retention, such as 30-day abuse monitoring logs, can still apply without ZDR. | Is ZDR or modified abuse monitoring actually enabled? |
| "Your data is never used for training." | A use promise, not a storage promise. | How long is it kept, and by whom? |
| "ZDR with all our LLM providers." | Possibly not the speech-to-text provider, which handles the audio. | Who transcribes the audio, and what do they keep? |
| ZDR claimed for a feature that keeps history or files | Stateful features are outside ZDR at OpenAI and Anthropic. | Which endpoints or features does this product call? |
What should the contract say?
An example clause to adapt with your counsel, not a recommended standard:
AI subprocessors. Supplier shall process Customer Personal Data
through third-party AI model and speech-to-text providers only
under arrangements in which those providers (a) do not use the
data to train or improve models and (b) do not retain prompts,
inputs, audio or outputs after processing, except where required
by law or for abuse investigations as described in the
provider's published terms. Supplier shall list each such
provider, the service used and the retention terms in Annex [X],
and shall give Customer [30] days' notice before any change.Does ZDR matter to candidates?
It narrows one copy of their interview, which is worth having. But when a candidate asks what you hold or asks for deletion, the copies that matter are the vendor's and yours; see candidate data access requests. When you answer a candidate's "where does this go?", describe the whole flow, not only the model provider.
We make Interview Signal, which takes a different route: transcripts, quotes, candidate names, notes and submittals stay on the recruiter's computer, audio goes to a speech-to-text provider in short chunks during the call and is not retained, and our server keeps only counts and durations. You should still ask us the same questions; the security page lists the subprocessors.
Not legal advice. Provider terms change; check each linked page before you rely on it.
Questions people ask
What does zero data retention mean for an AI note taker?
It usually means the note taker's AI model provider, such as OpenAI or Anthropic, has agreed not to keep the prompts and outputs it processes for that vendor after returning a response. It is an arrangement between the vendor and its provider. It does not mean the note taker itself keeps nothing: the vendor can still store recordings, transcripts and summaries on its own servers.
Does zero data retention mean my interview recordings are deleted?
No. Zero data retention covers what an AI model provider keeps after processing a request. Recordings, transcripts, summaries and scores that the note-taker vendor stores for you follow the vendor's own retention settings, and copies you paste into your ATS or email follow yours. Check each separately.
How can I verify an AI note taker's zero data retention claim?
Ask the vendor to name each AI provider and the platform it uses, confirm in writing that zero data retention or the provider's equivalent is enabled for the account that processes your data, list any features that fall outside it, cover its speech-to-text provider too, and put the commitment in the data processing agreement. For Azure, Microsoft documents a ContentLogging value of false that shows abuse-monitoring storage is off.
Is zero data retention the same as not training on my data?
No. A provider can promise not to train on your data and still keep it for a period, for example OpenAI's API keeps abuse monitoring logs for up to 30 days by default and Anthropic deletes API inputs and outputs within 30 days by default. Zero data retention is about storage; no-training is about use. You want both answers in writing.