What is zero data retention in AI?

Zero data retention is an arrangement with an AI provider under which the provider stores none of your prompts or the model's responses once the response has been returned. Without it, the default on Anthropic's and OpenAI's API platforms is to keep inputs and outputs for up to 30 days. Zero data retention is approved per organization and covers listed API features only.

Does Claude Enterprise include zero data retention?

No. Anthropic's API documentation says the Claude Team and Claude Enterprise product interfaces are not eligible for zero data retention. Enterprise admins can set a retention period for chats instead. Zero data retention applies to Anthropic's API and to Claude Code used with a commercial API key, after approval by Anthropic's sales team.

Does Harriet offer zero data retention?

Yes. Zero data retention is available on request with Harriet and is agreed as part of the rollout together with the processing region. Today it covers Anthropic models only, which run EU-hosted with zero data retention available. Bring your own provider keys and your existing agreement carries over, or use Harriet managed credits and rely on the arrangement Harriet holds.

Is a no-training promise the same as zero data retention?

No. Anthropic, OpenAI and Google all say they do not train on business data by default, and all three still keep API inputs and outputs for a period unless zero data retention has been agreed. Training is about what the provider learns from your data. Retention is about whether a copy of your data exists on the provider's systems at all.

What is zero data retention and how can my organization get it?

Zero data retention, usually shortened to ZDR, means the AI provider keeps none of your prompts or the model’s answers once the reply has been sent.

What is zero data retention?

Zero data retention is an agreement under which the provider does not store your inputs or outputs at rest after the response is returned. Anthropic’s API documentation puts it in exactly those words. Without it, the default is a holding period. Anthropic says API inputs and outputs are deleted within 30 days of receipt. OpenAI says abuse monitoring logs, which can contain prompts and responses, are kept for up to 30 days. Google says prompts may be logged for abuse monitoring under its cloud terms.

“We do not train on your data” is a different promise. All three labs make the training promise to business customers by default. Retention is about whether a copy of your data exists on the provider’s systems at all, for how long, and who can look at it.

Does Claude Enterprise have zero data retention?

No. Anthropic’s API and data retention page says the Claude Team and Claude Enterprise product interfaces are not ZDR-eligible. The one exception is Claude Code on Claude Enterprise, which has its own ZDR offering for qualified accounts and needs separate enablement by Anthropic. What Enterprise admins get instead is a retention setting for chats, files and projects, which applies unless a user deletes something sooner.

ZDR itself sits on the API. It covers the Claude Messages API and Claude Code used with a commercial organization’s API key, is requested through Anthropic’s sales team, and is switched on per organization. That is the route Harriet uses for Claude, which is why zero data retention is available through Harriet when it is not available in the Claude app itself.

Does ChatGPT Enterprise have zero data retention?

No. OpenAI’s enterprise privacy page, updated January 8, 2026, says ChatGPT Enterprise workspace admins control how long data is retained, and that deleted conversations are removed from OpenAI’s systems within 30 days. The same page says API inputs and outputs may be kept for up to 30 days, and that zero data retention can be requested for eligible endpoints “if you have a qualifying use-case”.

OpenAI’s data controls documentation adds the detail. Approved API organizations choose between Zero Data Retention and a lighter option called Modified Abuse Monitoring. Chat completions and the Responses API are eligible. Conversations, files and vector stores keep application state until it is deleted, whatever the setting. Both controls are “subject to prior approval by OpenAI and acceptance of additional requirements”.

What about Google Gemini?

Google’s zero data retention page for its Gemini Enterprise Agent Platform, last updated October 1, 2026, describes ZDR as something the customer achieves by turning things off rather than something switched on in one place. A customer in scope for abuse monitoring requests an exception. Request and response logging stays off. The Interactions API is called with store set to false, because it defaults to true. Grounding with Google Search keeps query logs for up to three days and cannot be disabled, so Google points ZDR customers at Web Grounding for Enterprise instead.

How does my organization get zero data retention?

In practice it takes four steps.

Get API access. ZDR exists on the API platforms, so your people need to reach the models through an API rather than a consumer or team app. That can be your own developer account or a platform that holds the keys and the agreement for you.
Ask, and get approved. Anthropic reviews requests per organization through its sales team. OpenAI requires prior approval and extra commitments, after which a Data Retention tab appears in the organization settings. Google takes the abuse monitoring exception request through the account team. Neither Anthropic nor OpenAI publishes the eligibility criteria.
Check the model and feature list. Anthropic designates Claude Fable 5.1, Claude Mythos 5.1, Claude Fable 5 and Claude Mythos 5 as Covered Models that require 30-day retention and are outside ZDR unless Anthropic authorizes it. Code execution, Claude Managed Agents and the Console are not covered either. On OpenAI, anything that stores state until deleted stays stored. On Google, grounding with Search keeps its logs.
Decide whose keys and whose agreement. With your own keys the retention arrangement is the one you hold with the provider. Through a platform, it is the one the platform holds, so ask to see it.

Can my organization get zero data retention through Harriet?

Yes. Zero data retention is available on request with Harriet, and it is agreed as part of the rollout together with the processing region. Today it covers Anthropic models only. Anthropic models run EU-hosted with zero data retention available. Other providers on the router set their own rules, and zero data retention for them is not offered through Harriet today.

There are two ways to hold it. Bring your own Anthropic, OpenAI or Google keys and your existing agreement with the provider, including any ZDR arrangement you already have, carries over. Use Harriet managed credits and the arrangement is the one Harriet holds with the provider, so there is no application of your own to make. Either way the models are the same and spend is visible per team.

What the arrangement covers through Harriet:

The processing region, chosen per team, with OpenAI and Anthropic models hosted in the EU.
Zero data retention on the model side, on request, for Anthropic models.
A record of where each connected system sends data, shown before a team can enable it.
A DPA with EU and UK transfer safeguards and a SOC 2 Type II report in the Trust Center.

The detail of how region, retention and connected systems are settled is on the data sovereignty page.

What is still kept under zero data retention?

Safety records, flagged content and anything a connected tool keeps. Anthropic keeps User Safety classifier results to enforce its usage policy, and content its systems flag can be held for up to two years. OpenAI scans image and file inputs on submission and retains anything its classifier flags, with or without ZDR. Both say data processed by third-party tools, including MCP servers, follows those tools’ own retention policies. Google keeps grounding logs regardless.

A model under ZDR still sends data to whatever a connector or agent reaches, and in a real rollout that is where most of the exposure sits. ZDR on the model is one line in the data-flow story; the connectors are the rest of it, which is why Harriet shows where each one sends data before it can be switched on and keeps every tool call on the record.

If you are working through the wider question of where AI processing happens, EU data sovereignty for AI covers the route as well as the model, and what security asks before approving AI has the questions a security review will put to any vendor, zero data retention among them.