Zero data retention for AI: what it covers

Zero data retention depends on the service and endpoint. Separate prompt retention, abuse monitoring, application logs, saved files and training controls.

Zero data retention limits what an AI service keeps from specified requests after processing. Its protection depends on the provider’s agreement, the account configuration and the endpoint or feature you use. It doesn’t automatically remove your store’s conversation history or the records in a helpdesk.

For a retailer, the practical task is to follow a customer’s information through the whole request. A model provider can discard the message while your application still keeps a copy in an error report.

What zero data retention means

Zero data retention (ZDR) is a promise about defined data within a defined service. Read it as a statement about particular content and processing records, with stated exceptions. It isn’t a promise that every system involved in a conversation stores nothing.

Separate the controls before comparing providers:

  • A no-training commitment controls whether customer data is used to improve models. It doesn’t establish when stored content is deleted.
  • Encryption protects data while it moves or remains stored. Encrypted records still exist.
  • Deleting a saved chat affects the product and records covered by that deletion process. It doesn’t configure the application programming interface (API) used by a separate shopping agent.
  • A ZDR arrangement limits retention for its covered requests. Operational metadata and features that need storage require separate attention.

“Not used for training” and “not retained” answer different procurement questions. Record the answer to each before treating either as a reason to send customer information.

Follow the records beyond the model

A retail request passes through systems with separate storage choices. The provider’s ZDR arrangement applies at its own boundary. Other systems need their own controls.

The following is a fictional route for an order-support question. It illustrates places to inspect, without describing any particular agent’s suppliers or retention policy.

StageWhat can pass through itRecords to account for
Customer’s browserQuestion, conversation history and uploaded contentBrowser storage and any session recordings
Retailer’s applicationMessage, account context and order lookupSaved conversations, request logs and retry queues
Model endpointSelected message content and tool resultsCovered content, feature state and safety exceptions
Connected toolOrder lookup or another permitted actionTool requests and its own audit records
Support and monitoring systemsEscalated conversation or failure reportTickets, traces and attachments

Suppose a customer asks why a parcel is late. The application sends the order status and the question to a covered model endpoint. It also forwards the conversation to staff. Provider ZDR doesn’t delete the staff ticket or the application’s saved conversation.

The same applies when a browser monitoring tool records a message before the application sends it anywhere. A promise at the model endpoint doesn’t govern that earlier copy.

Keep a record owner for each stage. Note what it stores, who can read it and the event or retention period that removes it. Include any service that receives tool-call arguments or attachments.

OpenAI and Claude controls

OpenAI and Anthropic publish separate API retention arrangements. Consumer chat settings don’t establish that a retailer’s API requests receive those protections.

OpenAI zero data retention

OpenAI requires prior approval and additional requirements for ZDR. Approved customers configure organisation and project controls. A project can inherit the organisation setting, select a control explicitly or disable it.

ZDR excludes customer content from abuse-monitoring logs and forces store=false for Responses and Chat Completions. Its endpoint table distinguishes eligible calls from stateful features.

Responses and Chat Completions are eligible, subject to feature limitations. Files, Batches, Vector Stores and Conversations are outside ZDR eligibility.

Image and file inputs have a safety exception: images flagged for potential child sexual abuse material can be retained for manual review. OpenAI also publishes model-specific safety retention provisions that can change an approved customer’s coverage with advance written notice. These boundaries appear in its API data controls.

Two features need particular attention. OpenAI’s background mode temporarily stores response data for roughly ten minutes to support asynchronous execution and polling, including for ZDR projects. Its prompt caching documentation describes model-dependent storage of encrypted cache representations, including retention for up to 24 hours. Match the model and cache settings to the required data boundary.

Claude zero data retention

Anthropic’s ZDR agreement guidance makes approval specific to each organisation. It identifies eligible commercial APIs separately from consumer products. It also preserves legal and misuse exceptions and retention of User Safety classifier results.

For covered API calls, Anthropic says prompts and responses aren’t stored at rest after the response returns. Messages and Token Counting support ZDR for eligible features. The same Messages endpoint can use features that fall outside the arrangement.

Claude API featurePublished ZDR position
Standard Messages and Token CountingEligible with a covered model
Batch processing, Files and code executionNot eligible
Structured outputsQualified eligibility: a schema can be cached for up to 24 hours since last use
Prompt cachingEligible: in-memory cache representations persist for the cache lifetime

Anthropic says non-eligible features aren’t automatically blocked under ZDR. Using them puts that data under the feature’s own retention policy. ZDR organisations must also route web-application calls through their back end because direct browser requests are rejected. Its API retention documentation explains those boundaries.

Model choice adds another condition. Anthropic’s Covered Models policy requires 30-day retention for designated models unless the organisation has an expressly authorised exception. Enabling retention for one workspace changes the protection of requests sent through that workspace. An organisation’s general ZDR approval alone doesn’t establish coverage for every model.

Availability, approval and use are separate checks

A provider can advertise ZDR while your organisation has no approval. Your organisation can have approval while an application uses a different project or an excluded feature. A successful response establishes that the request worked. It doesn’t establish its retention treatment.

Keep the agreement and the effective account setting together with the application’s endpoint and model configuration. Include fallback routes used after a timeout. A fallback that switches providers or features creates a different data path to assess.

Questions to settle before sending data

Turn the retention promise into a set of questions your team can answer for each request path. Use synthetic customer details when inspecting how the application behaves.

  1. What content leaves the application? Include system instructions, earlier conversation turns and tool results, as well as the latest customer message.
  2. How are attachments sent? Distinguish an inline photo or document from a file uploaded to a persistent Files service. Include filenames and document metadata.
  3. Which organisation, project or workspace receives the call? Match the credential’s scope to the approved retention setting without copying credentials into the assessment.
  4. Which features create state? Account for batches, background work, caches and containers. Include saved conversations and externally connected tools.
  5. What survives a failed request? Inspect application errors, retry queues and monitoring reports. A client timeout doesn’t establish that the provider never received or processed the request.
  6. What exceptions remain? Record safety handling and legal retention, together with any model-specific conditions. Separate content from billing or usage metadata.
  7. How do you remove other copies? Assign deletion and access controls to application records, support tickets and backups under their own retention rules.

Use one synthetic request to follow the normal path, then trigger a controlled failure in a test environment. Confirm that the error report omits message content and that retries use the intended route. These checks test your application’s handling. The provider’s agreement and effective configuration establish its promised boundary.

Choose controls for the remaining records

Combine provider ZDR with data minimisation and your own retention rules. Limit staff access to the records needed for the job, and avoid copying a full customer record into each model request.

Consider a fictional retailer handling a delayed-delivery question. Its order system already knows the customer’s delivery address. The agent needs the dispatch status to explain the delay, so the application sends that status with a temporary order reference. It leaves the full address and payment details in the order system.

The retailer uses an approved organisation and an eligible model feature for that request. It logs a request identifier and outcome without the conversation body. A failure report contains a redacted error, so a retry doesn’t create another copy of the customer’s question in monitoring.

If the customer asks for staff, the application creates a ticket with the details staff need. The ticket follows the retailer’s support retention policy and is visible only to the relevant team. Provider ZDR leaves that record untouched.

Before accepting “zero retention” in a supplier assessment, name the service and content the promise covers. Then assign controls to every remaining record. That gives your team an actionable boundary for the customer information it sends.

Give your store its own shopping agent

Customers ask in their own words. Your agent finds, compares and checks out from your catalogue, right on your site.

See it in actionFree trial · Simple setup