Zero Data Retention: Which AI Providers Offer It
What zero data retention means for AI, which providers offer it and on what terms, the exceptions it leaves, and what to ask an AI vendor.
- Published
- Updated
Zero data retention means an AI provider keeps no copy of your prompts or outputs after responding, apart from stated exceptions. OpenAI and Anthropic have approval-based arrangements, while Google Cloud, Microsoft and AWS apply product-specific controls. The provider name alone does not establish what happens to a confidential document.
For a legal team buying an AI feature, the useful question is whether the exact model, account and feature have the promised protection. A vendor can use an eligible model through an account or workflow that still stores your information.
What Zero Data Retention Means
Zero data retention, often shortened to ZDR, concerns what the provider retains after processing a request. The provider still receives and processes the information needed to answer. Its agreement defines the covered content and any safety or legal exceptions.
Default retention can include a period for abuse monitoring. For example, OpenAI’s API documentation specifies up to 30 days for those logs, subject to longer-retention exceptions. Approved ZDR removes customer content from that logging pathway, with feature-specific limits.
A promise against model training answers a separate question. A provider can exclude your documents from training while keeping them for support, conversation history or safety review. Likewise, Australian hosting describes location, without establishing when stored content is deleted.
Consider this hypothetical vendor assessment. A legal application sends a letter of advice through a ZDR-enabled API, then saves the letter and answer in its matter history. The model provider’s protection can apply while the application retains both documents. Your retention review must cover each copy.
Which AI Providers Offer It
Availability depends on the delivery product, approval and the features used. The table concerns business APIs and cloud services. A consumer chat account has its own terms.
| Provider | Availability and qualification | Covered route and limits |
|---|---|---|
| OpenAI | Eligible API customers require prior approval and additional requirements. Controls apply at organisation or project level. | Eligible endpoints include Responses and Chat Completions, with limitations. Files, vector stores and batches have separate storage requirements. |
| Anthropic | Approved commercial organisations can request ZDR through sales. Each organisation needs separate enablement. | Eligible Claude API use and specified Claude Code arrangements qualify. Batch processing and the Files API fall outside ZDR. Covered Models have separate retention requirements. |
| Google Cloud documents a configuration-based route to ZDR for its managed AI platform. It requires action across the applicable retention settings. | The Vertex AI retention documentation now redirects to Gemini Enterprise Agent Platform. Abuse logging, caching and stored interactions require separate attention. Some grounding and advanced features retain data. | |
| Microsoft | Eligible Azure customers can apply for modified abuse monitoring under additional Limited Access criteria. Advanced models can have stricter requirements. | Approval changes the abuse-monitoring treatment for the covered service. Stored application features require their own assessment. It is not a blanket promise for Microsoft 365 Copilot. |
| AWS Bedrock | Bedrock has model-specific retention modes. A compatible model and effective configuration determine whether no retention is available. | The none mode requires zero retention. Some models require retention or AWS review instead. Model-provider access and AWS retention are separate questions. |
OpenAI’s Endpoint Restrictions
OpenAI’s data controls documentation lists eligibility by endpoint. ZDR forces the storage setting to false on Responses and Chat Completions. Ineligible endpoints can still retain application state even within an enabled project.
Image safety exceptions and model-specific safety arrangements can change retention. Prompt caching can also retain derived representations. A contract that names only “OpenAI zero data retention” leaves those details unresolved.
Anthropic’s Product and Model Restrictions
Anthropic’s commercial ZDR terms identify eligible APIs and specified Claude Code access. They retain exceptions for legal compliance and misuse or harm. User Safety classifier results remain retained.
Claude zero data retention also depends on feature eligibility. Anthropic’s API documentation excludes stateful features such as batch processing and file storage. Using an excluded feature can put that data outside the arrangement without blocking the request.
Anthropic’s Covered Models policy requires 30-day retention for designated models, subject to specified programme exceptions. An organisation’s general ZDR status therefore does not establish that every Claude model qualifies.
Google Cloud and AWS Defaults Need Qualification
Neither Google Cloud nor AWS Bedrock supports a blanket conclusion that every request has no retention by default. Google’s documentation identifies default caching, optional request logging and stored interactions. Some grounding features impose storage that cannot be disabled while using that feature.
AWS documents an effective mode selected from project, account and model settings. Setting store=false alone does not guarantee ZDR. A model that requires a more permissive mode cannot run under none.
Microsoft’s Approval Covers a Defined Service
Microsoft’s abuse-monitoring documentation describes the application route and eligibility conditions. Approval removes the human-review abuse-monitoring process for covered use. Content classification and automated safety processing remain relevant.
Keep that approval tied to the Azure service and deployment it covers. A subscription to a different Microsoft AI product does not inherit the same arrangement.
What Zero Data Retention Does Not Cover
ZDR leaves several retention questions for the wider system. Start with features that need stored data. Uploaded files, saved conversations and asynchronous jobs can have their own retention schedules, even when the underlying text-generation endpoint qualifies.
Safety processing also needs careful reading. A classifier can inspect a prompt while the provider retains only a result or score. Some policies permit flagged content retention or additional model-specific review. Those exceptions affect whether a particular document is suitable for the workflow.
Usage metadata can remain after prompt content is deleted. Account identifiers, token counts and billing records can still reveal activity. Ask whether diagnostic logs include document names or fragments of content, and who can access them.
The application vendor adds another retention layer. It can keep uploads, generated answers and backups independently of the model provider. External search or connected services can receive data under separate terms.
For the broader relationship between retention, training and product settings, use the AI privacy guides. Keep the evidence for your particular deployment with the procurement record.
Questions to Ask a Vendor
Ask the vendor to answer these questions for the feature you will actually use:
- Which model provider, model version and hosting product process our information?
- Which organisation, project or subscription has ZDR enabled, and what written confirmation establishes that?
- Which endpoints and features does the application call, including file storage and background processing?
- Which safety, legal or model-specific exceptions permit content retention or review?
- What usage metadata remains, and can logs contain confidential content?
- What does your own application keep, including backups, support records and deleted matter data?
- What happens if the application switches model or falls back to another provider?
Use an AI risk assessment to connect those answers to the documents and staff who will use the feature. Record the agreed retention scope in the data processing agreement, including the vendor’s own copies. Approve the deployment against that written scope before sending confidential matter material.
FinTrace
Statements in. Evidence out.
FinTrace turns your bank statements into one searchable Excel ledger and a written findings report, with every finding cited to its source page.
FinTrace
See how FinTrace works