Zero data retention, explained: what happens to your data when AI reads it

Zero data retention (ZDR) means the model provider does not store your prompts or the model’s answers after the request is processed. Confirm it before client data goes to a model. “Available” is not the same as the setting that is actually applied.

What happens by default

The provider receives text, processes it, and returns an answer. What is kept depends on the configuration. Amazon says Amazon Bedrock gives you explicit control over whether the prompts and outputs from your inference requests are retained. An unconfigured region remains at inherit and falls back to each model’s default. Setting store=false does not guarantee zero data retention.

• Default storage. By default, Amazon Bedrock does not store model inputs or outputs. For specific models it may be required to store inputs and outputs for abuse detection.
• Copilot storage. Microsoft says the stored data includes the user’s prompt and Copilot’s response. Admins can set retention policies for those chats.
• A copy you do not control. For client financials, legal matters, or personal information, any stored copy is outside your control.

What ZDR covers

Amazon describes a mode in which no request or response data is written to durable storage by AWS or shared with the model provider. If someone then invokes a model that requires retention, Amazon Bedrock will block the request and return an error.

• No durable copy. No persistent copy of client data in the provider’s logs.
• No training. Your data is not used to train or improve the provider’s models.
• A smaller footprint. Less retained material if the provider is breached or subpoenaed.

Microsoft says prompts, responses, and data accessed through Microsoft Graph aren’t used to train foundation models. That is the no-training half. The privacy page says the prompt and the response are stored.

On Amazon, guaranteed zero retention is the mode in which no request or response data is written to durable storage, which that page distinguishes from a store flag left false.

What ZDR doesn’t cover

• Your vendor. ZDR is about the model provider. The vendor still stores data to do its job. Ask separately how that store is isolated and encrypted.
• Processing. Data exists in memory while the model works. ZDR is about what is kept afterwards.
• Every route. An unconfigured region falls back to the model’s default. A store flag set false does not by itself guarantee zero retention.
• Gateways. Ask whether any hop between your vendor and the model keeps its own log.

How to confirm it’s actually applied

1. A signed agreement. Is there a signed ZDR agreement with each model provider that sees the data?
2. No training. Is training on our data contractually excluded?
3. No fallback. Are routes to providers without ZDR turned off?
4. No extra logs. Is logging off at every hop, including any gateway?

Then ask whether the system refuses a model that fails any of the four. Amazon describes one version: a zero-data-retention configuration blocks a model that requires retention. A written policy depends on people remembering. A block does not.

Why it matters for businesses that hold client information

Engagement letters and confidentiality duties govern how client information is used. The FTC Safeguards Rule’s objectives include the security and confidentiality of customer information. Its scope includes tax preparation firms.

For tax return preparers, 26 U.S.C. § 7216 makes it a misdemeanor to disclose any information furnished for, or in connection with, the preparation of any such return, or to use any such information for any purpose other than to prepare, or assist in preparing, any such return, except where the statute allows it.

How Precision AI OS handles it

Language models run only under rules we write into the code. Only models on a list we review can be called. Every request asks our gateway for zero data retention, refuses data collection and is limited to the provider endpoints named for that model, with the gateway’s own fallback off. If a model is unavailable, the same request may go once to a second listed model, which can be a different provider, unless that is switched off for the agent.

We rely on the gateway’s published zero-retention listings; we don’t yet hold a signed zero-retention agreement with each provider. Providers may hold a prompt briefly, including in a short-lived cache, to process requests.

Every model call is recorded: which skill made it, which model and provider answered, and what it cost. Scanned documents and pictures are read by a language model under the same rules. See how we protect your data.