← Blog

Zero Data Retention for lawyers — what it actually means

Published 4 May 2026 · 6 min read · By GD · LexCodex

When a lawyer sends privileged contract text to an AI service, two questions matter. Is the text stored anywhere, and can it be used to train future models? If the answer is unclear, the service is in practice unusable for client work.

Zero Data Retention (ZDR) is the term for the commitment that addresses both questions. But the term is used with different meanings by different vendors. This text describes what ZDR actually means at LexCodex, at which layers it applies, and where the limits are.

Two layers of retention

A document you upload passes through two systems. Each has its own retention rules.

Layer 1

The LLM provider

LexCodex uses an LLM engine from a provider that is certified to SOC 2 Type II and ISO 27001. Text sent there is processed on the provider's infrastructure. The provider offers an enterprise term called Zero Data Retention: inputs and outputs are not stored once the response has been returned. No logs, no cache, no use for model training.

The term is publicly documented in the provider's commercial terms and data processing addendum. For business partners, ZDR is enabled by enterprise contract. It is not active by default on the public consumer API.

Layer 2

LexCodex's own infrastructure

Once the LLM provider has produced the response, it lands at LexCodex. Our own policy applies here: uploaded documents and AI responses are not stored after the analysis is complete. When you close the tab the data is gone. We hold no logs containing document content, no backups of analyses, no archives.

What we do store is session metadata: who was logged in, which tool was used, how long the request was. It is used for rate limiting and debugging. No content, no citations, no conclusions.

What this means in practice

For a lawyer or in-house counsel:

A common misconception

"The vendor doesn't train on my data" is not the same as ZDR. Many consumer AI services have a "no training" term but still store inputs and outputs for several days for quality monitoring or abuse detection. For client material where confidentiality matters, that is not enough.

The distinction in plain language:

Questions to ask an AI vendor

Before making an AI service part of the firm's workflow, it is reasonable to ask:

  1. Do you have ZDR terms from the underlying LLM provider? Is it active on our account?
  2. How long are requests and responses stored at your end before deletion?
  3. Are there backups or archives of user data? For how long?
  4. Which logs are kept? Do they contain content or only metadata?
  5. Can you sign a DPA under GDPR Art. 28?

For LexCodex the answers are public in our DPA. You don't need to ask, you can read them before creating an account.

Read more

How LexCodex avoids hallucinations · EU AI Act for lawyers

Security & Trust → Back to blog