What zero data retention means

What Is Private AI for Law Firms, and What Does Zero Data Retention Actually Mean?

Updated November 16, 2026

Short answer: "Private AI" has no fixed meaning, so judge a tool by three things: whether anyone trains on your data, whether the vendor stores it, and whether the AI model provider behind the vendor stores it. "Zero data retention" usually describes only the last of these. A legal AI platform can truthfully say it has zero data retention with OpenAI or Anthropic while storing your prompts, documents and answers on its own servers, with backups. This page explains the terms, shows the pattern in vendors' own words, and says what to ask before you rely on any of them.

Three promises that are easy to confuse

Vendors use three kinds of privacy promise. They are different, and a tool can make one without the others.

Promise What it means What it does not mean
No training Your data is not used to train or improve an AI model That your data is not stored
Vendor retention control You, or your admin, choose how long the vendor keeps your data That anything is deleted straight away, or at all
Zero data retention A named party keeps no copy of the request or the answer after processing it That every party in the chain keeps nothing

The third promise always has a subject. "Zero data retention" with whom? The answer is usually "the model provider", the company whose AI model the vendor calls.

What zero data retention means, in vendors' own words

Spellbook gives the clearest description of the usual arrangement: "Spellbook has negotiated agreements with both OpenAI and Anthropic for zero data retention (ZDR). This means customer data included in requests and responses with these LLMs is not persisted and exists only in memory in order to process a request" (Spellbook security).

That is a promise about OpenAI and Anthropic. In the same place, Spellbook says: "Spellbook uses cloud providers with data centers in Canada and US for storing and processing customer data." "Other third party services are used that may store or process customer information." So the model providers keep nothing, and Spellbook and its subprocessors store your data.

Zero data retention is also not something a firm can switch on for most chat products. Anthropic says its "Claude Teams and Claude Enterprise product interfaces" are "not ZDR-eligible", and that for its APIs, "To request ZDR for your organization, contact the Anthropic sales team. ZDR is enabled per organization" (Anthropic). OpenAI offers it on eligible API endpoints "subject to prior approval by OpenAI", and not for the ChatGPT apps.

The pattern: zero retention upstream, storage at the vendor

Across the legal AI vendors we checked on November 11, 2026, the same pattern appears. The model provider keeps nothing; the vendor keeps your content.

CoCounsel (Thomson Reuters). Upstream: "All interactions with GenAI LLM providers use zero-retention API calls", and "where applicable, we have established systemic controls to turn off third parties' abuse-monitoring solutions". At Thomson Reuters: retention is "customized to fit customer needs and set up at the organization account level with the support of a Thomson Reuters client manager. We then delete user content and prompts in accordance with the customer's retention policy." And "We host backup data in separate availability time zones in the U.S." (Thomson Reuters)

Clio's built-in AI and Vincent by Clio. Upstream: "AI features process files under zero data retention agreements", and Vincent uses "zero-retention agreements with large language model providers". At Clio: your matter data is stored by design, because Clio is your system of record. "stored data is encrypted with AES-256", and "We perform regular automatic data backups, use geo-redundant AWS infrastructure across multiple regions" (Clio security).

Spellbook. Upstream: zero data retention with OpenAI and Anthropic. At Spellbook: storage in Canada and the US, plus other third-party services, as quoted above.

None of this is hidden or wrong. A research tool that saves your work, a drafting tool with project history, or a practice-management system that holds your files must store them to do its job. The point is that "we have zero data retention" on its own does not tell you whether your client's content is stored. It tells you about one company in the chain.

Where the other tools sit

  • Harvey and Legora say they do not train on your data and let you set retention, but neither offers zero data retention on its public pages. Legora encrypts data "at rest with AES-256", so it is stored. Harvey says you can "set retention policies, delete data anytime" (Harvey, Legora).
  • Lexis+ with Protégé says "We never use customer data to train AI models and maintain robust data retention and deletion policies", stores documents in "secure Vaults", and states no zero data retention arrangement on the page we read (LexisNexis).
  • ChatGPT, Claude and Copilot business plans store chats, with retention your admins can control on some plans. The ChatGPT Enterprise page quotes each.
  • Tools that clear client content sit at the other end. Faradex holds files and conversations in memory, and they are gone when the Workspace is reset. Its AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them. Proton says of Lumo for Business: "After Lumo processes your query and generates a response, the data is erased" (Proton).

So what is "private AI" for a law firm?

Our working definition: a tool where your firm can answer, in writing and from the vendor's own terms, four questions about client content.

  1. Does anyone train on it? Every vendor we checked says no by default. Get it in the contract.
  2. Who stores it, and for how long? The vendor, the model provider, both or neither. Include backups.
  3. Who can see it? Which of the vendor's staff, for what reasons, under what process. Legora, for example, says access is "only granted to engineers with written customer approval for support-related issues."
  4. Who could be made to produce it? If content is stored anywhere, someone could be ordered to preserve or produce it. The subpoena page explains why that matters.

A tool can be private enough for your firm and still store content, if you are comfortable with who stores it and for how long. What matters is that you know the answers before a client's information goes in.

What to tell clients, and where the ethics rules come in

The ethics side of this, including when a client's informed consent is needed and why boilerplate in an engagement letter may not be enough, is covered on our companion site. See ABA Formal Opinion 512 explained, confidentiality and privilege and state-by-state rules. The ABA Model Rules bind no one until a state adopts them, so your state's rules control.

What this site adds is the vendor side: what each tool actually keeps. Two practical points for engagement letters and AI policies:

  • Describe the tool by what it keeps, not by its marketing label. "We use an AI tool that does not train on client data and deletes conversations after 30 days" is a statement you can check. "We use private AI" is not.
  • Keep your own file. Whatever the tool keeps, your firm's client file is your responsibility. With a tool that clears client content, everything you must keep has to be saved to your own systems.

Questions to send before you buy

For the other term vendors use, see What is confidential AI? The 12 questions to ask a legal AI vendor turn this page into a list you can send to any vendor, each tied to what ABA Formal Opinion 512 asks lawyers to understand.

This guide is general information, not legal advice. Check each vendor's current terms on its own pages before you buy.

About Faradex. Faradex is a private AI workspace for small and mid-sized firms that handle confidential client work. Each firm gets its own Workspace on dedicated infrastructure in the United States. Files and conversations are held in memory and are gone when the Workspace is reset. AI model: Currently Opus via AWS Bedrock, more information is available on our website. AWS does not store prompts or responses, and model providers have no access to them. Pricing: from $120 a month, no seat minimum; see faradex.ai/pricing.