Bring Your Own Key: What Zero Data Retention Actually Means for Privilege and Confidentiality

Every legal AI feature, whether review, redlining, extraction, or drafting, works the same way underneath: client documents are sent to an AI model, and text comes back. The legal questions that matter all live in that one sentence. Whose model? Under what terms? And what happens to the document after the answer comes back?

Two phrases keep coming up when firms try to answer those questions: "bring your own key" and "zero data retention." Both sound technical. Both are actually simple, and both matter for privilege and confidentiality. This guide explains them in plain English and gives you the questions to ask any legal AI vendor.

The usual arrangement, and why it worries firms

Most AI tools, from consumer chat apps to many legal platforms, run on the vendor's own account with the AI provider. Your documents go to the vendor, the vendor sends them on to the AI company under the vendor's agreement, and results come back. Three concerns follow.

Storage terms you never negotiated. Consumer AI subscriptions typically keep chat history and may use content to improve their services. When a lawyer pastes an agreement into a personal chat account, the firm has effectively handed client material to a third party under terms nobody at the firm has read.

An extra party in the chain. A confidentiality analysis has to cover every hand a document passes through. When the vendor sits between you and the AI company, that is one more party holding client material.

The training question. Without clear terms saying otherwise, content sent to AI services may be used to train future models. For client documents that risk is simply unacceptable, not because leakage is likely, but because the duty of confidentiality does not run on likelihoods.

Bar guidance across jurisdictions has settled in the same place: using AI is allowed, but lawyers must take reasonable steps to protect confidentiality. That means actually knowing the answers to the questions above.

What "bring your own key" means

Start with the term itself. An "API key" is like an account number: it is how a company opens its own direct account with an AI provider such as Anthropic (which makes Claude), Google (Gemini), or OpenAI (which makes ChatGPT's models).

"Bring your own key" (BYOK) means the legal platform does not send your documents through its own account with the AI company. It uses your firm's account instead. Three things shift in the firm's favor.

1. Your contract governs. The agreement between your firm and the AI provider, the one your firm actually reviewed and signed, controls what happens to document text. Business accounts work under different rules than consumer chat apps: at the major providers, business API content is not used for training by default.
2. Your choice of provider. Firms have preferences, and sometimes client instructions, about where data may go. With your own key, the firm picks the provider and can switch later.
3. Your off switch. Cancel the key and every AI feature stops immediately, under the firm's control, not the vendor's.

One practical note: a platform that holds your key must store it carefully. The key should be stored encrypted, never written into logs, and never visible in the browser. Ask about all three.

What "zero data retention" means

Zero data retention (ZDR) is a setting on the firm's account with the AI provider. With it in place, the provider processes each request and then discards it. What you sent and what came back are not stored, not used for training, and cannot be produced later, because they no longer exist.

For a legal team, the plain-English version is this: ZDR turns the AI model from a filing cabinet into a pipe. Documents flow through and are gone. A subpoena served on the provider cannot produce documents the provider never kept. For a privilege analysis, sending material to a service that processes and discards it under a confidentiality agreement is a far stronger position than depositing it, open-ended, into a third party's data store.

What zero data retention does not do

Honest limits, because vendors often blur them:

• It covers the AI provider, not the legal platform. The platform still stores your documents (it has to, to be useful), and that storage stands on the platform's own security: encryption, access controls, and an audit trail.
• It does not make the AI's output correct. It answers where data goes, not whether the work is good. The duty of competence is untouched.
• It is a setting, not a law of nature. It should be verifiable in the provider agreement your firm holds, which is exactly what having your own key makes possible.

The full picture: your key, no retention, and a record

Confidentiality is about where data goes. Accountability is about what was done with it. A defensible AI setup needs both:

• Your own key puts the provider relationship, the terms, and the off switch in the firm's hands.
• Zero data retention means processed documents are discarded, not accumulated.
• An audit trail records every AI action on the matter, who ran it, what changed, and why, so supervision can be shown rather than claimed.

The three reinforce each other. It is hard to claim careful, confidential AI use when the work ran through a personal chat subscription with default storage and no record of what the AI touched. It is easy when every run used the firm's key, under a no-retention setting, and landed as a named, reviewable entry in the matter history.

Eight questions to ask any legal AI vendor

1. Whose account with the AI provider processes our documents, ours or yours?
2. If yours: under what storage and training terms, and can we read them?
3. If ours: how is our key stored? Encrypted? Ever written to logs?
4. Can we choose the AI provider, and switch later?
5. Is zero data retention set up, and where is that documented?
6. Where do you store our documents, and how are they encrypted?
7. Is every AI action on a matter recorded with a name, the exact change, and the reason?
8. Can we run the software on our own servers if a client requires it?

A vendor comfortable with legal customers will have crisp answers to all eight.

How gitmatter is built

gitmatter runs on bring-your-own-key by design. The firm's key, whether for Claude, Gemini, OpenAI, or OpenRouter, is stored encrypted, used with zero data retention, and never written to logs or exposed to the browser. Every AI action runs inside the matter and lands in the audit trail with a name, a reason, and the exact change. And for firms whose clients require it, gitmatter is open source and can run entirely on the firm's own servers.

If your firm is writing its AI policy right now, the fastest way to pressure-test it is to see this setup working end to end. Book a demo and bring the eight questions.