Law firms and legal teams can license de-identified legal work product to AI labs, which want contract drafts, redlines and the record of how a negotiation moved, but nearly all of it belongs to clients. Under ABA Model Rule 1.6 a lawyer needs the client's informed consent to reveal information relating to a representation, and we never touch privileged material unless the client has consented in writing.
This is general information, not legal advice. Your own ethics counsel and your state bar's rules decide what your firm can do.
What labs want from legal work
Labs building AI for legal work have plenty of public law: statutes, court opinions, published filings. What they lack is the private side of practice, the documents that show how lawyers actually draft and negotiate.
- Drafts and redlines. Successive versions of a contract with every change tracked, which shows what was proposed, what the other side struck and what survived.
- Clause choices. Which positions a firm takes on indemnity, liability caps or termination, and how often it gives ground.
- Matter history. How a matter moved from intake to close, as events in order, without the content of advice.
- Templates and playbooks. A firm's own precedents and negotiation guides.
A sequence of redlines with a final signed version is an episode: a starting position, a series of moves and an outcome. That structure is what makes legal work product useful for training, and it rarely exists anywhere public. The legal industry page shows an example record before and after scrubbing.
Confidentiality: Rule 1.6
ABA Model Rule 1.6 says a lawyer shall not reveal information relating to the representation of a client unless the client gives informed consent, the disclosure is impliedly authorized to carry out the representation, or a listed exception applies. Licensing documents to an AI lab is not part of carrying out the representation, so informed consent is the route. Most states have adopted a version of Rule 1.6, with local differences.
The duty is broad. It covers information relating to the representation whatever its source, which is wider than privilege, so a document can be confidential under Rule 1.6 even when no privilege attaches to it.
What ABA Formal Opinion 512 adds
In July 2024 the ABA issued Formal Opinion 512 on lawyers' use of generative AI. It applies the existing duties of competence, confidentiality, communication and reasonable fees to AI tools. On confidentiality, it says a lawyer must get the client's informed consent before inputting information relating to the representation into a self-learning AI tool that could disclose it to others.
Opinion 512 is about lawyers using AI tools, and selling documents to a lab goes further. If the ABA expects informed consent before client information goes into a tool that might learn from it, deliberately licensing that information for training clearly needs consent too.
Privilege
Attorney-client privilege and work product protection are separate from the duty of confidentiality, and they carry an extra risk. Disclosing privileged material to a third party can waive the privilege, which can hurt the client in litigation long after the data deal is forgotten.
Our rule is simple: we never touch privileged material unless the client has consented in writing. In practice, most firms leave privileged communications out entirely and work with material where the client's consent is clear and the privilege risk is low, such as executed agreements, redline histories and the firm's own templates.
How firms get client consent
A workable consent process looks like this:
- Choose the matters and document types you would include. Start with transactional work, where documents are drafts and agreements, and leave out litigation files and privileged advice.
- Write a specific consent: which documents, used for training AI models, de-identified before they leave, the kinds of buyers allowed and the term.
- Get each client's informed consent in writing before any of their documents are included.
- Record every consent. Honor withdrawals for future batches, and say plainly that documents already licensed can't be removed from a trained model.
Clients who decline are left out. Labs check rights in diligence, and a smaller dataset with clean consent is worth more than a larger one with doubtful rights. During our rights review we read engagement letters and outside counsel guidelines with you, because some corporate clients' guidelines forbid any secondary use of their documents.
Your own work product
Some material belongs to the firm rather than any client: its own templates, clause libraries, playbooks and internal training materials. These still need scrubbing, since precedents often began life as a client document, but the client consent issue may not arise. Check each item's origin before including it.
How legal documents get scrubbed
Scrubbing runs on your machine; raw documents never leave it through our SDK. Legal documents are dense with identifiers, so the scrub works harder than it does on structured data.
| Content | What happens |
|---|---|
| Party names, people, signatories | Replaced with consistent pseudonyms such as PARTY_A |
| Addresses, emails, phone numbers | Removed |
| Deal values and fees | Rounded into ranges |
| Dates | Generalized to the month or quarter |
| Matter and case numbers | Replaced |
| Distinctive facts | Generalized or the document is excluded |
Distinctive facts are the real risk. A merger between two companies in a niche industry can be identified from the deal structure alone, even with every name removed. After scrubbing, we test whether any document can be singled out from a combination of details, and documents that can are removed. You review the scrub report before anything is offered. The de-identification vs anonymization guide explains why names alone aren't enough.
What legal documents sell for
We found no public price list for law firm work product licensed for AI training, and no disclosed deal between a law firm and a lab with a stated price. We won't invent one. Our calculator gives an indicative range for contracts and case files based on years of records, team size and how clear your rights are, and we confirm it on a call. What AI labs pay for legal data collects the price signals that exist from nearby markets.
The recurring part
A practice keeps producing drafts. After the first license sells the history, the SDK can run monthly, pull only documents and versions created since the last run, scrub them on your machine and upload a new batch. Each license with a refresh term pays per accepted batch, and the SDK needs no work from you once scheduled. Client consent still governs each matter, so new matters join a batch only when consent is on file. See recurring data revenue.
Find out what your work product is worth
The valuation calculator takes two minutes; pick "Contracts & case files". We check engagement letters and consent on the call, before any document moves. For the full process, start at selling data to AI companies.
Frequently asked questions
Can a law firm sell client documents to an AI company?
Only with the client's informed consent under Rule 1.6 or your state's version of it. Licensing documents for training is not part of carrying out the representation, so implied authorization doesn't cover it.
Does de-identifying a document remove the need for consent?
You shouldn't rely on that. Rule 1.6 covers all information relating to the representation, and identifying details can survive scrubbing in the facts of a matter. Ethics counsel should decide, and getting consent is the safe course.
Will you handle privileged documents?
Not unless the client has consented in writing. Disclosure to a third party can waive privilege, so most firms exclude privileged communications and work with executed agreements, redline histories and their own templates.
What does Opinion 512 say about this?
ABA Formal Opinion 512 says lawyers need informed client consent before putting client information into a self-learning AI tool. It addresses lawyers' use of AI tools, and licensing documents for training goes further, so consent is plainly needed.