Skip to content
News

GUIDES · November 19, 2026 · 7 min read

AI in Law Firms: What Actually Works, and What Gets Firms in Trouble

Two risks, and the famous one is the smaller one. Fabricated citations are embarrassing and caught by checking. The confidentiality breach is quiet and is not.

What is already routine

  • •First drafts of correspondence, from your own notes rather than from the file.
  • •Explaining an unfamiliar area of law well enough to know what to look up properly.
  • •Restructuring and tightening your own writing.
  • •Summarising a public judgment or a published standard.
  • •Generating the questions you have not thought of yet.

None of these require a client in the prompt, which is why they are uncontroversial and why they are where most of the real value has turned out to be.

The citations problem

Models produce plausible case references that do not exist. Courts have sanctioned lawyers for filing them, repeatedly and in several jurisdictions, and the pattern is always the same: the citation looked right and nobody pulled the case.

This is a solved problem in the sense that the solution is known and boring. Every authority gets verified in a real database before it goes anywhere. A firm that does that has removed this risk entirely.

The confidentiality problem

Nobody gets sanctioned for this. They get a complaint, eighteen months later.

Pasting a client's file into a general assistant is a disclosure to a third party. Privilege is the client's, not the firm's, and the duty survives the retainer. No convenience argument touches that, and no enterprise agreement changes whether you were entitled to make the disclosure — it changes what the vendor may do afterwards.

And the pattern in every sector is the same: the tool a firm approves is not usually the tool being used. Most pasting happens on personal accounts, where the firm has no visibility and no agreement at all.82% — the share of those pastes made from unmanaged, personal accounts — outside any enterprise oversight. LayerX, Enterprise AI & SaaS Data Security Report

What a workable position looks like

  • •Verify every authority in a real database. Non-negotiable and cheap.
  • •Decide explicitly whether client material may go to a model, and if so which one on what terms — rather than leaving each fee earner to decide in the moment.
  • •Provide something good enough that nobody needs the free tab. The alternative to a good approved tool is not abstinence.
  • •Where client material is involved, remove the identifying detail before it is sent. A document about a substituted person carries no privilege to waive.
  • •Tell clients what you do. Some will care; the ones who do would rather be asked than find out.

That fourth point is what this company builds: identifying detail replaced on the machine, the real names restored in the answer, the file itself never uploaded.

Try Secure AI free

Frontier AI chat — encrypted by default, with your identity stripped before any request reaches the provider.

Not legal advice, and deliberately general. Your regulator, your insurer and your risk partner decide your position.

Common questions

Can lawyers use ChatGPT?

For work that contains no client information, routinely — research framing, drafting from your own notes, tightening your own prose. For client material it is a disclosure to a third party, and the duty of confidence sits with you and survives the retainer.

Why do lawyers keep getting sanctioned over AI?

Fabricated citations. Models generate plausible references that do not exist, and the filings that drew sanctions all share one feature: nobody verified the authority in a real database. Verifying every cite removes this risk completely.

Does an enterprise agreement make client files safe to paste?

It governs what the vendor may do with the file, which is real and worth having. It does not decide whether you were entitled to disclose it. Those are separate questions and the second one is yours.

What about redacting the file first?

That is the approach that holds. A document in which every identifying detail has been replaced is not a disclosure about your client, and the substituted version still reads as a document so the model can actually work on it.

Related

More from News
Secure AI

Use any AI. Stay anonymous.

© 2026 Secure Artificial Intelligence, Inc.Private details never reach the AI.
AI in Law Firms: What Actually Works, and What Gets Firms in Trouble · Secure AI