Legal has adopted AI faster than its risk profile suggests it should, and the sanctions issued to lawyers who filed briefs containing invented case law are the reason the profession now treats this carefully.
The constraint that makes legal different
In most sectors a confidently wrong output is a quality problem. In legal it can be a professional conduct matter with consequences for the individual practitioner. Verification of any authority cited is therefore non-delegable — not because the model is unreliable, but because responsibility for the filing cannot transfer to software.
Where agents earn their place
- Intake — capturing matter details, conflict-checking against existing records, routing by practice area.
- Document review — first-pass extraction across large disclosure sets, flagging for human attention rather than deciding.
- Drafting from precedent — assembling from the firm's own approved templates rather than from general training data.
- Knowledge access — answering from the firm's matter history instead of sending associates hunting.
- Administrative load — time capture, file notes and correspondence logging.
Privilege and confidentiality
A retrieval system reaching matter files inherits the access model beneath it. If ethical walls exist between teams and the permission model does not enforce them structurally, the system will cheerfully cross them and present the result fluently. Auditing that model is the first task, before any capability discussion.
Why conventional vendor pitches fall short
Legal AI is frequently sold on research speed — the capability with the highest verification burden and the most public failures. The durable return sits in intake, document handling and administrative load, where errors are caught internally and the time recovered is real.
What this means for your firm
Start with intake and document handling, keep verification human, and audit the permission model before anything reaches matter files. Our guide to AI governance covers the review and audit structure this requires.