If you want to know how law firms use AI in 2026, ignore the conference keynotes. The version on stage is a managing partner describing a future that has not arrived. The real version is quieter, narrower, and a lot more useful. At four-to-eight-attorney firms, AI is doing a small number of specific jobs well and a larger number of jobs badly enough that the firm quietly switched them off after a month.
This note is the short, free version of what we found across two of our reports. It will not make you an expert. It should save you from buying the wrong thing.
The jobs that stuck
Three uses show up again and again at small firms, and they have one thing in common: each one replaces a task a paralegal hated doing, not a task that required judgement.
The first is intake triage. When a prospective client fills in a form or leaves a voicemail, the software drafts a first response, runs a preliminary conflict check against the firm's own matter history, and flags the matter type. A human still decides whether to take the case. The software just gets the firm to that decision faster.
The second is document review on a firm's own files. Not the dramatic e-discovery version from the trade press. The boring version: pulling the three relevant clauses out of a forty-page lease so the associate reads three pages instead of forty.
The third is the first draft of routine correspondence. Engagement letters, status updates, the polite chase for an unpaid invoice. The partner edits every one before it goes out. The blank page is the part the software removed.
The jobs that got switched off
Legal research without a human checking every citation is the one that bites. Several firms told us the same story in slightly different words: an associate trusted an AI research summary, a cite turned out to be wrong, and the firm now has a rule that every authority gets opened and read before it goes in a filing. That rule is correct. It also erases most of the time the tool was supposed to save.
Client-facing chatbots that answer legal questions are the other common retreat. The liability is obvious once you say it out loud. A tool that tells a prospective client something close to legal advice, branded with your firm's name, is a problem you do not want. The firms that tried this mostly pulled it back to "we will have someone call you," which is a scheduling tool wearing a chatbot costume.
What separates the firms getting value
The firms getting real value are not the ones that spent the most. They are the ones that picked one workflow, ran it the way the tool was designed, and resisted the urge to bend it back into their old process on day three.
The pattern underneath that is dull and reliable. A firm that adapts its intake process to the software gets the speed. A firm that asks the software to imitate its existing, idiosyncratic intake process gets a worse version of what it already had, plus a subscription fee.
If you take one thing from this note, take that. The deciding factor is rarely the tool. It is whether the firm commits to the new workflow long enough to get past the awkward first fortnight.