Best AI for Lawyers: What a Small Firm Should Hand Over First

A two-person firm bills for judgment. It does not bill for reformatting a discovery index, or for reading forty pages of medical records to find the three that matter, or for writing the fourth version of the same client status email this month. The gap between the work clients pay for and the work that fills the day is where the question of the best AI for lawyers actually starts, and it is a much narrower question than the marketing around legal tech suggests.

Two things tend to go wrong. Some firms ban the whole category after reading about a sanctioned filing. Others let a general chatbot draft something client-facing and find out the hard way what it invented. Both reactions come from treating this as one decision instead of a set of separate ones, each with a different amount of risk attached.

Where the hours actually leak

Before comparing anything, it helps to know which parts of the week are candidates. In a small firm, the recurring time sinks are usually some mix of:

• Reading a long record, deposition transcript, or document production to find the handful of relevant passages.
• Writing routine correspondence: engagement letters, status updates, scheduling emails, standard demand letters that follow a house template.
• Turning messy intake notes into a structured file the rest of the team can work from.
• Reconstructing timekeeping narratives at the end of a week nobody wrote contemporaneous notes for.
• Summarizing a matter for a colleague, or for yourself, six months after you last touched it.

None of those are the practice of law in the sense a bar would recognize. All of them are billable hours that clients increasingly push back on. That is the honest target.

The citation problem, stated plainly

Any list of the best AI for lawyers that skips this is not worth reading. A general-purpose chat model produces text that resembles its training data. Case citations have a very regular shape, so a model asked for supporting authority will produce something with a plausible case name, a plausible reporter volume, and a plausible year, whether or not that case exists. Courts have sanctioned lawyers over exactly this, and the pattern repeats often enough that it is no longer novel.

The working rule is simple and absolute: never cite anything you have not opened and read in the actual reporter or database. Not the summary the model gave you, the case itself. If that sounds like it removes most of the time savings from legal research, it does, which is why research is the last thing a small firm should hand over, not the first.

Purpose-built legal research products work differently. Tools like Westlaw's CoCounsel, Lexis+ AI, and Harvey retrieve from a real, licensed case database and cite what they actually found, rather than generating citations from scratch. That closes most of the fabrication gap, though it does not remove the obligation to read the case. It also costs meaningfully more than a general assistant subscription, which matters for a firm deciding where to spend first.

Confidentiality narrows the shortlist before features do

Client information is not ordinary business data. Before a matter file goes anywhere near a tool, two questions need answers you can point to in writing.

First, does the plan you are on train on your inputs? Consumer tiers of the major assistants have historically been more permissive here than their business and enterprise tiers, and the terms change. The paid business tiers of Claude, ChatGPT, and Copilot generally commit to not training on customer content, but the commitment lives in the terms of service for your specific plan, not in a blog post about the product.

Second, what does your jurisdiction's bar say? Several state bars have issued formal opinions on generative AI covering competence, confidentiality, supervision, and whether fees can be charged for time a tool saved. These are short documents and reading the one for your state is a better use of an hour than any tool comparison, including this one.

Four jobs worth handing over, ordered by how little a mistake costs

Start where a wrong answer is visible and cheap, not where the time savings look biggest.

Summarizing something you are going to read anyway. A summary of a long transcript, produced before you read it, tells you where to slow down. If the summary is wrong, you find out within the hour, because you are reading the source regardless. This is close to zero-risk and it genuinely compresses a long afternoon.

First drafts of routine correspondence. Client updates, scheduling, standard letters that follow a template your firm already uses. Feed the tool three or four of your own past letters so it writes in your firm's register rather than generic legal-adjacent prose. You edit and send. Nothing leaves without a lawyer reading it.

Structuring intake and file organization. Turning a page of handwritten intake notes into a consistent matter summary, or organizing a document production into a labeled index. Errors here surface quickly and cost minutes, not a client.

Digesting records against a specific question. Not "tell me what this says," but "list every passage in these records that mentions a prior back injury, with page numbers." A question that specific is checkable, because you can go straight to the page numbers and confirm.

The review step that decides whether this saves anything

Here is the arithmetic that gets skipped. A task is only worth handing over when checking the output is faster than doing the work yourself. Drafting a client update from scratch takes fifteen minutes; reviewing an AI draft of it takes four. That is a real gain. Verifying a research memo means reading every cited case anyway, which is most of the original work plus the reading of the memo. That is a loss dressed up as a win.

Apply that test task by task and the list of things worth automating gets shorter and much more reliable. It also protects you from the failure mode where a firm feels faster for a month and then discovers the review was never really happening.

Picking without turning it into a procurement project

For most small firms the sensible order is: one paid business-tier general assistant, used on the four jobs above and nothing client-facing without review. Give it six to eight weeks. Only when someone can point at a specific research bottleneck that a general assistant cannot safely touch does a purpose-built legal research subscription earn its price. Buying the expensive research tool first is the most common way a firm ends up with a licence nobody opens.

The tools also need one thing no vendor supplies: a written rule for your firm about what may be used where, who reviews what, and what never goes into a prompt. One page is enough. Without it, the policy is whatever the least cautious person in the office decides on a busy Thursday.

Most firms find the tool choice is the easy part. The written policy, the template library the drafting actually pulls from, and the intake structure that makes any of this reliable are what take real time to set up properly. That part can be built for you.