AI for Lawyers in 2026: What a Small Firm Can Safely Hand Off

AI can now answer a law firm's phones after hours, summarize a matter file, draft a first pass at a contract or brief, suggest time-entry descriptions and help with client email. It can't sign anything, exercise judgment for you, or be trusted with a citation you haven't read. Under the ABA's ethics guidance, every output stays your responsibility.

For a solo or small firm, the hours that disappear aren't the ones in court. They go to intake calls, conflict checks, status emails, scheduling with opposing counsel, and writing up time after the last call. Below, that work is taken one task at a time, with the legal AI tools built for each part, the ethics rules that decide how you can use them, and a plain account of where Merlin helps and where it doesn't.

Where a small practice loses its non-billable hours

Most of a lawyer's job is judgment work. O*NET's profile of lawyers lists interpreting laws and rulings, analyzing probable outcomes using precedent, drafting documents such as wills, deeds, leases and contracts, and negotiating settlements. It also lists supervising legal assistants and performing the administrative and management functions of the practice. The Bureau of Labor Statistics counts about 863,700 lawyer jobs in 2025, with about 11% self-employed, and notes that some lawyers work more than 40 hours a week (BLS). In a small office, the admin around the legal work tends to land on the lawyer:

• Intake and conflicts: the prospective client who calls at 7 p.m., the details you need before a consultation, and the conflict check before you hear too much.
• Engagement: fee agreements and engagement letters, signed and filed before work starts.
• Deadlines and scheduling: response dates, hearing dates, and the back-and-forth with opposing counsel and court reporters over a deposition date.
• Client updates: "any news on my case?" emails that each take ten minutes to answer properly.
• First drafts: demand letters, discovery responses, standard contracts and the cover email that goes with them.
• Time and billing: reconstructing the day's work into entries a client will accept.

AI for lawyers, task by task

Most legal AI is built into software a firm already pays for: practice management, research databases and Word add-ins. General assistants cover the email and scheduling around them.

Intake when you can't pick up

A prospective client with a fresh problem calls the next firm on the list if nobody answers. Smith.ai's legal page describes 24/7 answering that mixes AI with live agents, practice-specific intake questions, conflict checks, consultation booking, and routing that treats prospects, current clients, court staff and opposing counsel differently. It bills per call rather than per minute and doesn't publish prices on that page.

If an AI chatbot or voice agent speaks to prospective clients, say so. The Florida Bar's Opinion 24-1, approved in January 2024, says a lawyer must tell prospective clients they are communicating with an AI program and not a lawyer or staff member. Check whether your own state bar has said something similar.

Research you can verify

Legal research is where AI is most useful and most dangerous. CoCounsel Legal from Thomson Reuters answers questions grounded in Westlaw and Practical Law content, drafts with linked citations, and sells plans for firms under 10 attorneys (CoCounsel Legal). LexisNexis says Lexis+ with Protégé drafts from a single prompt and lets you check authority through linked citations and Shepard's (LexisNexis, August 2026).

Purpose-built doesn't mean error-free. In a 2024 Stanford RegLab and HAI study, Lexis+ AI and Ask Practical Law AI gave incorrect information more than 17% of the time, and Westlaw's AI-Assisted Research more than 34% of the time (Stanford HAI). Thomson Reuters disputed the method, and the products have changed since. Use the results as a reason to pull and read every case, not as a current score.

Drafting and contract review

Contract work suits AI because the firm's own positions can be written down. Spellbook reviews and redlines against playbooks you define and drafts clauses, and says its model providers don't train on or keep your data (Spellbook). Gavel Exec works in Microsoft Word, offers a trial of 25 free queries per user, then costs $160 a month per user for 1,000 completions, according to Gavel's site in October 2026. For litigation documents, MyCase IQ, included with the MyCase Advanced plan, summarizes contracts and pleadings and builds timelines from case data, and MyCase itself warns that output can be inaccurate and should be reviewed (MyCase).

Turning a day's work into time entries

Reconstructing a day from memory loses billable time. Smokeball's AutoTime captures work as you do it, and its AI writes descriptions for pending entries based on your last 100 edits. The feature is in beta and runs from the desktop app only (Smokeball help). Read each entry before it goes on a bill. A description the AI wrote is still a statement you are making to your client.

The ethics rules that decide how you use AI

No rule bans AI in law practice. The ABA's Formal Opinion 512, issued in July 2024, applies the existing Model Rules to generative AI, and state bars have been adding their own guidance since. These are the duties that change how you work day to day:

• Competence, Rule 1.1. Comment 8 already asks lawyers to keep up with "the benefits and risks associated with relevant technology" (ABA Journal). Opinion 512 says you needn't become an AI expert, but you must understand the tool you use and never rely on its output without an appropriate degree of independent review.
• Confidentiality, Rule 1.6. Paragraph (c) requires reasonable efforts to prevent inadvertent or unauthorized disclosure of client information. Opinion 512 asks you to read a tool's terms and privacy policy, and requires informed consent before you put confidential information into a self-learning tool, one that can learn from what you type into it. Boilerplate in an engagement letter isn't enough (summary by Frankfurt Kurnit).
• Supervision, Rules 5.1 and 5.3. Lawyers who manage a firm need clear policies on which AI tools may be used and for what, training for lawyers and nonlawyer staff, and vetting of the vendors themselves (Frankfurt Kurnit). A paralegal's AI-drafted motion is still your filing.
• Fees, Rule 1.5. You may bill the time you actually spent, including reviewing AI output. You may not bill for learning a tool unless the client asked, and AI built into everyday software counts as overhead, not an expense (Frankfurt Kurnit).
• Candor to the court, Rule 3.3. Opinion 512 warns against submitting misstatements of law or fact produced by AI, and says lawyers may not abdicate judgment to a tool.

What happens when citations aren't checked

The best-known case is Mata v. Avianca: in June 2023 Judge P. Kevin Castel of the Southern District of New York fined the plaintiff's lawyers $5,000 under Rule 11 after they filed a brief with nonexistent cases produced by ChatGPT. In February 2025 a federal judge in Wyoming sanctioned three Morgan & Morgan lawyers in Wadsworth v. Walmart over a motion in limine citing eight cases that didn't exist: $3,000 and revoked pro hac vice status for the drafting lawyer, and $1,000 for each of the two who signed without checking (Bloomberg Law).

It keeps happening. Damien Charlotin's database of court decisions addressing AI hallucinations listed 2,217 cases worldwide on October 25, 2026, with 879 involving lawyers rather than self-represented parties (AI Hallucination Cases). Some judges also have their own rules: since May 2023, Judge Brantley Starr in the Northern District of Texas has required a certificate that any AI-drafted language was checked by a person against print reporters or traditional legal databases (Volokh Conspiracy). Check the standing orders of every judge you appear before.

What AI should not do in a law practice

Some parts of practice are judgment, and the rules keep them with a licensed person.

• Advice to a client. Whether to settle, what to plead, what risk to accept. AI can organize the facts; it can't own the recommendation.
• Anything filed or signed. Read every case and quotation in full before it goes to a court or to the other side.
• Deadline math from a chat window. Use rules-based docketing, then check it. A general chatbot guessing a response date is how a default happens.
• The conflict decision. Software can search names. Deciding whether a conflict can be waived is yours.

Where Merlin fits in a law practice

Merlin handles the email and calendar side of a practice. It is a general Mac assistant that suits lawyers whose firm email lives in Gmail, with a Google Calendar beside it, and it remembers each client, opposing counsel and court reporter you correspond with, along with what you've told them. It is not practice-management, research or billing software, it doesn't answer phones or texts, and it doesn't work with Outlook or Microsoft 365. If your firm's mail runs on Microsoft, Merlin won't fit.

• Clients waiting on you. Merlin's People list flags a person who emailed you 2 to 30 days ago and is still owed an answer, so the client asking about their hearing date doesn't wait a week.
• Replies you approve. Ask in Merlin's chat to answer opposing counsel about deposition dates, and the draft appears in the chat for you to change; nothing goes to the other side until you approve it.
• Briefs before a call. About 10 minutes before a meeting on your Google Calendar, a Meeting Brief summarizes earlier email threads with the people attending, and the commitments made in them.
• Recorded calls, on your terms. Merlin can record calls on your Mac, and paid plans transcribe them with speaker labels. Audio stays on your Mac; transcripts and notes are kept with your Merlin account.
• Routines. A routine can, say, list every client email still unanswered each Friday. Routines run without asking each time and can send email if you allow it, so keep them to summaries for client matters.

Treat Merlin the way Opinion 512 tells you to treat any tool. Read its terms and privacy policy, decide which Google accounts to connect, and get informed consent where your analysis says you need it, before client mail flows through it.

First steps with AI for your firm this week

Start small, and put what you decide in a one-page policy, which Opinion 512 expects of firm managers anyway.

1. List the tools already in use. Ask everyone in the office, including staff, which AI tools they use and what client information has gone into them.
2. Read the terms of each one. Check whether inputs are used for training, how long they're kept and who can see them. Drop any tool you can't answer those questions for.
3. Turn on what you already pay for. Your practice-management or research subscription may already include AI summaries or drafting.
4. Pick one task and one rule. For example: AI may draft the first version of a demand letter, and a lawyer reads every cited case before anything is filed or sent.
5. Update your engagement letter and intake script. Explain plainly how you use AI and get specific consent where it's needed, beyond boilerplate. If a bot answers your phones, have it say so.

For more on notes from calls, see our comparison of AI note-taking tools, and if a message to the wrong recipient is your worry, the guide to recalling a sent email.