AI for Legal Work in 2027: What Lawyers Should Use for Each Task (and Where It Fails)

Monday, 8:15 a.m., at a three-attorney firm that does employment, family and small business work. There are 22 new emails since Friday, six of them from prospective clients who filled out the website form over the weekend. A motion to dismiss response is due Thursday, the exhibits run 410 pages, and nobody has opened them yet. Somewhere in the inbox, a client is asking for the third time whether "anything happened" on her case. That is the moment when "should I use AI for legal work?" stops being a CLE topic and turns into a scheduling question.

The short version: yes, but not the way the vendor demos sell it. The AI that actually earns its keep at a small firm does not write your brief for you. It reads fast, puts facts in order, builds a first outline and cleans up prose. The part that can get you sanctioned or reported to the state bar (case citations, deadlines, the numbers in a settlement demand) stays with you. This guide walks through a normal week at a small firm, task by task, and tells you what I would use for each one, what it costs and where it tends to go wrong.

AI for legal work: what it does well and what it does badly

One rule organizes everything else in this guide. AI for lawyers is strong at reading and organizing and weak (sometimes dangerous) at recall and calculation.

• Does well: summarizing a deposition transcript, pulling every claim and request for relief out of a complaint, flagging risky clauses in a commercial lease, rewriting a paragraph that runs nine lines, turning a client's rambling email into a dated timeline, drafting deposition outlines from a set of documents.
• Does badly: citing cases from memory, giving you a reporter citation you have not pulled yourself, counting court days under your local rules, computing damages with several moving parts, knowing about a statute amended after the model's training cutoff.

If you keep only this list, you already avoid the two mistakes that keep showing up in the news about lawyers getting scolded by federal judges.

The firm's week, task by task

I built this table around a firm of one to three attorneys handling litigation plus some transactional work, which is the most common profile of someone searching for AI for lawyers in the US. The times are estimates from people who do this work every week, not a study.

Notice that the "what stays with you" column is never empty. That is not distrust of the tool. Your signature on a filing carries Rule 11 (or your state's version of it), and no terms of service from any AI company change that.

Reading long records: the best use by far

If I had to pick a single task to start with, it would be this one. You upload the deposition or the exhibit binder (after stripping what does not need to be there, covered in the ethics section), ask for a structured summary, and you get an hour back in the first week. The prompt I use:

You are assisting a defense attorney in an employment case in [STATE]. Read the attached deposition transcript of the plaintiff and give me:
1. Employment facts: job title, start date, end date, how the employment ended, pay rate the plaintiff claims
2. A numbered list of every adverse action the plaintiff describes, with page:line
3. Every person the plaintiff names as a witness or decision maker, with page:line
4. Statements that contradict the complaint (I will paste the complaint below)
5. Dates I should verify by hand
Rules: cite page:line for every fact. If something is not in the transcript, write "not found". Do not cite case law.

I ran it on a made-up 96-page transcript built for this test. Real output, pasted without edits:

1. Employment facts: Plaintiff worked as an inventory coordinator at [COMPANY] from June 6, 2022 (14:3) until April 18, 2026 (14:9). She testified she was "let go on a phone call" by the regional manager (51:12). Claimed rate: $24.50 per hour (15:20).

2. Adverse actions described: (1) Removed from the weekend lead rotation in January 2026 (33:4 to 33:19). (2) Written warning for tardiness on February 11, 2026 (38:2). (3) Denied a transfer request to the Plano location (41:7). (4) Termination on April 18, 2026 (51:12).

3. People named: regional manager [NAME] (22:8, 51:12); shift supervisor [NAME] (33:6); HR generalist [NAME] (44:15); coworker [NAME], described as present for the February incident (38:21).

4. Contradictions with the complaint: Complaint paragraph 19 says the written warning was issued after she reported the safety issue. At 38:2 to 38:9 she places the warning on February 11 and the safety report "sometime in March." Complaint paragraph 24 says she never received a performance review; at 29:14 she says she "got one review, in 2024, and it was fine."

5. Dates to verify: February 11, 2026 (warning) against the date of the safety report; April 18, 2026 (termination) against the EEOC charge filing date.

Items 1 through 3 I would have found myself, in 90 minutes instead of ten. Item 4 is the kind of thing that slips past you at 11 p.m.: the sequence of events in her own testimony undercuts the retaliation timeline in the complaint. I still checked every page:line. In an earlier test the model wrote 41:7 for something that was on 42:7. With the transcript open next to it, that takes a minute to fix.

Complaints, motions and oppositions: structure yes, final text no

For drafting, "write my opposition" is the wrong request. Ask for the skeleton: each argument the other side makes, the record cite that answers it, the points where your client is weakest, and a list of propositions you will need authority for. Then you write, faster and in your own voice, which the judge in your county may already recognize. The full prompts for complaints, oppositions, discovery and demand letters are in the post with AI prompts for lawyers.

Case law: where AI can get you in real trouble

This is the serious risk. Language models invent cases that look perfect: plausible party names, a reporter citation in the right format, even a pin cite. In Mata v. Avianca (S.D.N.Y. 2023), two New York lawyers were sanctioned US$ 5,000 for filing a brief with six fake decisions that ChatGPT had produced. Courts around the country have since sanctioned others for the same thing, and many federal judges now have standing orders that require you to disclose or certify AI use in filings.

The practical rule: use a general chatbot to suggest search terms and possible lines of argument, then do the research in Westlaw, Lexis or a free database like CourtListener. If a legal research tool with its own case database shows you a case, open the full opinion and read the relevant part before you cite it. Every time.

Deadlines and math: do not delegate

Calendaring deadlines and computing damages are not chatbot work. The model gets court days wrong, does not know your local holidays or your judge's standing order, and adds badly when there are many line items. For deadlines, use your docketing software and the rules themselves. For damages and support figures, your spreadsheet or the state's official calculator. AI can explain the result to the client afterward. It should not produce it.

Which AI tools for lawyers are worth it (with prices)

Prices checked on official sites in late 2026, in US dollars as shown there. They change, so confirm before you sign up.

If the budget covers only one tool, I would take Claude Pro and use NotebookLM for free on records. ChatGPT and Claude do nearly the same job, and paying for both rarely makes sense for a solo. I would not buy a "legal AI" subscription from a vendor you never heard of before testing it on one of your own closed files. Many of them are a general model with a new logo and a higher price.

Ethics: what the ABA and your state bar expect

No US jurisdiction bans lawyers from using generative AI. What applies are the duties you already have. In July 2024 the ABA published Formal Opinion 512 on generative AI tools, and several state bars (California, Florida, New York, Texas and others) have issued their own guidance. They point to the same Model Rules:

• Rule 1.1, competence. You need a reasonable understanding of what the tool can and cannot do. Comment 8 already asked you to keep up with technology.
• Rule 1.6, confidentiality. Before you paste client information into a tool, you need to know whether the vendor stores it, trains on it or shares it. For some uses, the ABA opinion says informed client consent is needed.
• Rule 1.4, communication. Depending on the matter, you may need to tell the client you use AI, especially if it affects fees or how their information is handled.
• Rule 1.5, fees. You can bill for the time you actually spent, not the time the task "would have taken." Charging the client for learning a tool you will use across all cases is a problem.
• Rules 5.1 and 5.3, supervision. AI output is treated like work from a nonlawyer assistant. You review it before it leaves the office.
• Rules 7.1 and 7.3, advertising and solicitation. Marketing copy written by AI still cannot be misleading or promise results.

In practice, that turns into five habits:

1. Redact before you paste. Replace names with [CLIENT], Social Security numbers and account numbers with [ID]. Legal analysis almost never needs them.
2. Turn off training. ChatGPT and Claude both have a setting so your conversations are not used to improve the model. Business and enterprise plans come with it off and with contractual terms you can show a client.
3. Be extra careful with sensitive matters. Medical records, family law and immigration files are full of information that should never sit in a consumer account.
4. Put it in the engagement letter. One sentence saying the firm uses AI tools with attorney review covers transparency for most matters.
5. Check your judge. Before filing, look for a standing order on AI in that courtroom.

Common mistakes when a firm starts using AI

• Using the chatbot as a case finder. The number one mistake and the one that ends up in the news.
• One-line prompts. "Write a motion to compel" gets you generic filler. Give the role, the jurisdiction, the format and what it must not do.
• Pasting the whole file with client names into a free account without checking the data settings.
• Forgetting the jurisdiction. If you do not say "Texas state court" or "Ninth Circuit," you get a blend of rules from everywhere.
• Letting the tool set the tone. The client hired you. If an email sounds like a machine, they notice.

The bigger win is not in drafting, it is in intake

After a few months of using AI for writing, a different number shows up. The hours that disappear at a small firm are not in the briefs. They are in intake: website forms from people outside your practice areas, calls asking "how much for a consultation," prospects who send four paragraphs about their divorce at midnight. A well-built AI agent does the screening, sorts by practice area, collects conflict-check information, asks for the right documents and hands you a two-line summary. I walk through the whole setup, with monthly cost and the exact instructions the agent uses, in the post on ChatGPT for lawyers.

A 30-day plan to get started

If you have never used AI in an organized way, this is the path I would follow:

1. Week 1: pick one tool (Claude or ChatGPT), turn off training on your data and use it only to summarize depositions and exhibits. Compare each summary with your own read.
2. Week 2: add fact timelines for complaints and a clarity pass on your own drafts.
3. Week 3: build a folder with the ten prompts that worked. The AI prompts for lawyers on this blog are a good starting point.
4. Week 4: count how many intake emails and calls last month did not need a lawyer. If it was more than 40, it is time to look at an intake agent.

A month in, the question changes from "does AI for legal work actually help?" to "what else can I get off my desk?" That is the right question.

Advertisement

Want to skip this part?

Ready-made AI agents for law firms: nothing to build, no prompts to write, they already work for you.

• Screens new clients from website chat and texts, by practice area, around the clock
• Collects conflict-check names, asks for the right documents and books the consultation on your calendar
• Flags emergencies and hands you the conversation, and never gives legal advice

I want the ready-made agents

Step by step

1. Pick one tool and turn off training on your data
2. Use it only to summarize depositions and exhibits in week one, and compare with your own read
3. Add fact timelines and a clarity pass on your drafts
4. Save the ten prompts that worked in a shared folder
5. Count intake contacts that did not need a lawyer and decide on an intake agent