AI for Lawyers (2026): The Four Jobs It Does, the One It Never Does, and Which Plan Keeps Client Data Out of a Training Set

TL;DR: Most "AI for Lawyers" Advice Is Written by a Company Selling an AI Lawyer

AI for lawyers, as of 2026, is four jobs and a rule. The four jobs: drafting from your own templates, first-read summaries of documents you will read anyway, research drafts you verify authority by authority, and intake and follow-up. The rule: nothing with a citation in it leaves the firm until a human has opened every case. The plan you buy decides the data terms (Claude Team and ChatGPT Business do not train on your data by default; consumer plans can), and the never-paste list is shorter than you fear and longer than a vendor will tell you. That is the whole page. The rest is the detail a solo or small-firm owner-attorney actually needs, from someone who installs this in owner-led businesses for a living and has read the sanctions orders so you do not have to.

Scope note: this is the attorney's own workflow. The firm-level question (what to install first, intake, the phone) is on the AI for law firms page, and the Claude-specific version (plans, legal plugins, connectors) is Claude for lawyers.

The four jobs AI does for a lawyer every day

1. Drafting from your templates. Engagement letters, status updates, fee reminders, demand letters that take positions the firm has taken a hundred times. You give the assistant the template, the facts of this matter, and the firm's voice, and it gives you a first draft in the time it took to describe it. This is the highest-volume, lowest-risk job in the practice and the one to start on, because the output is something you were going to write anyway and will read before it goes out.

2. First-read summaries. A three-hundred-page production, a lease with fourteen amendments, a deposition transcript. The assistant produces the map: parties, dates, defined terms, the sections that touch your issue. You still read the document. You read it faster, and you read the right parts first. Anthropic now sells this shape explicitly, with document-management connectors and Word and Outlook add-ins on the business plans; the Claude for lawyers page covers which plan carries which.

3. Research drafts. Not research. Drafts. The assistant proposes the frame, the likely authorities, and the counterarguments, and you verify every authority in a citator before it appears in anything with your bar number on it. Treated this way it is a junior associate's first pass; treated as a finished memo it is the thing the next section is about.

4. Intake and follow-up. The web form, the voicemail transcript, and the referral email become a structured intake note (matter type, urgency, conflicts check fields, next step) and a draft reply for your review. In a solo practice this is the job that moves revenue, because the consult that gets a same-day reply is the one that signs.

The one job it never does: the authority you did not open

In 2023 a federal court in New York sanctioned attorneys who filed a brief citing cases that did not exist, produced by a general-purpose assistant and never checked; the case is Mata v. Avianca, and it has been cited in every bar CLE on the subject since. The lesson is not "do not use AI." The lesson is that a language model produces text that reads like authority whether or not the authority exists, and the only defense is the one you already owe your client: open the case, read the holding, check that it is still good law. The verify rule is mechanical and it is absolute: no citation, quotation, statute, or rule number goes into a filing, a letter, or a client email until a human has confirmed it in a source you would cite in court. Build it into the workflow (a checklist line, a second reader, a citator pass) rather than into your intentions.

The confidentiality question is a plan question

ABA Model Rule 1.6 does not care which vendor you use; it cares whether you took reasonable steps to keep client information confidential, and ABA Formal Opinion 512 (July 2024) applies that to generative AI directly. In practice, as of 2026, the decision is made at checkout. The consumer plans at both major vendors (Claude Free, Pro, and Max; ChatGPT Free, Go, Plus, and Pro) can use conversations to improve models unless you turn that setting off. The workspace plans do not by default: Claude Team and Enterprise, ChatGPT Business and Enterprise, and both APIs. A solo attorney who pays $20 a month for a consumer plan and pastes a client's medical records into it has a Rule 1.6 problem that the $20 business seat would have prevented. The Claude Team plan covers two to one hundred and fifty seats at $20 a seat billed annually ($25 monthly), which is the smallest firm Anthropic will sell a workspace to, and it is the plan I put small firms on.

The never-paste list, even on a business plan: anything under a protective order or filed under seal; anything a client has not consented to share with a vendor under your engagement terms; anything covered by a separate regime with its own contract, which for health information means a HIPAA business associate agreement rather than a plan setting; and credentials, account numbers, and Social Security numbers, which no draft needs. Everything else in an ordinary matter is the material the workspace plans are built to hold.

General assistant or legal-specific tool?

Ask an assistant "what is the best AI for lawyers" and it names CoCounsel Legal, Westlaw and Practical Law, Spellbook, and Harvey. Those are real tools for real jobs: research inside a citator, contract review inside Word with a clause library, drafting inside a document-management system with the firm's precedents indexed. They are sold by sales teams, priced by quote, and bundled into research subscriptions a solo practice may already pay for. The general assistant on a business plan does the four jobs above for $20 a seat and does them across every matter type. The honest sequence for a small firm is general assistant first, for the drafting and intake volume, then one specialist tool for the single workflow where you bill enough hours to notice the difference, usually research or contract review. Firms that buy the specialist tool first tend to use it for the two hours a week it fits and go back to typing for the rest.

What it costs

As of September 2026, from the vendors' own pricing pages: Claude Pro is $20 a month ($17 on annual billing) and includes Cowork; Claude Team is $20 a seat billed annually or $25 monthly, for teams of two to one hundred and fifty; ChatGPT Business is $20 a user billed annually or $25 monthly with a two-user minimum. The legal-specific tools publish no list price. For a solo attorney the whole workflow on this page costs less than one billable hour a month, which is why the question is never the price; it is whether the verify rule survives the third week.

The 20-minute setup that survives the third week

1. Buy the business plan, not the consumer one. Two seats if you have an assistant or a paralegal; the no-training default is the point.
2. Write a two-line firm policy. Line one is the never-paste list above. Line two is the verify rule. Put both where the assistant is opened.
3. Load the firm's voice and templates. A project or a workspace folder with your engagement letter, three status updates you actually sent, and the demand letter you are proudest of. The drafts get better the moment the assistant has read how you write.
4. Start with follow-up drafting, not research. Highest volume, lowest stakes, no citations. Two weeks of client updates and fee reminders prove the tool and the habit.
5. Count the hours. Each Friday, tally the drafts you approved and what typing them would have cost. Three weeks in, that number decides whether the second job (first-read summaries) gets added or the whole thing quietly dies.

Will AI replace lawyers?

Not the part clients pay for. Judgment, the license, the standing to sign the filing, and the liability that comes with it stay with the attorney; what moves is the typing and the first read. The practical effect in a small firm, as of 2026, is that the same hours go to higher-value work and the response time on intake drops from days to the same afternoon. The risk is not replacement. It is the attorney who treats a first draft as a finished one, and the firm that puts a consumer plan in front of a paralegal with no policy. Both are avoidable in twenty minutes, which is what the list above is.

For the solo or small-firm owner reading this

This works or quietly stops working on three things: the plan (business, not consumer), the policy (two lines, visible), and the first workflow (follow-up drafting, not research). If you would rather have it installed than assemble it, that is what the Claude for Small Business install does for owner-led firms: the workspace, the connectors, the firm's voice loaded, the policy written, and a handoff in about two weeks, from $4,500; the $397 Playbook is the do-it-yourself version, and the $2,500 Assessment is the written plan for a firm that wants to see the whole map first. If the question is broader than the tool, I answer it as an AI consultant for owner-led businesses, and the first call is free.