Skip to content
Yaakov Tzedek
Guides

Claude for Lawyers, the Complete Guide: From Pleadings to an Automated Practice

79% of legal professionals already use AI, yet more than half say their firm has no AI policy. This guide shows how a law firm builds a sound way of working with Claude: confidentiality and privilege first, then a Skill for pleadings, then automated correspondence, with the professional-ethics rules and the fake-citation sanctions that have already cost lawyers tens of thousands of dollars.

By Yaakov Tzedek · · 14 min read

Claude for lawyers: the complete guide to ethics, documents and safe practice

Lawyers are the profession that gains the most from Claude, yet most firms are still using it the wrong way. According to Clio's Legal Trends Report, 79% of legal professionals already use AI in some form, but 53% say their firm has no AI policy or they are unaware of one: no confidentiality rules, no method, and real exposure on ethics and privilege. This guide fixes that: how a law firm builds a sound way of working with Claude, from the professional-ethics rules, through a pleadings Skill that saves hours on every matter, to automations that manage correspondence. Everything is current as of November 2026, with sources.

Claude for lawyers: the short version

Claude, from Anthropic, is an AI assistant that excels at exactly what legal work is made of: long texts, precise drafting and work on documents. Used properly, it cuts a first draft of a pleading from hours to minutes, reviews a contract and flags risks, distills a year of correspondence into a page, and prepares demand letters on the firm's letterhead, while the lawyer remains the one who checks, approves and signs. Three conditions: an account that does not train on your data, input rules that respect privilege, and verification of every authority before filing.

What do the numbers say about lawyers and AI?

The legal market has passed the point of no return, and the data tells a consistent story:

  • 79% of legal professionals use AI in some form, Clio Legal Trends Report (2025, the same share as in 2024)
  • 30% of US lawyers use AI in their practice, nearly triple the 11% in 2023, according to the ABA's 2024 Legal Technology Survey Report
  • And on the other side: 53% say their firm has no AI policy or they do not know of one, Clio. Clio's 2026 report on solo and small firms puts the no-policy share at 57% for solo firms and 55% for small firms.

That gap, between wide use and no method, is exactly where a firm that works properly gains an edge over competitors.

What changed in 2026

Two things sharpened this year, and both change how a firm should work. First, the Claude desktop app with Cowork became the practical entry point for most firms, because it works on a real case folder on your computer, reads every document in it and prepares drafts without pasting anything into an open chat. (Anthropic is now folding Cowork into a single unified Claude desktop experience, so the label may look different in your app.) Second, sanctions for fabricated case law kept climbing (more than 2,300 documented cases worldwide), which turned "the AI drafts, the lawyer is responsible" from a recommendation into an operational requirement.

What that means in practice for a firm starting now:

  • Pick one environment: the desktop app for case work, not the browser. Build your Skills and connectors there and work on the folders there.
  • An account that does not train: a business account (Team or Enterprise), or training turned off on any personal account, before the first document goes in.
  • A standing verification layer: every authority is checked against Westlaw, Lexis or your official database before filing, and the standing instructions say "when there is no source, write 'needs verification'."

If you want this as an organized method rather than trial and error, start with our courses or book a consulting session to tailor it to your firm. For the advanced automation layer, see our guide to Claude Code skills.

Confidentiality and privilege first: what do the ethics rules say?

Bar associations and regulators around the world have issued guidance on generative AI, and the core principles are remarkably consistent. The most cited is the American Bar Association's Formal Opinion 512, "Generative Artificial Intelligence Tools" (July 29, 2024). It applies the existing Model Rules to AI:

  1. Competence (Rule 1.1). You need a reasonable understanding of what the tool can and cannot do. You do not need to become an expert, but "the machine did it" is no defense.
  2. Confidentiality (Rule 1.6). The duty to protect client information extends to AI tools. Before entering information about a representation into a "self-learning" tool (one that may use inputs to train or improve itself), the opinion says you should obtain the client's informed consent, and a generic boilerplate clause in an engagement letter is not enough.
  3. Communication (Rule 1.4). In some situations you must tell the client you are using AI, for example when it affects the representation or the client asks.
  4. Supervision (Rules 5.1 and 5.3). Firm leaders must set clear policies and train lawyers and staff on permitted use.
  5. Candor and meritorious claims (Rules 3.1, 3.3). Every output is checked before it reaches a court.
  6. Fees (Rule 1.5). You may bill for the time you actually spend, not for time the AI saved, and general AI tools are usually overhead rather than a separate charge unless agreed in advance.

A brief note on jurisdiction: ABA opinions are influential but not binding. Many US state bars, regulators in other countries and national bar associations (for example, the Israel Bar Association's ethics committee) have published their own guidance along the same lines: full responsibility stays with the lawyer, every output is verified, client-identifying details are protected, and client consent is recommended or required. Check the rules that actually govern you.

What happens to the information you enter into Claude?

This is where the account type is critical. According to Anthropic's Privacy Center, the picture as of November 2026 is:

Account typeIs your data used for training?
Consumer (Free, Pro, Max)Only if the Model Improvement setting is on. Check your privacy settings and turn it off
Team and Enterprise (commercial)No by default. Commercial terms do not use inputs or outputs for training
API (systems the firm builds)No. Inputs and outputs are deleted within 30 days and not used for training

Two caveats worth knowing: on any plan, conversations you rate with thumbs up or down can be kept for up to 5 years and used for training, and content flagged by safety systems can be kept longer.

The working rules we implement in firms: a business account (or at the very least training turned off in the privacy settings on every machine); replacing names with labels ("the client", "the opposing party") in anything sensitive; removing ID numbers, addresses and account details when the analysis does not need them; and an informed-consent process with clients that goes beyond a boilerplate line in the engagement letter where your rules require it. In most cases that is all you need to work both correctly and calmly.

The three layers of Claude, and what each gives a firm

Definition: Claude has three working environments, and the difference between them is the difference between "another chatbot" and an automated practice: Claude in the browser (conversations about text), Claude Cowork in the desktop app (work on real folders and documents on your computer), and Claude Code (an agent that builds automations and systems).

Claude in the browser: drafting and questions

The familiar layer: general legal questions, polishing drafts, translating a document, brainstorming arguments. One tip already here: choosing the right model saves money. Sonnet is excellent for most day-to-day work and stretches a subscription; save the heavy models for complex analysis.

Claude Cowork: where a law firm really starts to benefit

In the desktop app, Claude works on real folders on your computer. Open a working folder for each matter, and Claude reads everything in it (pleadings, contracts, correspondence, transcripts) and works on it: producing drafts, answering questions from the material, preparing ready Word files. No uploading files one by one into a chat, and none of the upload limits that frustrate browser users.

Claude Code: the automated practice

Claude Code is the layer that turns repetitive actions into processes that run on their own: a daily check for unanswered correspondence, a status dashboard for every matter in the firm, deadline reminders, weekly reports for partners. It is the second stage, after the firm already works smoothly with Cowork.

What does "before and after" look like in a real firm?

Take a representative ten-lawyer litigation firm. This is what the rollout we guide looks like, step by step:

Week 1, infrastructure: the right account, training off, written input rules (one page), and the "iron rule" in Claude's standing instructions: no document and no email is sent without a lawyer's approval; everything is prepared as a draft only. That one sentence prevents most possible failures.

Weeks 2 to 3, the first win: pick one painful process (for most litigation firms, pleadings) and build a Skill for it (see below). The measurable result at firms that work this way: a first draft that used to take half a day drops to under an hour, including review.

Month 2 onward, expansion: connect email (a connector), add Skills for contracts and demand letters, and explore automations in Claude Code. Every stage is measured: how many hours saved, on which matters.

9 Claude use cases that work in law firms

1. How does Claude prepare a draft pleading?

Give it two or three strong pleadings from the firm as examples, the case material (cleaned of unnecessary details) and instructions, and it returns a complete draft in the firm's structure and language. The secret is not the one-off request but the Skill: you save all the rules once, and from then on every draft comes out to the same standard. Full walkthrough below.

2. How do you review a contract and spot risks?

Enter the contract and ask for a review against the firm's checklist: liability and indemnity, jurisdiction clauses, exit mechanisms, indexation, guarantees. Claude returns a risk table referencing clause numbers, and the lawyer gets a map for focused reading instead of a blind read of 40 pages. The same approach works on leases, founders' agreements and employment contracts.

3. How do you distill a year of correspondence into a status picture?

A matter that runs for years accumulates hundreds of emails and letters. In Cowork, put the correspondence in the matter folder and ask for a timeline of key events, each party's commitments, and the open issues. Before a hearing or settlement meeting, that is the difference between half a day of preparation and half an hour.

Work against intuition here: do not ask Claude "what does the case law say" and copy the authorities. Download the relevant decisions from your research database, give them to Claude as files and ask for analysis: what the holding is, where the distinctions lie, how it applies to our facts. Claude is brilliant at analyzing text in front of it and dangerous when asked to recall from memory. That is exactly what the warning section below is about.

5. How do you manage correspondence with opposing parties and institutions?

Firms correspond endlessly with the same parties: opposing counsel, experts, banks, insurers, agencies. With an email connector (Gmail or Outlook/Microsoft 365), Claude prepares reply drafts from the matter's context and drafts escalating reminder letters ("second request", "final notice before proceedings"), and everything stays as drafts for approval, thanks to the iron rule.

6. How do you summarize hearings, mediations and meetings?

Record (subject to the applicable rules), get a transcript, and Claude produces a summary, decisions, tasks by owner, and follow-up points. A client intake meeting becomes a tidy internal memo the same day.

7. How do you prepare demand letters quickly and consistently?

A good demand letter is template plus facts plus exactly the right tone. A demand-letter Skill (with two or three examples from the firm) produces a draft in minutes, letterhead included, and the lawyer just refines the tone and signs.

8. How do you improve new-client intake?

An intake questionnaire filled in by the client becomes, in Claude's hands, a matter-opening memo: a factual summary, follow-up questions worth asking, missing documents, and an initial timeline estimate. The lawyer arrives at the first meeting far better prepared.

9. The step zero everyone skips: ingesting messy raw material

In the real world, material does not arrive as clean PDFs. Clients send crooked phone photos of documents, faded fax scans, screenshots and old correspondence in every possible format. If you feed Claude a pile like that and expect precise analysis, you will be disappointed, and rightly so: output quality will not exceed input quality. The right process has two stages. First, ask Claude to go through the folder, transcribe each document into readable text, give each file a clear name (date, document type, party) and explicitly mark passages that cannot be deciphered. Only then work on the clean versions. The rule: anything marked "illegible" goes back to the client with a request for a new copy, and is never filled in from imagination. Firms that adopt this habit discover that many "AI errors" were actually scanning errors.

Hands-on: building a pleadings Skill, step by step

A Skill is Claude's "saved brain": a set of instructions saved once and applied every time, like a senior employee who already knows the firm's method. Here is how to build one for pleadings:

  1. Gather 2 to 3 winning examples: pleadings the firm is proud of, cleaned of identifying details.
  2. Open a conversation and upload the examples with an explanation: what is good about them, the structure, which phrases are mandatory and which are forbidden.
  3. Iterate. Ask for a test draft on invented facts, comment, improve, until the result meets the standard.
  4. Tell it: "Save this as a Skill" with a clear name ("Pleading, property damage") and an instruction for when to use it. Common trap: after creating it you must click the Save button that appears at the top. Without it the Skill is not saved, and that is the number one reason for "my skill disappeared."
  5. Use it: on every new matter, "Use the property damage pleading Skill and prepare a draft from the material in this folder."

That is how you build a firm-wide Skill library: pleadings by practice area, contracts, demand letters, intake memos. Every new lawyer who joins gets a "senior employee" from day one.

The three rules that make this safe

  • The iron rule in the standing instructions: "You never send a document or email; you prepare drafts for approval only."
  • One folder per matter: in Cowork, select the working folder before starting the conversation. It is the context, so conversations about different matters do not mix.
  • Ask for a plan before execution: on big tasks, "Before you start, tell me what you intend to do and why." You see the direction before investing.

Hallucinations and fake case law: the warning you must read

This is not a theoretical warning: a global tracking database lists 2,306 cases of AI-hallucinated content submitted to courts (as of November 16, 2026), and in the first quarter of 2026 alone, US courts imposed about $145,000 in sanctions over fabricated citations. A few landmarks every lawyer should know:

  • Mata v. Avianca (New York, 2023): the case that opened the era. Lawyers filed a brief citing invented cases and were fined $5,000. (ABA overview)
  • The Morgan & Morgan matter (February 2025): eight of nine authorities cited in motions did not exist; three lawyers were sanctioned.
  • Whiting v. City of Athens (Sixth Circuit, March 2026): more than two dozen fake citations and misrepresentations in appellate briefs. Each of the two attorneys was ordered to pay $15,000 in punitive sanctions, plus the opposing side's appellate attorney fees and double costs. (Sixth Circuit Appellate Blog analysis)

The three rules that prevent this (and exactly what professional-ethics guidance already requires): work from sources you provide, not from the model's memory; verify every authority against Westlaw, Lexis or your official database before filing; and a standing instruction, "when you have no reliable source, write 'needs verification' and do not fill in details."

What does it cost, and what is the return?

The right entry path for a firm of up to 10 lawyers: start small, on a plan built for business, and expand based on actual use. Prices from Anthropic's pricing page, November 2026:

  • Pro ($20 per month, or $17 per month billed annually): for a solo lawyer evaluating the tool. When the usage limit runs out in the middle of a workday, that is the signal to upgrade.
  • Max ($100 per month) and Ultra ($200 per month): individual plans with roughly 5x and 10x Pro usage. Useful for a heavy individual user, but they are consumer plans, so check the training setting.
  • Team (2 to 150 users): the natural fit for a firm. Standard seats are $25 per user per month ($20 billed annually), Premium seats with more usage are $125 ($100 annually). It includes Claude Code and Cowork, admin controls, and no training on your data by default, which solves the training issue at the root.
  • Enterprise: for larger firms that need advanced security and compliance controls, priced per seat plus usage.
  • Savings in practice: have the team use Sonnet with medium thinking for day-to-day work and save the heavy models for genuinely complex analysis. That small habit stretches a subscription a long way.

And the right comparison: an hour of legal work is worth hundreds of dollars in most markets. A tool that saves each lawyer 3 to 4 hours a week on drafting, distilling and correspondence pays back its monthly cost in the first week, and everything after that is operating profit.

Where to start: a 30-day path

  1. Today: turn off training in the privacy settings and write a one-page input-rules document for the firm (based on the ethics rules that govern you).
  2. Week 1: install the desktop app for every lawyer, set the iron rule, and run a first exercise: distilling the correspondence of one matter.
  3. Weeks 2 to 3: build the first pleadings Skill and connect an email connector, on a single pilot matter.
  4. Week 4: measure how many hours were saved, and decide on the next expansion (more Skills? Claude Code automations? training the whole team?).

If you want to shortcut the process, in a one-on-one consulting session we build this together on your real matters, from setting the rules to a working Skill, and for firms that want a full rollout across the team there is end-to-end AI implementation for organizations. You can also start learning at your own pace with our courses.

Bottom line

Lawyers who adopt Claude properly do not replace legal judgment; they free it from grunt work. The numbers are clear (79% are already using AI), the ethical framework exists (ABA Formal Opinion 512 and its counterparts elsewhere), and the pitfalls are known and avoidable (an account that does not train, verified authorities, the iron rule). What separates a firm playing with a chatbot from a firm saving dozens of hours a month is not technology. It is method, and now it is written down here.

Looking for this kind of training for your profession? We run AI lectures and workshops by profession, for HR, marketing, executives, public sector, teachers and business consultants, built on real processes from each field, just like this guide.

Disclosure: this guide was written by the tzedek.me team led by Yaakov Tzedek, who has guided more than 100 organizations through AI implementation and teaches lawyers and law firms to work with Claude, including lectures for bar association district chapters. Nothing here is legal advice; the binding ethics rules are those published by the bar or regulator in your jurisdiction.

Frequently asked questions

Are lawyers allowed to use Claude?

Generally yes, under the same professional duties that govern any tool. In the US, ABA Formal Opinion 512 (July 29, 2024) says lawyers may use generative AI if they meet their duties of competence, confidentiality, client communication, supervision, candor to the tribunal and reasonable fees. Other bar associations and regulators have issued similar guidance. The common principles: full professional responsibility stays with the lawyer, every output must be checked, client-identifying information needs protection, and informed client consent may be required before confidential information goes into certain tools. ABA opinions are not binding, so check your own jurisdiction's rules.

Does Claude train on what the firm enters?

It depends on the account type. On consumer accounts (Free, Pro, Max), chats are used for training only if the Model Improvement setting in your privacy settings is on, so check it and turn it off. On commercial accounts (Team and Enterprise), Anthropic does not use inputs or outputs for training by default, and on the API, inputs and outputs are deleted within 30 days and not used for training. For a law firm the recommendation is clear: a business account, and training turned off on any personal account.

Can Claude write a pleading?

Claude can produce a complete first draft of a pleading from the firm's examples and the case file, and with a Skill (a saved set of instructions) it does so in the same structure and legal language every time. But it is a draft. Every professional-ethics framework puts full responsibility for each document on the lawyer, so every draft is reviewed, its authorities verified and edited before filing.

What is the difference between Claude in the browser, Claude Cowork and Claude Code for a lawyer?

Claude in the browser is the familiar chat, good for drafting and questions. Cowork in the desktop app works on real folders on your computer: it reads an entire case file, produces documents and prepares email drafts. Claude Code is the advanced layer, an agent that builds standing automations: correspondence tracking, a case dashboard, reminders. Most firms start with Cowork and move to Claude Code when they want processes that run on their own.

How do I make sure Claude does not invent case law?

Three rules: (1) never ask it for authorities from memory; give it the relevant decisions as files and ask it to work only from them; (2) verify every case number, quotation and citation against Westlaw, Lexis or your jurisdiction's official database before filing; (3) set a standing instruction that when it has no source, it writes 'needs verification' instead of filling in details. US courts have already imposed sanctions ranging from thousands to tens of thousands of dollars over fabricated citations; this is not a theoretical warning.

How much does it cost to bring Claude into a 10-lawyer firm?

For a firm, the Team plan is the natural fit: Standard seats are $25 per user per month billed monthly ($20 billed annually), and Premium seats with more usage are $125 ($100 annually), for 2 to 150 users, with training off by default and admin controls. A solo lawyer can start on Pro ($20 per month). The right comparison is not the subscription cost but an hour of legal work: if a tool saves each lawyer even 3 to 4 hours a week on drafting, correspondence and summaries, the return is many times the cost.

Do I need a technical background to use Claude at the firm?

No. Working in the browser and in Cowork is in plain natural language. What you do need is a method: written rules for what may and may not be entered, Skills for recurring processes, and correct connections to the tools you already use. That takes a lesson or two to learn, and this guide gives you the foundation.

What should never be entered into Claude?

Client-identifying information or details that could link to a client, unless your confidentiality analysis and, where required, the client's informed consent allow it. In practice: replace names with labels ('the client', 'the company'), remove ID numbers and addresses where they are not needed, work on a business account that does not train on your data, and for especially sensitive documents consider local processing only.

I created a skill in the browser and cannot find it in the desktop app. Why?

Because the browser and the desktop app have been separate environments: skills, connectors and settings created in one do not always sync to the other. This is the trap new users fall into most often, together with forgetting to click Save after creating a skill. The recommendation for a firm: pick one environment as the main workspace, almost always the desktop app because it works on folders, and build all skills and connectors there.