---
title: "Claude for Legal: A Plaintiff Firm Guide to Plugins, Cost and Risk"
seoTitle: "Claude for Legal: 12 Plugins, Pricing and Setup (2026)"
url: https://www.masstortmarketingagency.com/blogs/claude-for-lawyers
canonical: https://www.masstortmarketingagency.com/blogs/claude-for-lawyers
published: 2026-10-22
modified: 2026-10-22
author:
  name: Tarun
  role: Founder, Mass Tort Marketing Agency
publisher:
  name: Mass Tort Marketing Agency
  url: https://www.masstortmarketingagency.com
description: |
  What Claude for Legal costs and does: 12 free plugins, 20+ connectors, paid plan prices, setup steps and ABA 512 risks for plaintiff PI and mass tort firms.
keywords:
  - claude for legal
  - claude for lawyers
  - claude legal plugin
  - claude vs chatgpt for lawyers
  - claude for legal vs harvey
  - claude for legal pricing
  - how to use claude for legal
  - claude for legal github
  - which claude is best for legal documents
  - is claude ai good for legal
banner: /images/blog-content/claude-for-lawyers/hero.jpg
ogImage: /images/og/claude-for-lawyers-og.jpg
license: |
  Cite freely with attribution to Mass Tort Marketing Agency. Verbatim quoting
  permitted with citation back to the canonical URL.
---

> **Quick answer.** Claude for Legal is Anthropic's open-source suite of 12 practice-area plugins and 20+ MCP connectors for legal work, announced in Anthropic's [Claude for the legal industry launch post](https://claude.com/blog/claude-for-the-legal-industry) on May 12, 2026. It adds 90+ named agents and works inside Word, Excel, PowerPoint and Outlook. The plugins are free under the Apache-2.0 license, but using them as documented requires a paid Claude plan. Every output is a draft for attorney review, never a finished legal conclusion.

**Key takeaways**

- Claude for Legal is a bundle of plugins, agents and connectors that runs inside Claude. Anthropic launched it on May 12, 2026 with 12 plugins, from Commercial Legal to Legal Builder Hub.
- Cost, as of October 2026: the plugins are free, but you need a paid plan. Pro is $17 a month billed annually or $20 monthly; Team standard seats are $20 annual or $25 monthly; Enterprise is $20 per seat plus usage at API rates, with a 20-seat minimum.
- None of the 12 plugins is built for personal injury or mass tort. Litigation Legal and Legal Clinic adapt best to plaintiff work.
- Client files belong on commercial terms (Team or Enterprise), not consumer plans. Only Enterprise can enable the HIPAA-ready configuration.
- Every output is a draft. Connect a research tool first, verify every citation, and read ABA Formal Opinion 512 before you roll anything out.
- The recommended default model as of October 2026 is Claude Opus 5.5.

This guide is written for personal injury and mass tort firms first and for any lawyer second. It explains what Claude for Legal is, how it arrived over 2026, what each of the 12 plugins and the litigation agents do, and which connectors exist. Later sections cover pricing and plan choice, installation, plaintiff-side workflows with a prompt library, accuracy and ethics rules, privilege and HIPAA, comparisons with other tools, and a 90-day rollout plan. Multi-tool roundups live in [our broader guide to legal AI tools](/blogs/legal-ai-tools); this page stays on Claude.

A disclosure before we start. Mass Tort Marketing Agency is a marketing and lead-generation agency for plaintiff firms, best known for [personal injury lead generation](/personal-injury-lead-generation). We do not sell Claude or any AI software, we are not a law firm, and nothing here is legal advice. Facts on this page were last verified on 22 October 2026.

One naming note. In September 2026 Anthropic began merging Cowork and chat into "one Claude", and that rollout is still in progress, so you will see both names below.

## What Claude for Legal is and what it is not

Claude for Legal is Anthropic's name for a bundle of legal plugins, connectors and agents that run inside Claude. It is not a separate model, a separate app or a legal research database.

That distinction drives most of the practical questions a firm will ask: what it costs (the price of Claude itself), where client data goes (the terms of your Claude plan), and how accurate it is (as accurate as the model, the instructions and the sources you connect).

### The one-paragraph definition

Claude for Legal is Anthropic's open-source collection of 12 practice-area plugins, 90+ named agents and 20+ MCP connectors that turn Claude into a structured legal assistant. Each plugin interviews you about your practice, writes a practice profile, and then runs skills such as demand drafting, chronologies and contract review as drafts for attorney review.

Anthropic announced the suite in its [launch post of May 12, 2026](https://claude.com/blog/claude-for-the-legal-industry), filed under the Claude Cowork product. The post describes "20+ new MCP connectors that link Claude to the software the legal industry already relies on, and 12 new plugins tailored to specific legal work and practice areas." It also names partnerships with the Free Law Project and the Justice Technology Association, "and others". The Free Law Project runs CourtListener, which matters later because it is one of the free research connectors.

The same post opens with a usage claim worth reading as Anthropic's own: it says legal professionals "have become the most engaged Claude Cowork users of any knowledge-work function." Anthropic has not published the underlying numbers, so treat it as a vendor statement.

Three days later, on May 15, 2026, Anthropic published [Deploying Claude across the legal industry](https://claude.com/resources/articles/deploying-claude-across-the-legal-industry). That guide sets out a product matrix (chat, Cowork, the Microsoft 365 add-ins and the Claude Platform), the plugin set and "a three-phase adoption plan." It is the closest thing Anthropic offers to an official rollout document, and it is worth reading alongside our own 90-day plan later in this guide.

### Claude for Legal versus using Claude for legal work

People who search "claude for legal" mean one of two things, and the difference changes the answer.

The first meaning is using Claude, the general assistant, for legal tasks. Lawyers have pasted contracts, deposition transcripts and records into Claude to summarize or draft since well before 2026. In that mode you write every prompt yourself and supply every rule each time: your jurisdiction, your house style, your escalation thresholds, what not to say in a demand. Nothing persists unless you build it.

The second meaning is the product launched in May 2026. With a plugin installed, a short setup interview writes a practice profile, and every skill in that plugin reads it. A demand-letter skill already knows your firm's tone, your review chain and that a human must approve before anything is sent. The structure also adds gates, connectors and named commands.

A short way to put it: without plugins, you are the system. With plugins, the system remembers your rules and enforces checkpoints, and you review the output.

Much of what still ranks for "claude for lawyers" covers only the first meaning. The [rankings.io guide](https://rankings.io/blog/claude-for-lawyers/), for example, is a widely ranked 2025 guide, re-dated in 2026, that predates the product. That is why it still refers to older Claude models. Both meanings are useful, and this guide covers both, but anything about plugins, agents and connectors applies only to the second.

### Who built it and how it is licensed

Claude for Legal is built and maintained by Anthropic, which describes itself as ["a public benefit corporation dedicated to securing [AI's] benefits and mitigating its risks."](https://www.anthropic.com/company) The code lives in a public GitHub repository, [anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal), described as "A suite of plugins for legal workflows." It was created on April 21, 2026, carries the Apache-2.0 license, and had nearly 10,000 stars as of October 2026.

Apache-2.0 is a permissive open-source license. A firm, a vendor or a legal aid clinic can read the plugins, copy them, change them and use the changed versions, including commercially, as long as it keeps the license notices. For a law firm, the practical value is transparency: you can read exactly what a skill tells Claude to do before you trust it with a matter.

The README is unusually direct about limits. Its guardrail notice reads: "**Every output from these plugins is a draft for attorney review — not legal advice, not a legal conclusion, not a substitute for a lawyer.**" It lists the guardrails as "source attribution on every citation, conservative defaults on privilege and subjective legal calls, jurisdiction assumptions surfaced, and explicit gates before anything is filed, sent, or relied on." It also states: "These plugins do not represent Anthropic's legal positions."

On the people side, Anthropic's [legal solutions page](https://claude.com/solutions/legal) lists Mark Pike as Associate General Counsel at Anthropic. Legal news site Artificial Lawyer describes him as the lawyer leading the Claude for Legal rollout.

![Diagram of Claude for Legal layers: the model, plugins with skills and agents, connectors and the apps lawyers use](/images/blog-content/claude-for-lawyers/01-claude-for-legal-architecture.jpg "How the pieces fit: one model, plugins that read your practice profile, connectors to your systems, and the surfaces where lawyers work.")

### Constitutional AI and Claude's limits in plain English

Constitutional AI is Anthropic's method of training Claude against a written set of principles, a "constitution", that describes how the model should behave. Anthropic published [Claude's new constitution](https://www.anthropic.com/news/claude-new-constitution) on January 22, 2026, and the full text is public at [anthropic.com/constitution](https://www.anthropic.com/constitution).

For a lawyer, the useful way to read this is narrow. The constitution shapes behavior: how Claude handles harmful requests, how it weighs helpfulness against caution, and how candid it should be about uncertainty. It is not a fact-checking system. Nothing in a constitution guarantees that a case citation exists, that a quote is accurate, or that a statute has not been amended.

Claude can and does produce errors, including confident ones. The plugins reduce that risk with source tagging and review gates, but they do not remove it. Our section on accuracy limits and verification, later in this guide, covers the error types, the sanctions cases and the published error-rate research in detail.

There is also a simpler limit: time. As of October 2026, Anthropic's [training-data cutoff page](https://support.claude.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data) puts the current models' reliable knowledge at June 2026, and it warns that models "may not be aware of events or information that occurred after their respective cutoff dates." A ruling from last month, a new MDL, or an amended rule will not be in the model's memory. That is the strongest practical argument for connecting a research tool before anyone relies on a legal citation.

## Claude for Legal timeline: January to October 2026

Claude for Legal arrived in stages over 2026, starting with a single legal plugin in January and reaching a 12-plugin suite in May. The model underneath changed four more times after launch.

The table lists each milestone with its date and the fact-sheet reference we verified it against. Dates are from Anthropic's own announcement and marketplace pages unless noted.

| Date | What shipped | Source F# |
|---|---|---|
| Nov 25, 2024 | Anthropic open-sources the Model Context Protocol | F3 |
| Jan 22, 2026 | New Claude constitution published | F2 |
| Jan 30, 2026 | Cowork plugin support; first "Legal" plugin among 11 | F7 |
| Feb 2026 | Harvey connector announced (Feb 24) | F35, F125 |
| Apr 9, 2026 | Cowork generally available on paid plans | F39 |
| Apr 16, 2026 | Claude Opus 4.7 | F50 |
| Apr 21, 2026 | claude-for-legal GitHub repo created | F18 |
| May 7, 2026 | Claude for Word, Excel, PowerPoint GA; Outlook beta | F42 |
| May 12, 2026 | Claude for Legal: 12 plugins, 20+ connectors | F5 |
| May 15, 2026 | Legal deployment guide | F17 |
| May 28, 2026 | Claude Opus 4.8 | F50 |
| Jul 24, 2026 | Claude Opus 5 | F50 |
| Aug 2026 | 8am MyCase connector added | F33 |
| Aug 26, 2026 | Help Center article on privilege and configuration | F65 |
| Sep 2026 | Filevine connector added | F32 |
| Sep 16, 2026 | Cowork and chat become "one Claude" (rolling out) | F40 |
| Sep 22, 2026 | Claude Opus 5.5 | F50 |
| Oct 7, 2026 | Claude Haiku 5.5 | F50 |

The pace is the headline. In ten months the product went from one general "Legal" plugin to a 12-plugin suite, the Microsoft 365 add-ins reached general availability, two case management connectors relevant to plaintiff firms appeared, and the flagship model was replaced four times. Anything you read from early 2026 describes a different product.

![Timeline of Claude for Legal milestones in 2026 from the first legal plugin to Opus 5.5](/images/blog-content/claude-for-lawyers/02-launch-timeline.jpg "Claude's legal tooling shipped in stages through 2026. Anything written before May describes a different product.")

That includes Anthropic's own launch material. On May 12 the launch post called Claude Opus 4.7 "our most capable publicly available model for legal reasoning and long-document work." That was true in May. As of October 2026, Anthropic's [models overview](https://platform.claude.com/docs/en/about-claude/models/overview) recommends starting with Claude Opus 5.5 for most workloads, so the launch-post line is dated.

A second correction concerns the start date. It is easy to assume the first legal plugin arrived in February, when the Harvey connector was announced. Anthropic's [Cowork plugins announcement](https://claude.com/resources/articles/cowork-plugins) is dated January 30, 2026, when Anthropic added plugin support to Cowork "as a research preview for all paid Claude users" and open-sourced 11 plugins, one of them "Legal — Review documents, flag risks, and track compliance." Claude for Legal, the 12-plugin suite, followed on May 12.

The practical lesson for a firm: put a quarterly review on the calendar. Each quarter, someone should check the model lineup, the plan page, the connector directory and the repo's changelog, then update your internal AI policy and any saved prompts that name a specific model or feature.

## The 12 Claude for Legal plugins, explained

Anthropic's 12 practice-area plugins each package skills, agents and a setup interview for one kind of legal work. Most target transactional, regulatory and in-house practice; Litigation Legal and Legal Clinic are the ones closest to plaintiff work.

### What a plugin contains: skills, agents and a practice profile

A plugin is a folder of plain-text files. Inside it are three kinds of building blocks.

Skills are the core. Anthropic's [Skills Help Center article](https://support.claude.com/en/articles/12512176-what-are-skills) defines them as "folders of instructions, scripts, and resources that Claude loads dynamically." In the repo each skill lives at `<plugin>/skills/<skill>/SKILL.md`. A skill tells Claude how to do one job, such as drafting a demand or building a chronology, including what to ask, what to check and what to output.

Agents are named workflows with slash commands, such as `/litigation-legal:chronology`. The repo's README lists 90+ of them across the plugins.

The practice profile is a plain-English `CLAUDE.md` file written by a setup interview and stored at `~/.claude/plugins/config/claude-for-legal/<plugin>/CLAUDE.md`. Every skill in the plugin reads it, and the README notes, "It survives plugin updates." Connectors are configured per plugin in a `.mcp.json` file.

The README sums up the format: "Everything is markdown and JSON. No build step." That matters to a non-technical firm. A managing partner or paralegal can open a skill, read what it instructs Claude to do, and edit it without a developer. Each plugin starts with that setup interview, which the launch post says "learns your practice: your playbook, your escalation chain, your risk calibration, your house style." The install and interview steps are covered in the setup section later in this guide.

The table lists the 12 plugins named in the launch post, what each is built for, and my rating of its relevance to a plaintiff PI or mass tort firm.

| Plugin | Built for | Plaintiff-firm relevance |
|---|---|---|
| Commercial Legal | Commercial contracts and negotiation | Low: vendor contracts only |
| Corporate Legal | Corporate and transactional work | Low |
| Employment Legal | Employment law matters | Low to medium: firm HR, some employment claims |
| Privacy Legal | Privacy and data protection | Medium: intake data and website privacy |
| Product Legal | Product counseling for companies | Low |
| Regulatory Legal | Regulatory tracking and compliance | Low to medium |
| AI Governance Legal | AI policy and governance | Medium: firm AI use policy |
| IP Legal | Intellectual property | Low |
| Litigation Legal | Matters, demands, chronologies, depositions | High |
| Law Student | Legal education | Low |
| Legal Clinic | Client intake, deadlines, supervised review | High: high-volume client work |
| Legal Builder Hub | Finding and vetting community skills | Medium: building custom PI skills |

These ratings are mine, not Anthropic's. The pattern is clear, though: ten of the twelve plugins serve corporate, in-house, regulatory or educational users. A plaintiff firm will spend most of its time in two plugins and use Legal Builder Hub to bring in or vet custom skills.

The older Anthropic plugin is still worth knowing. The original [Legal plugin in the Claude marketplace](https://claude.com/marketplace/plugins/legal), labelled "Anthropic verified" and "Made by Anthropic", offers `/review-contract` (clause-by-clause review against your playbook "with GREEN/YELLOW/RED flags and specific redline suggestions"), `/triage-nda`, `/vendor-check`, `/brief` and `/respond`. Its page states: "All outputs should be reviewed by licensed attorneys." For a plaintiff firm it is mostly useful for reviewing vendor, co-counsel and referral agreements.

### Litigation Legal: the plugin plaintiff firms should start with

Litigation Legal is the closest Anthropic gets to plaintiff work. Per the launch post, it "manages matter intake and portfolio tracking, legal holds, demand letters, subpoena triage, chronologies, deposition preparation, privilege logs, and brief drafting."

Several of those map directly onto a PI or mass tort file. Demand letters are the center of most pre-suit PI practice. Chronologies are how a case manager turns hundreds of pages of medical records into a treatment story. Deposition preparation and brief drafting matter once a case is in suit. Matter intake and portfolio tracking fit the case-opening and case-status work that consumes so much staff time at volume. Legal holds and subpoena triage matter for preserving evidence and responding to third-party requests.

It is still a general litigation plugin. The README is explicit that none of the litigation agents is built for personal injury or mass tort, and the gaps show. There is no medical-billing logic that reconciles charges against payments and write-offs. There is no lien workflow for Medicare, Medicaid, ERISA plans or hospital liens. There is no tort qualification grid that checks a claimant against a specific litigation's product, exposure window and injury list. Those pieces have to come from custom skills or from people.

My recommendation is to start here anyway, because the structure (practice profile, gates, connectors) is sound and editable. Our section on Claude workflows for personal injury firms, later in this guide, shows how to adapt each skill to plaintiff work.

### Legal Clinic and the client-letter pattern

Legal Clinic was built for legal aid and law school clinics. The launch post says it "manages client intake, deadline tracking, case memos, and the supervisor review queue."

Two of its agents deserve a plaintiff firm's attention. The Client Letter Drafter, run with `/legal-clinic:client-letter`, handles "Routine client correspondence — appointment confirms, doc requests, updates." Clinic Intake provides "Structured client intake with cross-area issue spotting and conflict flags."

The reason to look closely is volume. A clinic and a mass tort firm share a problem: many clients, limited staff, and a duty to keep each client informed. The supervisor review queue is also the right shape for a firm where paralegals draft and attorneys approve. I think the client-letter pattern, a drafted update that waits for human approval, is the safest first use of Claude in client communication. The mass tort workflows section later in this guide applies it to status updates and document requests.

### Legal Builder Hub, community skills and the CoCounsel plugin

Legal Builder Hub is the plugin for finding, vetting and installing skills that Anthropic did not write. The README calls it the "trust layer" for community skills. It runs a security review (a hidden-content scan and injection detection), uses a restrictive-by-default allowlist, applies a license gate and a freshness gate, re-scans at every update and keeps an install log. The README is also frank: "Nobody certifies community skills."

The launch post names Legal Quants and Lawvable as sources of community-maintained skills. The repo also includes one partner-built plugin, `cocounsel-legal` from Thomson Reuters, which "Requires a CoCounsel Legal subscription with the MCP connector enabled." With Anthropic's 12, that makes 13 plugin entries in the repo.

My advice: treat any community skill like third-party software. Read the `SKILL.md` before installing it, run it on a closed file first, and record who approved it.

## The 90+ named agents that matter to litigators

The repo's agents table lists 93 named agents across the plugins (our count). Artificial Lawyer's coverage rounds that to "over 90". About 15 sit in Litigation Legal, and those are the ones a plaintiff firm will use.

Artificial Lawyer's headline on 1 June 2026 put it as ["Claude For Legal Has Over 90 AI Agents"](https://www.artificiallawyer.com/2026/06/01/claude-for-legal-has-over-90-ai-agents/). The count matters less than the mix. Most agents serve contract, privacy, regulatory and corporate teams.

### Litigation agents mapped to plaintiff work

The table lists the Litigation Legal agents from the README, the command where the README gives one, what each does, and a possible plaintiff use. The last column is my mapping, not Anthropic's.

| Agent | Command | What it does | Possible plaintiff use (author's mapping) |
|---|---|---|---|
| Demand Letter Drafter | `/litigation-legal:demand-draft` | Drafts a demand with FRE 408 awareness and a send gate; .docx output | Pre-suit PI demand draft |
| Demand Intake | `/litigation-legal:demand-intake` | Named in README; no description | Collect damages and treatment facts |
| Demand Received Triage | None listed | Named in README; no description | Sort defense or carrier letters |
| Chronology Builder | `/litigation-legal:chronology` | Builds or updates a chronology from declared sources and uploads | Treatment timeline from records |
| Deposition Prep | `/litigation-legal:deposition-prep` | Outline tied to case theory with docs and impeachment | Prep for defense medical experts |
| Matter Intake | None listed | Named in README; no description | Case opening after retainer |
| Matter Briefing | None listed | Named in README; no description | Handoff memo to trial counsel |
| Portfolio Status | None listed | Named in README; no description | Status across a mass tort inventory |
| Legal Hold | None listed | Named in README; no description | Preservation letters for evidence |
| Privilege Log Reviewer | None listed | Named in README; no description | Check defense privilege logs |
| Subpoena Triage | None listed | Named in README; no description | Third-party records subpoenas |
| Docket Watcher | None listed (scheduled agent) | Named in README; no description | MDL docket monitoring |
| Brief Section Drafter | None listed | Named in README; no description | Motion sections for review |
| Claim Chart Builder | None listed | Charts a patent or civil cause of action | Elements chart for each claim |

The README also lists an Outside Counsel Status agent, which fits in-house legal teams better than plaintiff firms, so I left it out of the table.

Read the table as a starting map. Where the README gives only an agent's name, check its skill file for the exact behavior before relying on it. The four agents with commands (demand drafting, demand intake, chronology and deposition prep) are the ones most plaintiff firms will test first, because they line up with daily work. Docket Watcher is the most interesting for mass tort practice, since it runs on a schedule rather than waiting for a prompt. Matter Intake and Portfolio Status fit case management, but both work best when a case management connector supplies the data.

### What none of the agents does

None of the agents is built for personal injury or mass tort, and the README says so. The gaps are worth listing because they are exactly the jobs that drive plaintiff economics.

No agent values a case. None calculates medical specials against a fee schedule, reconciles billed versus paid amounts, or projects future care. None resolves liens or tracks Medicare conditional payments. None scores a claimant against tort qualification criteria. None fills an MDL plaintiff fact sheet or a census form. None checks proof of use or proof of injury against a litigation's requirements.

Each of those needs a custom skill, a separate specialized tool, or a person. A firm can write those skills on top of the Litigation Legal structure, and our section on building a custom mass tort skill, later in this guide, shows how.

The send gate on the Demand Letter Drafter is a good example of the design philosophy. It exists because settlement communications carry evidence-rule consequences under Federal Rule of Evidence 408, which the workflows section explains. Mark Pike described the gates to Artificial Lawyer this way: the plugins have "explicit gates before anything is filed, sent, or relied on. The lawyer reviews and verifies; the tooling is designed to make that review easier, never to skip it." That is a fair summary of how the agents should be used: they draft and check, and a lawyer decides.

## MCP connectors: what they do and which exist

Connectors use the Model Context Protocol to let Claude read from, and sometimes write to, the software a firm already uses, under each user's existing permissions. As of October 2026 the launch set spans research, e-discovery, document management and case management.

### What MCP is and how permissions work

Anthropic open-sourced the [Model Context Protocol](https://www.anthropic.com/news/model-context-protocol) on November 25, 2024, calling it "a new standard for connecting AI assistants to the systems where data lives, including content repositories, business tools, and development environments." A software vendor exposes its data through an MCP server, and Claude connects to it as a client. If terms like MCP, MDL or PFS are new, our [mass tort marketing glossary](/glossary) defines the common ones.

Permissions follow the user. Anthropic's [connectors Help Center article](https://support.claude.com/en/articles/11176164-use-connectors-to-extend-claude-s-capabilities) says: "Claude inherits each person's permissions from the connected service. If someone can't access a specific file, channel, or record in the source system, the connector can't reach it from Claude either." A paralegal walled off from a conflicted matter in your case management system stays walled off in Claude.

Read access and write access carry different risk. Anthropic's [Cowork safety guidance](https://support.claude.com/en/articles/13364135-use-claude-cowork-safely) says write tools "inherently carry more risk" and that "human oversight is recommended in high stakes scenarios." It also defines prompt injection as "malicious instructions … embedded in external content that Claude reads as part of a legitimate task." A medical record, an email or a web page can carry such text. Review each connector's scopes, and start read-only.

Some connectors are free; others are marked "customer subscription" in the README, meaning you need your own account and API key with that vendor.

### Connectors by category

The launch post grouped the connectors into nine categories. The table lists them as named, as of October 2026.

| Category | Connectors named at launch |
|---|---|
| Contract lifecycle | Definely, Docusign, Ironclad |
| Deal rooms | Box, Datasite |
| Document management | iManage, NetDocuments |
| Expert networks and skills | Lawve AI, The L Suite (Lloyd, TopCounsel) |
| E-discovery | Consilio, Everlaw, Relativity |
| "Fiduciary-grade workflows" | Thomson Reuters CoCounsel Legal |
| Research | Legal Data Hunter, Midpage, Trellis |
| Legal AI assistants | Harvey, Solve Intelligence |
| Public service | BoardWise, Courtroom5, Descrybe, Free Law Project (CourtListener) |

The launch list leans toward large-firm and in-house infrastructure: deal rooms, contract lifecycle and enterprise document management. Research and e-discovery are the categories most plaintiff litigators will touch. The README's own connector table goes further and also lists general workplace tools such as Slack, Google Drive, Aurora, Linear, Atlassian (Jira) and Asana, alongside the legal ones. The README names CourtListener, Trellis, Descrybe and Solve Intelligence as research tools and tells users to "Connect a research tool first", because citations from model knowledge alone are flagged `[verify]`.

### Connectors that matter to plaintiff firms

Five connectors stand out for plaintiff work as of October 2026. Check the directory before you plan around any of them, because the list changes monthly.

- **Filevine** (added September 2026): the [Filevine connector](https://claude.com/marketplace/connectors/filevine-api), made by Filevine, retrieves "projects, contacts, documents, notes, tasks, and other supported Filevine resources" while "respecting the permissions and access available to the connected user."
- **8am MyCase** (added August 2026): the [8am MyCase connector](https://claude.com/marketplace/connectors/8am-mycase) lets users "look up a matter, review time entries and billing, check tasks and deadlines, review leads, or read client portal messages."
- **CourtListener** (added May 2026): the [CourtListener connector](https://claude.com/marketplace/connectors/courtlistener) from the Free Law Project covers "millions of U.S. court opinions, PACER dockets, judge profiles, oral arguments, and citation data" and can "extract and verify legal citations against a canonical database."
- **CoCounsel Legal** (added May 2026): the [CoCounsel Legal connector](https://claude.com/marketplace/connectors/cocounsel-legal) needs a CoCounsel Legal license and is "offered as a pilot." Its terms bar routing Thomson Reuters content to another "Professional AI Tool" and warn against relying on "always allow" or "auto-approve" settings.
- **Harvey** (added February 2026): the [Harvey connector](https://claude.com/marketplace/connectors/harvey) lets a Harvey customer query Harvey and its Vault projects from inside Claude.

No official Clio, Smokeball or Litify connector appeared in Claude's directory when we checked on 22 October 2026, so recheck before you plan around one. Litify says on its [homepage](https://www.litify.com/) that it is built on Salesforce, AWS and Anthropic infrastructure, which is a different arrangement from a Claude connector. If you are still choosing a system of record, our guide to [case management software for law firms](/blogs/legal-case-management-software) compares the options.

Almost every path above, from Cowork plugins to the Microsoft 365 add-ins to these connectors, runs on a paid Claude plan, which is where cost comes in.

## Claude for Legal pricing as of October 2026

Claude for Legal's plugins are free and open source, but you need a paid Claude plan to run them: Pro from $17 a month billed annually, Team from $20 per seat, or Enterprise at $20 per seat plus usage, as of October 2026. Those figures come from [Anthropic's Claude pricing page](https://claude.com/pricing), and they change often enough that you should confirm them with Anthropic before you buy.

![Chart of Claude plans for law firms showing which tiers include shared projects, audit logs, custom retention and HIPAA](/images/blog-content/claude-for-lawyers/03-plan-matrix.jpg "Plan features that matter for client work, as of October 2026. Confirm current prices and features with Anthropic before you buy.")

The price question has two layers. The software Anthropic published for lawyers costs nothing to download. The plan that runs it, and the vendor subscriptions some connectors need, do cost money. The sections below separate those layers so a managing partner can budget for the real number.

### Is Claude for Legal free?

Partly. The plugins and connector configurations sit in a public GitHub repository under the Apache-2.0 license, so anyone can read, copy or adapt them at no charge. Running them the way Anthropic documents is a different matter. Cowork, where most lawyers install plugins, is [available on paid plans (Pro, Max, Team, Enterprise)](https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork) only. Claude Code and Claude for Microsoft 365 are not included on the Free plan either, according to the pricing page as of October 2026.

Some connectors add a second bill. The CoCounsel Legal, Harvey, Trellis and Everlaw connectors need the firm's own account with that vendor, and CoCounsel Legal specifically requires a CoCounsel Legal license and a Thomson Reuters account. Claude's side of the connection does not replace those subscriptions.

The Free plan is still useful for a first look. As of October 2026 it includes connectors, web search, file creation and up to five Projects, so a lawyer can test general Claude on public material at no cost. What Free cannot do is run the documented plugin install. Legal aid clinics, public defenders and nonprofit legal services groups have a separate route: [Claude for Nonprofits](https://www.anthropic.com/news/claude-for-nonprofits), launched December 2, 2025, offers discounted access of up to 75% on Team and Enterprise plans for eligible organizations.

### Individual plans: Free, Pro and Max

Individual plans are priced per person and run under Anthropic's consumer terms. The table below lists what each tier costs and what it includes for legal work, alongside the two business tiers for comparison, all as of October 2026.

| Plan | Price as of Oct 2026 | Key inclusions for legal work | Model training default |
|---|---|---|---|
| Free | $0 | Chat, connectors, web search, up to 5 Projects; no Claude Code, Microsoft 365 or Research | Opt-out |
| Pro | $17/mo annual ($200 up front) or $20 monthly | Claude Code, Projects, Microsoft 365, Claude in Chrome, more models | Opt-out |
| Max 5x / Max 20x | $100 / $200 per month, monthly only | 5x or 20x Pro usage, $100 or $200 in API credits | Opt-out |
| Team | Standard $20 annual / $25 monthly per seat; premium $100 / $125 | SSO, central billing, connector admin, shared Projects | None by default |
| Enterprise | $20/seat/month annual, 20-seat minimum, plus usage at API rates | Audit logs, custom retention, SCIM, HIPAA-ready offering | None by default |

The training column is the one most lawyers skip and should not. On Free, Pro and Max, training is an opt-out choice the user makes; on Team and Enterprise it is off by default. That single difference drives most of the plan advice in the next section.

Anthropic does not publish usage as a message count. The [Pro plan Help Center article](https://support.claude.com/en/articles/8325606-what-is-the-pro-plan) says the number of messages "will vary based on message length, including the length of files you attach." Pro has a session limit that resets every five hours, plus a weekly limit. Long medical records eat into that faster than short emails. The [Max plan](https://support.claude.com/en/articles/11049741-what-is-the-max-plan) raises the ceiling to 5x or 20x Pro usage. Anthropic also states it does not offer standard discounted pricing on Pro or Max, so the nonprofit program is the only published discount route.

### Team and Enterprise for law firms

Team is built for groups "of 2 to 150," per the pricing page as of October 2026. It includes SSO, central billing, admin controls for connectors, enterprise search and "No model training on your content by default." The [Team plan Help Center article](https://support.claude.com/en/articles/9266767-what-is-the-team-plan) adds domain capture, just-in-time provisioning, role-based permissioning, spend controls and "Everything in Pro," including Claude Code and Cowork. Team bills monthly or annually. ACH payment is not available on Team, which matters for firms whose accounting runs on bank transfers.

Enterprise layers on what a compliance officer asks for: role-based access and SCIM, audit logs and the Compliance API, custom data retention controls, network-level access control with IP allowlisting, and a HIPAA-ready offering. It comes in a self-serve version and a sales-assisted version. Both bill annually, and invoicing or net terms are available only on the sales-assisted route.

Anthropic does not publish an all-in Enterprise cost per lawyer, so treat any quoted figure with caution. The $20 seat price is fixed, but usage is billed at API rates and scales with the model and the task, so a firm summarizing thousands of pages a week will spend very differently from one drafting short letters.

Three rows on the pricing page separate the business tiers for legal work. Audit logs, custom retention and the HIPAA-ready offering are Enterprise-only; Team has none of them, and neither do the individual plans. Project sharing is available on Team and Enterprise but not on Free, Pro or Max. The full data-handling comparison, including retention periods, appears in the privilege and HIPAA section later in this guide.

### API pricing for firms that build

Firms that build their own tools on the Claude Platform pay per token instead of per seat. As of October 2026, the pricing page lists these rates per million tokens:

| Model | Input | Output |
|---|---|---|
| Fable 5.1 | $10 | $50 |
| Opus 5.5 | $4 | $20 |
| Sonnet 5.5 | $2 | $10 |
| Haiku 5.5 (prompts up to 100K tokens) | $0.10 | $0.50 |

Batch processing saves 50%, according to the same page. Anthropic's [Opus 5.5 announcement](https://www.anthropic.com/claude-opus-5-5) claims the model "costs 40% less to run than Opus 5" on typical workloads; treat that as a vendor claim, since it depends on the workload.

Who needs these numbers? Mostly firms with a developer who plan to build an intake summarizer, a chronology pipeline or a scheduled agent through Managed Agents, which Step 6 of the setup guide below covers. A purely illustrative calculation: if a record set came to 300,000 tokens and you sent it once to Opus 5.5, the input would cost 0.3 × $4 = $1.20 before any output charges. That is arithmetic on the published rate, not a quote, and real jobs involve repeated passes, output tokens and review cycles.

## Which Claude plan and model a plaintiff firm needs

For client files, a plaintiff firm should be on Team at minimum and on Enterprise if it handles medical records under a HIPAA obligation or needs audit logs and custom retention. Use Opus 5.5 as the default model as of October 2026.

That is my recommendation, not Anthropic's or any bar's. The reasoning rests on three published facts: consumer plans run under consumer terms, only Enterprise offers custom retention and audit logs, and only Enterprise can enable the HIPAA-ready configuration. The rest of this section turns those facts into a decision you can make in an afternoon.

### The plaintiff firm plan picker

The table below matches common plaintiff firm profiles to the plan I recommend, why, and the trap to watch for. Prices and features are as of October 2026.

| Firm profile | Plan I recommend | Why | Watch out for |
|---|---|---|---|
| Solo PI attorney testing on non-client material | Pro | Cheapest path to plugins and Microsoft 365 | Consumer terms; turn training off; no client data |
| Solo or 2-person firm with client files | Team | Commercial terms, no training by default | No audit logs, no HIPAA |
| 5 to 20-person PI firm | Team, premium seats for heavy users | Shared Projects, SSO, connector admin | Retention is not customizable |
| Firm receiving records under a BAA, or wanting audit logs | Enterprise | HIPAA-ready, audit logs, custom retention | 20-seat minimum; BAA excludes connectors and Cowork in the cloud |
| Mass tort firm screening thousands of claimants | Enterprise or API and Managed Agents | Volume, retention control, automation | Usage cost scales; build review gates |
| Legal aid or nonprofit | Team or Enterprise via Claude for Nonprofits | Up to 75% discount | Eligibility rules apply |

Use the picker from the top down and stop at the first row that describes you. The first question is whether client information will ever enter Claude. If the answer is yes, even occasionally, skip the Pro row. A solo attorney who only wants to try plugins on public statutes, sample forms or a firm blog draft can start on Pro, switch training off, and keep client files out entirely.

The second question is whether you need proof of who did what, or a retention period other than the default. If a carrier, a co-counsel agreement or your own policy requires audit logs, only Enterprise has them. The mass tort row deserves one more note: automated screening at volume still needs people behind it. A [live intake team](/services/intake-services) handles the calls, callbacks and judgment that screening output hands off to.

Check the Team minimum on the pricing page before you sign up. The page describes Team as "for teams of 2 to 150," which implies at least two seats; confirm the current minimum and seat types on the page itself, because the footer says prices and plans can change at Anthropic's discretion. Enterprise's 20-seat minimum is stated plainly, so a 6-lawyer firm that needs HIPAA will pay for 20 seats or need a sales conversation.

Plan for the upgrade path before you need it. Watch for three triggers that move a firm from Team to Enterprise: a client, carrier or co-counsel asks for an audit trail of who saw a file; the firm starts receiving records under a business associate agreement; or the default retention no longer fits the firm's records policy. Remember too that the HIPAA decision is click-to-accept by the Primary Owner and, per Anthropic, "a one-way decision," so make it deliberately. Run your budget at the seat price and then add a usage estimate, because Enterprise usage is billed on top of seats.

### Why consumer plans are the wrong default for client files

Free, Pro and Max run under Anthropic's consumer terms, and those terms let the user's chats train future models unless the user opts out. Anthropic's [August 28, 2025 consumer-terms update](https://www.anthropic.com/news/updates-to-our-consumer-terms) asked every Free, Pro and Max user to choose. If you allow training, retention extends to five years for new or resumed chats. If you do not, Anthropic keeps the existing 30-day retention period.

The same page says the change does "not apply to services under our Commercial Terms," which include Team, Enterprise and the API. That is the core reason I treat Team as the floor for client work: the protective default is built into the contract instead of depending on each lawyer finding a toggle.

ABA Formal Opinion 512 tells lawyers to read and understand the terms of use and privacy policy of any generative AI tool they use, or to consult someone who has. Under Model Rule 1.6(c), a lawyer must make "reasonable efforts to prevent the inadvertent or unauthorized disclosure" of client information; a consumer account with training switched on is hard to square with that. The full ethics analysis appears in the ABA Opinion 512 section later in this guide.

Incognito mode does not fix the problem. Incognito chats are not used for training, but they are still retained for 30 days by default. They are a privacy convenience, not a substitute for commercial terms.

### Which Claude model is best for legal documents

Anthropic's own advice is to "start with Claude Opus 5.5 for most workloads" and use Fable 5.1 for "demanding reasoning and long-horizon agentic work," per the [Claude models overview](https://platform.claude.com/docs/en/about-claude/models/overview). The table below lists the current lineup as of October 2026, with my view of where each fits in a plaintiff firm.

| Model | Released | Context window | Best legal use (author's view) |
|---|---|---|---|
| Fable 5.1 | September 2026 | 1M tokens | Hard reasoning and long agentic jobs; on Pro only via usage credits |
| Opus 5.5 | Sept 22, 2026 | 1M tokens | Default for drafting, review and chronologies |
| Sonnet 5.5 | Sept 28, 2026 | 1M tokens | Everyday extraction and summaries |
| Haiku 5.5 | Oct 7, 2026 | 1M tokens | High-volume triage, such as sorting intake notes |

Opus 5.5 is the sensible default for demand drafts, record review and anything a lawyer will sign. Sonnet 5.5 is cheaper on the API and fine for pulling dates or provider names. Haiku 5.5 suits sorting large piles where a person reviews every result. All four have a reliable knowledge cutoff of June 2026, so none knows about later rulings without a research connector. Research in the Claude app runs on Opus 5.5 whatever model you select.

One wrinkle matters for firms that negotiate zero data retention (ZDR). Fable 5.1 shares its underlying model with Mythos 5.1, and Anthropic designates these as covered models: prompts and outputs are retained for 30 days on every platform, and covered models cannot be used in ZDR organizations unless Anthropic authorizes it. On Pro, Fable is available only through usage credits.

Older guides still say Claude's best model is "Claude 3 Opus" and that it handles "200,000+ tokens in Claude 2.1." Both statements are obsolete. As of October 2026 the current models carry 1M-token context windows, and some older models offered in chat have 500K or 200K.

### File, page and context limits for medical records

Medical records are where file limits bite. Anthropic's [file upload Help Center article](https://support.claude.com/en/articles/8241126-upload-files-to-claude), dated July 23, 2026, sets these limits as of October 2026:

- Chat uploads: 500MB per file, up to 20 files per chat, PDFs up to 1,000 pages.
- Project files: 30MB per file; no cap on file count, but the total must fit the context window.
- PDFs of 100 pages or fewer: Claude reads both text and visuals.
- PDFs of 101 to 1,000 pages: text only, so visual-only content is not analyzed.

The last line is the trap for injury files. A 600-page hospital production with scanned itemized bills, checkbox intake forms or handwritten notes will lose whatever exists only as an image. Claude will not tell you what it skipped unless you ask. Some guides say Claude accepts "up to 5 documents at a time"; the real chat limit is 20 files.

A practical routine: run OCR on scanned records first, split each set into PDFs of 100 pages or fewer, and keep a page-index sheet that maps each split file back to the original Bates or page numbers. That way every fact in a chronology can be traced to a page a reviewer can open.

For a sense of scale, Anthropic's [context window article](https://support.claude.com/en/articles/8606394-how-large-is-the-context-window-on-paid-claude-plans) says a 200K-token window holds "about 500 pages of text or more"; the 1M windows on current models hold far more. For contrast, EvenUp's [Claude customer story](https://claude.com/customers/evenup) describes reading pages with real text first, then standard OCR, then Claude's vision for the scanned bills and checkbox forms, and says a case "often arrives as 1,000 pages or more." That layered approach, attributed to EvenUp, is the same logic a firm can copy by hand.

## How to install and set up Claude for Legal

You install Claude for Legal in seven steps: choose a commercial plan, install plugins from Cowork, connect a research tool, run each plugin's cold-start interview, add the Microsoft 365 add-in, optionally deploy through Claude Code or Managed Agents, and lock down retention and connector settings before client data goes in.

1. Pick the plan and set up the workspace.
2. Install plugins in Cowork.
3. Connect a research tool first.
4. Run the cold-start interview.
5. Install Claude for Microsoft 365.
6. Install from Claude Code or deploy Managed Agents.
7. Lock down settings before client data goes in.

![Seven-step setup flow for Claude for Legal from choosing a plan to locking down retention settings](/images/blog-content/claude-for-lawyers/04-setup-flow.jpg "Set up in this order. The research connector and the cold-start interview do the most to improve output quality.")

The order matters more than any single step. A research connector and a finished practice profile change the quality of every later output, while the security step has to be done before the first real client file is uploaded.

### Step 1: Pick the plan and set up the workspace

Choose the plan with the picker above, then decide who administers it. On Team, owners and admins control membership, and the plan supports SSO and domain capture so anyone signing up with the firm's email domain lands in the firm workspace instead of a personal account. On Enterprise, the launch post says admins enable the legal connectors and plugins in workspace settings, so nothing appears for lawyers until an admin turns it on.

Before anyone touches a client file, create a test workspace or a dedicated Project that holds only public or dummy material: a published statute, a blank form, a fictional fact pattern. Use it to learn the tools. The first week of any rollout produces mistakes, and they should happen on material that cannot breach a confidence.

### Step 2: Install plugins in Cowork

The GitHub README gives three steps for Cowork, as of October 2026: open the Cowork tab, click **Customize** in the left sidebar, then click **Browse plugins** and install the ones you want, or upload a custom plugin file. For a plaintiff firm, Litigation Legal is the natural first install, with Legal Clinic a useful second for its client-letter patterns.

One caveat. On September 16, 2026 Anthropic announced that [Cowork and chat are now one Claude](https://claude.com/resources/articles/cowork-is-now-claude), rolling out to Pro and Max first with Team and Free to follow. On accounts that already have the unified Claude, the Cowork tab may look different. The Customize path above is the one the README documents as of October 2026; if your screen differs, look for the customization or plugin menu in the unified app.

### Step 3: Connect a research tool first

The README is blunt: "Connect a research tool first." Citations that come through a research connector are tagged with their source. Citations drawn from the model's own knowledge are flagged `[verify]`, and if no research tool is connected, the reviewer note records that sources weren't verified. That flag is useful, but it only helps if someone reads it.

The README names CourtListener, Trellis, Descrybe and Solve Intelligence as research tools. CourtListener, which covers federal dockets and opinions, is public, with an optional API key, so a firm with no research subscription can still start here. Trellis covers state court dockets and motions but needs the firm's own subscription. For a broader look at the options, see our [legal research tools compared](/blogs/legal-research-tools) guide.

### Step 4: Run the cold-start interview

Every plugin starts with a setup interview, and the README says to run it before anything else: "Every other skill in a plugin reads from the practice profile it writes. Skipping setup is the single most common reason a skill produces generic output." The command is `/<plugin>:cold-start-interview`, for example `/litigation-legal:cold-start-interview`. The interview takes 10 to 20 minutes per plugin, and a quick-start option makes you productive in about 2 minutes so you can refine later.

The interview writes a plain-English profile to `~/.claude/plugins/config/claude-for-legal/<plugin>/CLAUDE.md`. In the README's words, "It survives plugin updates." You can open and edit that file directly, which means the profile is a living firm document, not a one-time form.

Preparation makes the interview worth the 20 minutes. Here is the checklist I would bring to it:

- The jurisdictions where you file and the courts you appear in most.
- Your case types and the statute-of-limitations rules you track for each.
- Your house style for demand letters: tone, structure, exhibits, signature block.
- Your escalation chain: who reviews what, and who signs.
- Your risk calibration: which outputs always need a partner's review.
- The documents that must never go out without attorney review, such as demands, releases and anything filed.

### Step 5: Install Claude for Microsoft 365

Claude for Microsoft 365 installs from Microsoft AppSource. The README gives the listing ID `wa200010453`. Per Anthropic's [Microsoft 365 announcement](https://claude.com/resources/articles/collaborate-with-claude-across-excel-powerpoint-word-and-outlook), Claude for Word, Excel and PowerPoint became generally available on May 7, 2026, and Claude for Outlook entered public beta for all paid plans. One AppSource listing covers Word, Excel and PowerPoint; Outlook has a separate listing.

Organizations can sign in with a Claude account, or route traffic through an existing LLM gateway to Claude models on Amazon Bedrock, Google Cloud's Vertex AI or Microsoft Foundry. The README says every contract-touching skill is written to work in the Word sidebar with tracked changes as the output mode. Your IT administrator may need to approve the add-in before lawyers can install it, depending on how the firm's Microsoft tenant is configured. Approve it for a pilot group first, so the people who completed the cold-start interview are the first to use Claude inside real documents.

### Step 6: Install from Claude Code or deploy Managed Agents

Firms with a developer can install the same plugins from Claude Code. The README's sequence is `/plugin marketplace add <path-to-this-repo>`, then a plugin install such as `/plugin install litigation-legal@claude-for-legal`, a restart, and the cold-start interview. Updates come through `/plugin update`. The repository lives at [github.com/anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal).

Scheduled agents can also be deployed through the Claude Managed Agents API with `scripts/deploy-managed-agent.sh <agent>`, for example `scripts/deploy-managed-agent.sh docket-watcher`. Subagent delegation is labelled a research preview and supports a single delegation level. Anthropic also offers the Commercial, Corporate, Litigation and Product plugins as cookbooks that can be deployed as Managed Agents in the Claude Platform. This route is for firms building automation, and it bills at the API rates listed earlier.

### Step 7: Lock down settings before client data goes in

Do this step before the first real file, not after. My checklist, built from Anthropic's own documentation as of October 2026:

- Confirm the workspace runs on commercial terms (Team, Enterprise or API), not a personal consumer account.
- Set organization retention on Enterprise; the minimum is 30 days, and projects are retained indefinitely by default unless you change it.
- Review each connector's scopes and disable write tools you do not need.
- Turn off "always allow" and auto-approve settings for connector actions.
- Decide on HIPAA before any protected health information goes in, since only Enterprise can enable it.
- Avoid giving Cowork access to local folders that hold sensitive files.
- Write and circulate an AI use policy (a template appears in the rollout section later in this guide).

The connector item deserves care, because connectors are how Claude reaches a firm's system of record. If you plan to connect your [practice management software](/blogs/law-practice-management-software), confirm which actions the connector can take and keep write permissions off until someone has tested them.

Some older guides say "Claude doesn't store your files permanently." That is an oversimplification. Consumer chats are kept 30 days, or five years if training is allowed. Enterprise projects are kept indefinitely by default. Incognito chats are still retained for 30 days. Retention is a setting to configure, not a promise to assume.

## Where Claude works: chat, Cowork, Word and Outlook

Claude for Legal runs in four places: the Claude app (chat and Cowork), the Microsoft 365 add-ins, Claude Code, and the Claude Platform for custom builds. Most plaintiff lawyers will spend their time in the first two, so this section focuses there and on the workspace features that sit inside them.

### Chat, Cowork and the move to one Claude

Anthropic says Cowork "brings Claude Code's agentic capabilities to knowledge work beyond coding." In practice that means Claude can work through multi-step tasks, use connectors and act on files, not only answer questions. As of October 2026 Cowork is available on paid plans on Claude Desktop for macOS and Windows, and on the web and mobile for Pro, Max and Team, and for Enterprise where an admin has enabled it. Anthropic made Cowork [generally available on all paid plans](https://claude.com/resources/articles/cowork-for-enterprise) on April 9, 2026, adding Enterprise role-based access controls, group spend limits and usage analytics.

On September 16, 2026, Anthropic announced that Cowork and chat "are now one Claude," rolling out to Pro and Max over several weeks. Team and Free will follow, and Enterprise admins will get at least 30 days' notice before anything changes. Claude Docs and Claude Slides launched the same day in beta on paid plans. Because the rollout is still in progress, your firm may see either the separate Cowork tab or the unified app.

Anthropic's [guide to using Cowork safely](https://support.claude.com/en/articles/13364135-use-claude-cowork-safely) warns that it "has unique risks due to its agentic nature and internet access," including prompt injection, where malicious instructions hide in content Claude reads. It recommends human oversight in high-stakes scenarios. Every plaintiff matter qualifies.

### Claude for Word, Excel, PowerPoint and Outlook

The [Claude for Microsoft 365 page](https://claude.com/claude-for-microsoft-365) says Claude for Word edits with tracked changes, responds to comment threads and updates content using company styles. In Outlook, replies are drafts that wait for you to send: "Nothing goes out or gets saved until you say so." The launch post adds that Claude carries context across Word, Outlook, Excel and PowerPoint, so one thread can span all four.

Three plaintiff examples, offered as my suggestions rather than Anthropic's: redline a defense-proposed settlement release in Word and review each tracked change; build a medical specials spreadsheet in Excel from a finished chronology, then check the totals against the bills; triage a morning of adjuster emails in Outlook into drafts that wait for a lawyer. One limit for firms handling health information: Claude in Office is a beta feature excluded from Anthropic's BAA, a point covered in detail in the privilege and HIPAA section.

### Projects, Skills, Research and scheduled tasks

[Projects](https://support.claude.com/en/articles/9517075-what-are-projects) act as matter workspaces; the launch post says matter teams get a persistent space where precedents and prior drafts are retained. Free users can create up to five. Paid plans get retrieval-enhanced project knowledge that expands capacity by up to 10x, and sharing with view or edit permissions is limited to Team and Enterprise. A new beta version of Projects is rolling out, starting with Claude Code.

[Skills](https://support.claude.com/en/articles/12512176-what-are-skills) are folders of instructions, scripts and resources that Claude loads when relevant. They are available on every plan, need code execution turned on, and on Team and Enterprise owners can provision them for the whole firm. A firm's demand-letter format is a natural candidate.

[Research](https://support.claude.com/en/articles/11088861-use-research-on-claude) runs multiple searches that build on each other, on paid plans with web search enabled; on mobile it searches only the web and skips connectors. [Scheduled tasks](https://support.claude.com/en/articles/13854387-schedule-recurring-tasks-in-claude-cowork), also on paid plans, run remotely, so they keep their cadence while your computer sleeps. The launch post mentions weekly update sweeps and intake triage; for a plaintiff firm, a weekly docket sweep is an obvious first task.

The setup is the same for every practice; the next sections show what a plaintiff firm actually does with it.

## Claude workflows for personal injury firms

In a personal injury file, Claude is most useful for turning large, messy records into structured drafts: treatment chronologies, demand-package sections, liability memos and correspondence, each reviewed by a lawyer before it leaves the firm. None of Anthropic's agents is built for personal injury, so the workflows below adapt the Litigation Legal plugin and plain prompts to plaintiff work.

The table maps each PI task in this section to the plugin or agent that helps, what you feed it, and the check a person must run before the output is used. Agent names and commands come from the [Claude for Legal GitHub README](https://github.com/anthropics/claude-for-legal), as of October 2026.

| Task | Plugin or agent | Input | Human check |
|---|---|---|---|
| Research and statute summaries | Research connector | Statute, issue, jurisdiction | Pull every cited case |
| Medical chronology | Chronology Builder | Indexed records | Spot-check page cites |
| Demand package | Demand Letter Drafter | Chronology, bills, liability facts | Lawyer sets the number |
| Liability and fault | Plain prompt | Police report, photos, statements | Confirm doctrine and facts |
| Insurance letters | Plain prompt or Outlook | Claim file, adjuster letters | Read before sending |
| Releases and liens | Claude for Word | Release, lien notices | Lawyer reviews every clause |
| Deposition prep | Deposition Prep | Transcripts, theory, exhibits | Attorney edits outline |

Read the table as a division of labor. The middle columns are where Claude saves effort; the last column is where the firm keeps responsibility. Every row ends with a person, and the rows where money or legal positions are involved (the demand, the release, the liens) end with a lawyer, not a paralegal. If a task has no realistic human check, it is not ready for Claude.

### Legal research and statute summaries with a connector

Research is the one PI task where you should never rely on Claude's memory. Anthropic's README tells users to "Connect a research tool first," and it says citations from model knowledge alone come back flagged `[verify]`. For a plaintiff firm, the [CourtListener connector](https://claude.com/marketplace/connectors/courtlistener) covers federal opinions, PACER dockets and citation verification, and Trellis covers state court dockets and motions.

A practical example: ask Claude to summarize your state's comparative-fault statute, list the leading appellate cases applying it, and give a pinpoint citation for each holding. Then pull every case it cites and read the relevant passage yourself. Anything still marked `[verify]` stays out of a brief until a lawyer has confirmed it.

### Medical record chronologies

The chronology is the backbone of a PI file, and it is the task Claude handles best when the records are prepared well. The Chronology Builder agent (`/litigation-legal:chronology`) "builds or updates a chronology from declared sources and uploads". Take a hypothetical rear-end collision client, 41, with an ER visit, an orthopedic consult, an MRI and 14 weeks of physical therapy. The steps:

1. Split and OCR the records first. As of October 2026, chat accepts PDFs up to 1,000 pages, but Claude reads visuals only in PDFs of 100 pages or fewer; above that it reads text only, so scanned bills can be missed.
2. Upload with a page index: file name, provider, date range and Bates or page range for each document.
3. Ask for one row per visit with date, provider, visit type, chief complaint, diagnosis, treatment, bill amount and a page cite.
4. Ask Claude to flag gaps in care longer than a number of days you set, and any mention of pre-existing conditions or prior injuries.
5. Reconcile billing totals against the provider ledgers by hand. Never let a model-generated total go into a demand unchecked.

Plaintiff vendors describe the same difficulty. EvenUp's CEO, quoted on [Anthropic's legal solutions page](https://claude.com/solutions/legal), calls PI law one of the toughest AI challenges because it requires "reasoning across large volumes of medical records and billing data". EvenUp's principal ML engineer adds, in the [EvenUp customer story](https://claude.com/customers/evenup), that a subtly wrong argument "looks just as confident and polished as a correct one". That is why the page cite in every row matters: it lets a paralegal check a row in seconds.

Medical records are protected health information, so which plan you use matters; the HIPAA and BAA question is covered in the privilege and HIPAA section below.

![Sample medical chronology table with dates, providers, treatment, bill amounts and page cites, labeled as sample data](/images/blog-content/claude-for-lawyers/05-medical-chronology-mock.jpg "A chronology layout worth asking for: one row per visit with a page cite, plus flags for gaps in care. Sample data only.")

### Demand packages and the FRE 408 send gate

The Demand Letter Drafter (`/litigation-legal:demand-draft`) drafts "with FRE 408 awareness and a send gate" and produces a .docx file. The send gate means the agent stops before anything goes out, so the lawyer approves the final letter.

Federal Rule of Evidence 408 is why that gate exists. Under [FRE 408(a)](https://www.law.cornell.edu/rules/fre/rule_408), offers to compromise a claim, and statements made during compromise negotiations, are not admissible to prove or disprove the validity or amount of a disputed claim, or to impeach with a prior inconsistent statement. Under 408(b), a court may admit them for another purpose, such as proving a witness's bias. In plain terms: what you write in a demand can still surface for other purposes, so every factual statement in it must be accurate. Many states have their own version of the rule, so check yours.

A typical demand package has these sections, and Claude can draft each one from the chronology and file:

- Liability: how the collision or incident happened and why the defendant is responsible.
- Injuries: diagnoses tied to records.
- Treatment: a narrative built from the chronology.
- Special damages: medical bills and lost wages, from ledgers you have reconciled.
- Future care: drawn only from a treating provider's or expert's opinion in the file.
- General damages narrative: the effect on the client's daily life, from client statements.
- Policy-limits or time-limited demand language, if the lawyer decides to use it.

The demand figure belongs to the lawyer. Do not ask Claude to pick a number or suggest a settlement range, because it has no verified verdict data and will produce a plausible-sounding guess. If your team wants a rough reference point, use a separate, human-run tool such as our [state-specific MVA case value estimator](/tools/mva-case-value-estimator), which is an estimator and not a valuation, and treat its output as one input among many. EvenUp's customer story reports large gains in settlement offers, but that is a vendor claim with no disclosed method, and no firm should expect the same results.

### Liability and comparative fault analysis

Liability memos work well when you tell Claude which doctrine applies instead of letting it assume. Name the state and its rule: pure comparative fault, modified comparative fault with a 50 or 51 percent bar, or contributory negligence. Then ask Claude to organize the facts by element (duty, breach, causation, damages) and to list, in two columns, the facts that help the client and the facts the defense will use.

Take a hypothetical premises case: a client slips on a wet grocery-store floor near a freezer case, with no warning sign visible in photos, but she was looking at her phone. Claude can sort the photos, incident report and statements into a fact map and point out what is missing, such as inspection logs. Verify the doctrine and any cases through the research connector before relying on the memo.

### Insurance correspondence without invented values

Claude is useful for the routine letters that fill a PI week: records and bill requests, letters of representation, responses to adjuster questions, replies to reservation-of-rights letters, and follow-ups on unanswered requests. In Outlook, replies drafted by Claude are drafts that "wait for you to send", which suits this work.

One rule matters here. Do not ask Claude to "reference comparable case values" or cite verdicts to an adjuster unless you supply the verified data yourself. Without a source, the model may invent figures that look real. Keep examples and templates generic: use "[Carrier]" and "[Adjuster]" placeholders, not real company names, and have a person read every letter before it is sent.

### Settlement releases and lien summaries

Defense releases are where a careful redline protects the client. Claude for Word edits "with tracked changes", and Anthropic's README says contract-touching skills use tracked changes as the output mode. Upload the defense release and ask Claude to flag, for the lawyer's review:

- Medicare and lien language, including who is responsible for satisfying liens.
- Confidentiality and non-disparagement clauses.
- Indemnity and hold-harmless provisions.
- Tax characterization of the payment.
- Releases broader than the claim, such as unknown or unrelated claims.

For negotiators, ask for a lien summary table: each lienholder, the amount claimed, the statutory or contractual basis that the lawyer supplies, and the documents attached. Claude should not decide whether a lien is valid or reducible. That is a legal judgment, and the summary is only an organized starting point for it.

### Deposition and trial preparation

The Deposition Prep agent (`/litigation-legal:deposition-prep`) builds an "outline tied to case theory with docs and impeachment". For a plaintiff firm, three uses stand out. First, summarize a defense medical expert's prior testimony that you supply, organized by opinion and page. Second, generate cross-examination topics where the expert's report conflicts with the treating records. Third, have Claude play the witness so a client or junior lawyer can practice answering questions. Each output is a draft; the attorney decides which lines of questioning to use.

The same approach carries into trial preparation. Ask Claude to turn the chronology into a short timeline exhibit draft, to list every exhibit that supports each element of the claim, and to draft plain-language witness outlines for treating providers based on their own records. Ask it, too, to list the weaknesses a defense lawyer would raise in closing, using only facts in the file. That list is often the most useful output, because it shows where the case needs more evidence before trial. Confirm every exhibit reference against the exhibit list the court will actually see.

## Claude workflows for mass tort intake and MDLs

Mass tort work is a volume problem: thousands of claimants, each with exposure, diagnosis and date facts that must match a litigation's criteria. Claude can summarize and pre-screen at that scale, but a trained human must make every qualification decision.

The scale is visible in the Judicial Panel on Multidistrict Litigation's own numbers. The table shows actions pending in six large MDLs as of 1 October 2026, from the [JPML report on pending MDL dockets by actions pending](https://www.jpml.uscourts.gov/sites/jpml/files/Pending_MDL_Dockets_By_Actions_Pending-October-1-2026.pdf).

| MDL | Docket | Actions pending at 10/1/2026 |
|---|---|---|
| Johnson & Johnson Talcum Powder | MDL-2738 | 69,340 |
| Bard Hernia Mesh | MDL-2846 | 23,324 |
| AFFF | MDL-2873 | 15,264 |
| Hair Relaxer | MDL-3060 | 12,381 |
| Depo-Provera | MDL-3140 | 6,412 |
| GLP-1 RAs | MDL-3094 | 4,025 |

The same report counts 39 MDLs with 1,000 or more actions pending. These numbers are not a market-size claim. They show the paperwork problem: every one of those actions has a client file, intake record, medical proof and, in many MDLs, court-ordered forms. A firm with even a few hundred claimants in one docket handles more documents than a PI practice sees in years, which is where structured AI drafts help most.

### Intake call summaries

Most mass tort cases start with a phone call. If the firm records and transcribes calls (with consent and recording rules handled by the firm under its state's law), Claude can turn a transcript into a structured summary with fixed fields:

- Caller name and relationship to the injured person.
- Injury or diagnosis, as the caller describes it.
- Product, drug or exposure, with brand names as stated.
- Dates: first use or exposure, last use, diagnosis.
- Treating providers and facilities.
- Prior representation or prior claims.
- Red flags: statute-of-limitations concerns, gaps in memory, conflicting answers.

Eve Legal, a plaintiff-side platform, says in its [Claude customer story](https://claude.com/customers/eve-legal) that it transcribes intake calls and "scores each one". That is the vendor's description, not an independent result, but it shows how plaintiff software is using the model.

The summaries that tend to help most are the ones that match the fields your intake team already uses to sign or decline a case. Keep the field list identical to your intake form, so summaries drop straight into the CRM. For the full intake process, see [our mass tort intake guide](/blogs/mass-tort-intake-guide); firms that would rather hand off the calls can review our [mass tort intake services](/mass-tort-intake).

### Qualification screening against tort criteria

Screening is where Claude can save the most time and do the most harm, so the design matters. Write the criteria grid yourself before you write any prompt. A typical grid has these columns:

- Product or exposure (named brand, generic, or site).
- Exposure window (start and end dates the litigation accepts).
- Qualifying diagnosis.
- Diagnosis date.
- Age at diagnosis, where criteria set one.
- Statute of limitations status in the claimant's state.
- Proof available (pharmacy records, medical records, receipts).

Then ask Claude to score each intake summary against the grid and output, per criterion, "meets", "does not meet" or "unknown", with the transcript line or document page that supports the answer. Route every "unknown" and every borderline case to a trained screener. Never let Claude decline a claimant automatically: a missed diagnosis date or a misheard brand name can turn a valid claim into a rejected one, and the claimant may never call back.

Criteria change as litigations develop, so keep the grid current. Our page on [qualification criteria for each active litigation](/blogs/mass-tort-qualification-criteria) is one reference point, and firms that want screening handled outside the firm can look at [claimant screening and lead verification](/services/mass-tort-lead-verification). California's 2026 Practical Guidance names "autonomously facilitating client intake" as an example of agentic AI that needs care; the intake compliance section below covers what it requires.

### Plaintiff fact sheets and MDL paperwork

Many MDLs require plaintiff fact sheets, profile forms or census data, and the form and deadlines are set by the court's case management orders. Claude can pre-fill a draft from the client file, but give it the exact form the court ordered, not a generic version, because terminology and questions vary from MDL to MDL. Use [plain-English definitions of MDL terms](/glossary) when training new staff on these forms.

Ask for three things in every pre-fill: an answer to each question, a source cite (document and page) for each answer, and a blank, not a guess, where the file is silent. Blanks become a document-request list for the client. Claude can also build checklists for document collection (pharmacy records, medical records, product proof) for each claimant.

The firm, not the tool, signs and verifies. Under [Federal Rule of Civil Procedure 11](https://www.law.cornell.edu/rules/frcp/rule_11), a filer certifies that factual contentions have evidentiary support after a reasonable inquiry, and client verifications carry their own obligations. A lawyer or trained paralegal reviews every pre-filled answer against its source.

### Proof of use and proof of injury review

Proof review is document-heavy and repetitive, which suits Claude. Typical documents include pharmacy printouts, product identifiers and lot numbers, purchase records, employment or site records for exposure torts, and pathology or imaging reports. Ask for an evidence table with four columns: document, page, what it shows, and which criterion it supports.

Watch the file format. As of October 2026, Claude reads visuals only in PDFs of 100 pages or fewer, and for other document types it extracts text only. A scanned pharmacy receipt inside a long PDF may not be read at all. Split scanned proof into short PDFs or images, and have a person confirm every "proof found" row against the original.

### Client updates and docket monitoring at volume

Mass tort clients wait years, and they call when they hear nothing. Legal Clinic's Client Letter Drafter handles "routine client correspondence," including document requests and updates. The same pattern works for status letters: a lawyer approves a template for each stage of the litigation, Claude fills in the client-specific details in plain language, and a person reviews and sends.

For the court side, the Docket Watcher is a scheduled agent that can also be deployed through the Managed Agents API. Paired with the CourtListener connector, it can track federal MDL dockets. Scheduled tasks run remotely on paid plans, even when your computer is asleep. State-court consolidations may need the Trellis connector instead. Treat any docket summary as an alert to read the actual order, not a substitute for it.

### Building a custom mass tort skill

Skills are "folders of instructions, scripts, and resources that Claude loads dynamically," according to [Anthropic's Help Center article on skills](https://support.claude.com/en/articles/12512176-what-are-skills). In Claude for Legal, each skill lives in its own `SKILL.md` file and reads the practice profile. A firm can write its own mass tort skill without code. What to encode:

- The intake script fields, in the same order as your intake form.
- The criteria grid for each litigation the firm accepts.
- Disqualifiers that always send a file to human review.
- The output template (summary, scoring table, document requests).
- The escalation rule: who reviews unknowns, and how fast.

On Team and Enterprise, owners can provision skills to all users. Treat the skill like any firm document: version it, assign an owner, and review it at least quarterly or whenever criteria change. Our [mass tort launch checklist](/resources/mass-tort-launch-checklist) lists the decisions to make before a firm opens a new tort, and most of them belong in the skill.

## PI and mass tort prompt library with checks

A good legal prompt names the task, the case type, the jurisdiction, the facts, the output format and the checks Claude must run, and tells Claude to say when it does not know. The library below follows that pattern, and every prompt has a human check attached.

### The prompt framework

Most experienced users structure legal prompts in a similar way: say what you want done, give the context, and say what the result should look like. I recommend a six-part version for plaintiff work:

1. Action: the task, in one verb ("build", "score", "draft", "summarize").
2. Case type: motor vehicle, premises, product liability, or the specific mass tort.
3. Jurisdiction: the state, or the MDL and court.
4. Facts: the documents you uploaded and any facts Claude should assume.
5. Format: table columns, headings, length, reading level, file type.
6. Verification: what Claude must cite, flag or leave blank.

End every prompt with the same line: "If the file does not say, write UNKNOWN; do not guess." That single instruction cuts down the most dangerous failure in plaintiff work, which is a confident answer filling a gap in the record.

Be specific about content, too. Anthropic's [legal summarization guide](https://platform.claude.com/docs/en/about-claude/use-case-guides/legal-summarization) says "There is no single correct summary for any given document" and tells builders to identify the specific information they want included. A chronology prompt that names its columns beats one that asks for "a summary of the records."

### Eight full prompts for plaintiff work

Each prompt uses hypothetical facts and placeholders in square brackets. Adapt the fields to your own forms.

**1. Medical chronology**

> Build a medical chronology for [client initials], a [age]-year-old [case type] client in [state]. Use only the uploaded records and the page index. One row per visit: date, provider, facility, visit type, chief complaint, diagnosis, treatment, bill amount, page cite. Sort by date. Do not combine visits. If a field is not in the records, write UNKNOWN; do not guess.

**Check:** A paralegal opens ten random page cites and confirms each row matches the record.

**2. Gaps-in-care flag**

> Using the chronology above, list every gap of more than [number] days between treatment dates. For each gap, give the dates on both sides, the providers, and any explanation found in the records (missed appointment, insurance issue, travel). Also list any record mentioning prior injuries or pre-existing conditions, with page cites. If no explanation appears, write UNKNOWN.

**Check:** The case manager confirms each gap with the client before the demand is drafted.

**3. Demand liability section**

> Draft the liability section of a demand letter for a [case type] claim in [state]. Use the police report, witness statements and photos uploaded. Describe how the incident happened, the defendant's duty and breach, and causation, citing documents by name and page. Do not state a demand amount or settlement range. Flag any fact the defense could dispute.

**Check:** The handling attorney confirms every factual sentence against its source before signing.

**4. Comparative-fault fact map**

> [State] applies [pure comparative / modified comparative with a 50 or 51 percent bar / contributory negligence]. Using the uploaded file, make a two-column table: facts that support the client and facts the defense will use to assign fault to the client. Cite the document and page for each fact. List evidence that is missing and should be requested.

**Check:** An attorney confirms the stated doctrine through the research connector or a current source.

**5. Intake call summary to fields**

> Summarize this intake call transcript into these fields only: caller, relationship to injured person, injury or diagnosis, product or exposure, first use date, last use date, diagnosis date, treating providers, prior representation, red flags. Quote the transcript line that supports each field. If the caller did not say, write UNKNOWN. Do not give the caller any legal opinion.

**Check:** The intake supervisor compares the summary with the recording on a sample of calls each week.

**6. Mass tort criteria scoring with UNKNOWN**

> Score this intake summary against the criteria grid below for [litigation name]. For each criterion, answer meets, does not meet, or UNKNOWN, and quote the supporting line or page. Do not give an overall decision. List every UNKNOWN as a follow-up question for the claimant. [Paste criteria grid.]

**Check:** A trained screener makes the sign or decline decision; nothing is declined by the model.

**7. Plaintiff fact sheet pre-fill with source cites**

> Pre-fill the uploaded plaintiff fact sheet for [MDL name] from the client file. Answer each question using only the file, with the document and page for each answer. Leave the answer blank, and add it to a separate missing-information list, wherever the file is silent. Do not paraphrase the form's questions or skip any.

**Check:** A paralegal verifies each answer against its cite, and the client reviews before signing.

**8. Client status update**

> Draft a status update for [client first name] about [litigation or case stage]. Use the approved template and the facts below. Write at an eighth-grade reading level, in under 200 words. Explain what happened, what happens next, and what we need from the client. Make no predictions about outcome, timing of payment or amount. [Facts.]

**Check:** The lawyer approves the letter before a person sends it.

### The 30-prompt index

The index below lists 30 shorter prompt ideas across the plaintiff workflow, each with the one check that makes it safe to use. Rows 28 to 30 touch marketing, which raises advertising-rule questions covered in the intake and marketing compliance section.

| # | Task | Feed Claude | Ask for | Human check |
|---|---|---|---|---|
| 1 | Intake summary | Call transcript | Fixed intake fields | Compare to recording |
| 2 | Conflict flags | Intake summary, party list | Possible conflicts | Run formal conflict check |
| 3 | SOL flag | Dates, state | Dates needing review | Lawyer calculates deadline |
| 4 | Follow-up questions | Summary with UNKNOWNs | Claimant question list | Screener edits |
| 5 | Retainer explainer | Approved retainer | Plain-language summary | Lawyer approves |
| 6 | Criteria scoring | Summary, grid | Per-criterion scores | Screener decides |
| 7 | Proof gap list | Scores, documents | Missing proof list | Paralegal confirms |
| 8 | Criteria update diff | Old and new grids | Changed criteria | Lawyer confirms |
| 9 | Batch re-screen list | Claimant list, new grid | Files to recheck | Screener reviews |
| 10 | Record index | Record set | Page index | Spot-check pages |
| 11 | Chronology | Indexed records | Visit-by-visit table | Check page cites |
| 12 | Gaps in care | Chronology | Gap list | Confirm with client |
| 13 | Prior conditions | Records | Mentions with cites | Attorney assesses |
| 14 | Billing summary | Bills, ledgers | Charges by provider | Reconcile by hand |
| 15 | Liability section | Reports, statements | Draft section | Verify each fact |
| 16 | Damages narrative | Client statements | Daily-life narrative | Client confirms |
| 17 | Specials table | Reconciled bills | Specials table | Match ledgers |
| 18 | Adjuster reply | Adjuster letter | Draft response | Read before sending |
| 19 | Release review | Defense release | Flagged clauses | Lawyer redlines |
| 20 | Expert testimony | Prior transcripts | Opinions by page | Attorney reads key pages |
| 21 | Cross topics | Expert report, records | Conflict list | Attorney selects |
| 22 | Discovery responses | Requests, file | Draft responses | Lawyer finalizes |
| 23 | Fact sheet pre-fill | Court form, file | Cited answers | Verify each cite |
| 24 | Docket digest | New docket entries | Plain summary | Read the order |
| 25 | Status letter | Template, facts | Client update | Lawyer approves |
| 26 | Document request | Missing list | Client letter | Staff reviews |
| 27 | FAQ answers | Approved facts | Client FAQ draft | Lawyer approves |
| 28 | Ad copy review | Draft ad | Possible rule issues | Compliance review |
| 29 | Landing page check | Page text | Unsupported claims list | Lawyer approves |
| 30 | Intake script review | Script | Advice-like lines | Compliance review |

The pattern across the index is that Claude organizes and drafts while a named role decides. Intake rows end with a screener, records rows end with a page check, and anything that leaves the firm ends with a lawyer. If a new prompt does not fit that pattern, rewrite it until it does.

### Prompts to avoid

Some requests are risky no matter how well they are written:

- Asking for a case value or settlement range: Claude has no verified verdict data and will invent a plausible number.
- Asking for case citations with no research connector: model-only citations come back flagged `[verify]` and may not exist.
- Asking Claude to answer a claimant's or prospect's legal questions: that risks unauthorized practice under [ABA Model Rule 5.5](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_5_unauthorized_practice_of_law_multijurisdictional_practice_of_law/); the intake compliance section covers the rules.
- Pasting client PHI into a consumer plan: consumer terms and retention differ from commercial ones (see the privilege and retention section).
- Asking it to "make the injuries sound worse": a demand must match the records, and exaggeration damages credibility.
- Letting anything send automatically: every letter, filing and client message needs a person to approve it.

### When to turn a prompt into a skill

My rule of thumb: if more than one person uses a prompt every week, turn it into a skill. A skill keeps the instructions, output template and checks in one versioned place and reads the firm's practice profile, so every user gets the same structure. Use Projects for matter-specific knowledge, such as one client's records or one MDL's orders; on Team and Enterprise plans, projects can be shared with view or edit permissions.

A simple example: prompts 1 and 2 above, used daily by three paralegals, belong in one "PI chronology" skill with the column list, the gap threshold and the page-cite rule written once. Prompt 8 stays a prompt until the firm approves its stage templates, then becomes a skill too.

Every check in this library exists because Claude can be wrong in specific, predictable ways, and the next section lays out those failure modes and a verification workflow for catching them.

## Accuracy limits and how to verify Claude's output

Claude can invent citations, misread long records and apply the wrong jurisdiction's law, and it does so in confident prose. Every legal output needs a human verification pass before anyone relies on it.

![Five-step verification workflow for AI legal drafts: source, law, jurisdiction, numbers and send gates](/images/blog-content/claude-for-lawyers/06-five-gates-verification.jpg "The Five Gates: no Claude draft leaves the firm until a person has checked its sources, law, jurisdiction, numbers and send decision.")

Anthropic says the same thing in its own repository. The Claude for Legal README states that "Every output from these plugins is a draft for attorney review — not legal advice, not a legal conclusion, not a substitute for a lawyer," and lists its guardrails as source attribution, conservative defaults, surfaced jurisdiction assumptions and "explicit gates before anything is filed, sent, or relied on" ([Claude for Legal on GitHub](https://github.com/anthropics/claude-for-legal)). This section explains what that review has to catch and gives you a workflow for it.

### How Claude gets legal work wrong

The best-known failure is the fabricated or misattributed citation. ABA Formal Opinion 512 cites a Stanford study finding that leading legal-research generative AI systems "hallucinate between 17% and 33% of the time" ([ABA Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf)). That figure describes legal-research AI tools in general, as studied then, not Claude specifically. The point is the category of risk, not a score for one model.

The second failure is stale law. Anthropic's Help Center says Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 have training data to June 2026, and that the models "may not be aware of events or information that occurred after their respective cutoff dates" ([how up to date Claude's training data is](https://support.claude.com/en/articles/8114494-how-up-to-date-is-claude-s-training-data)). A statute amended in August, a new appellate decision or a fresh MDL order will not be in the model's memory. Only a research connector or a person reading current sources closes that gap.

The third is jurisdiction. Tort law, comparative fault rules, damage caps and limitation periods vary by state, and a prompt that does not name the forum invites the model to assume one. The README says the plugins surface jurisdiction assumptions; your reviewer still has to read them and correct them.

The fourth is the long record. Anthropic's upload guidance says Claude reads text and visuals in PDFs of 100 pages or fewer, and text only for PDFs of 101 to 1,000 pages ([upload limits](https://support.claude.com/en/articles/8241126-upload-files-to-claude)). In a large scanned medical file, a handwritten note, a stamped bill or a checkbox form past page 100 may never be analyzed. Split long records, or confirm the file's text layer before you trust a chronology built from it.

The fifth failure is the hardest to see: reasoning that is plausible and wrong. EvenUp's principal ML engineer is quoted on Anthropic's site saying, "If it builds an argument that's subtly wrong, that mistake looks just as confident and polished as a correct one" ([EvenUp customer story](https://claude.com/customers/evenup)). Eve's CEO is quoted saying that "in litigation, an authoritative-sounding hallucination is worse than no answer" ([Eve Legal customer story](https://claude.com/customers/eve-legal)).

So, plainly: one competing guide says Claude is "built for transparency and trust" with "no hallucinated facts." That is not accurate. No current language model, Claude included, guarantees accurate citations or facts, and Anthropic's own repository flags citations from model knowledge alone with a `[verify]` tag. Treat every output as a first draft from a fast, well-read assistant who has never seen your case file and sometimes misremembers.

### Sanctions and Rule 11

The cost of skipping verification is already on the record. In *Mata v. Avianca, Inc.*, No. 22-cv-1461 (S.D.N.Y. June 22, 2023), lawyers "were sanctioned for submitting a brief that cited non-existent judicial opinions made up by ChatGPT," as [Texas Ethics Opinion 705](https://www.legalethicstexas.com/resources/opinions/opinion-705/) describes it. The [sanctions order](https://law.justia.com/cases/federal/district-courts/new-york/nysdce/1:2022cv01461/575368/54/) imposed a $5,000 penalty jointly and severally and required letters to each judge falsely identified as the author of the fake opinions.

The governing duty is not new. Under [Federal Rule of Civil Procedure 11(b)](https://www.law.cornell.edu/rules/frcp/rule_11), a lawyer who presents a filing certifies, after "an inquiry reasonable under the circumstances," that the legal contentions "are warranted by existing law or by a nonfrivolous argument" and that the factual contentions "have evidentiary support." An AI draft does not change who signs. Rule 11(c)(1) adds that, "Absent exceptional circumstances, a law firm must be held jointly responsible for a violation committed by its partner, associate, or employee." That makes verification a firm-level process, not a personal habit.

State courts have their own analogues to Rule 11, so check your state's version before assuming the federal text controls a state-court filing.

### The Five Gates verification workflow

The Five Gates is a review checklist I built for plaintiff firms using Claude. It is my framework, not an Anthropic feature or a bar requirement, though it lines up with the gates the plugins themselves describe. Every draft that will be filed, sent or relied on passes all five, in order, and the reviewer records what they checked.

The table below sets out each gate, the question it answers, who should own it and what evidence the file keeps.

| Gate | Question | Who | Evidence kept |
|---|---|---|---|
| 1. Source | Does every fact trace to a page in our file? | Paralegal | Page cites on each fact |
| 2. Law | Has every citation been pulled and read? | Associate | Research log, `[verify]` flags cleared |
| 3. Jurisdiction | Does the rule fit this forum and date? | Associate | Forum and date noted |
| 4. Numbers | Do bills, dates and totals reconcile by hand? | Case manager | Reconciliation sheet |
| 5. Send | Has a lawyer approved release? | Responsible attorney | Signed approval note |

Gate 1 catches the most common plaintiff-side error: a chronology entry, injury description or treatment date that is not in the records. Ask Claude to put a page cite after every fact, then spot-check the cites against the source documents. Gate 2 follows the README's advice to connect a research tool first. Citations that come through a connector arrive tagged with their source, while citations from model knowledge alone carry `[verify]`, and none of those flags should survive into a final draft.

Gate 3 matters most in multistate dockets. A comparative fault rule, a notice requirement or a statute of limitations that is right in one state can be wrong next door, and it can be right last year and wrong now. Gate 4 is a hand check of arithmetic: medical specials, lien balances, wage loss and date intervals. Language models are not calculators, and a demand that misstates the bills undermines everything else in it.

Gate 5 is the send decision. It mirrors what Mark Pike, who leads the Claude for Legal rollout, told [Artificial Lawyer](https://www.artificiallawyer.com/2026/06/01/claude-for-legal-has-over-90-ai-agents/): the plugins have gates before anything is filed, sent or relied on, and "The lawyer reviews and verifies; the tooling is designed to make that review easier, never to skip it." In the plugin workflows, the demand drafter's FRE 408 check is one such gate. The Five Gates make the same idea the firm's rule for every output, plugin or not.

Log the review. A short note in the case management system (who checked which gates, on what date, against which version of the draft) costs a minute and gives managing lawyers the record they need to show reasonable supervision. That record supports the supervision duties under Opinion 512, covered in the next section.

How much to spot-check is a judgment call, and I would not trust any published percentage for it. In my view, the sensible approach is to check everything at first, then sample less only for task types where your own logs show the drafts are consistently clean, and to return to full review whenever the model, the plugin or the task changes.

## ABA Opinion 512 and state bar rules on AI

ABA Formal Opinion 512, issued July 29, 2024, says lawyers using generative AI must meet their existing duties of competence, confidentiality, communication, supervision, candor and reasonable fees. Many states have issued their own guidance, and your state's rules control.

This section summarizes rules; it is not legal advice. The ABA Model Rules are a template, and each state adopts, modifies and interprets them, so check your state's version of every rule cited here.

### What Opinion 512 requires

Opinion 512 tells lawyers to "fully consider their applicable ethical obligations, including their duties to provide competent legal representation, to protect client information, to communicate with clients, to supervise their employees and agents, to advance only meritorious claims and contentions, to ensure candor toward the tribunal, and to charge reasonable fees." It does not create a new rulebook for AI. It applies the existing Model Rules to a new tool.

**Competence.** The opinion says uncritical reliance on AI output, "without an appropriate degree of independent verification or review of its output," could violate Model Rule 1.1, and that these tools "cannot replace the judgment and experience necessary for lawyers to competently advise clients." This sits on top of [Rule 1.1, Comment 8](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_1_competence/comment_on_rule_1_1/), which asks lawyers to keep abreast of "the benefits and risks associated with relevant technology." Knowing how Claude fails is part of competence, not an optional extra.

**Confidentiality.** For "self-learning" AI tools, Opinion 512 says "a client's informed consent is required prior to inputting information relating to the representation into such a GAI tool." It adds that "merely adding general, boiler-plate provisions to engagement letters purporting to authorize the lawyer to use GAI is not sufficient." The opinion also sets a baseline: lawyers should "read and understand the Terms of Use, privacy policy, and related contractual terms and policies" of any tool they use, or consult someone who has. All of this flows from [Model Rule 1.6(c)](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_6_confidentiality_of_information/), which requires "reasonable efforts to prevent the inadvertent or unauthorized disclosure of, or unauthorized access to, information relating to the representation of a client."

**Communication.** Even where Rule 1.6 does not require informed consent, the lawyer "must separately consider whether other Model Rules, particularly Model Rule 1.4, require disclosing the use of a GAI tool in the representation."

**Supervision.** Under Model Rules 5.1 and 5.3, "Managerial lawyers must create effective measures to ensure that all lawyers in the firm conform to the rules of professional conduct." For AI, those measures mean a written policy, training, approved tools and plans, and a review process such as the Five Gates. The supervision duty covers paralegals, case managers and intake staff who use Claude, not just lawyers.

### Billing and fees for AI work, including contingency cases

Opinion 512's fee guidance is specific. Lawyers who bill hourly "must bill for their actual time"; the opinion's example is that 15 minutes spent entering information into an AI tool may be billed, along with the time spent reviewing the draft. A lawyer "may not charge a client to learn about how to use a GAI tool or service that the lawyer will regularly use for clients." With flat fees, if the tool makes the work much faster, "it may be unreasonable under Rule 1.5 for the lawyer to charge the same flat fee." General office overhead may not be billed as an expense.

California's [2026 Practical Guidance](https://www.calbar.ca.gov/Portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf) takes a similar line: "Hourly billing must reflect the time actually spent," and subscription fees for general-purpose AI tools "typically constitute overhead expenses … and thus should be absorbed within the lawyer's fee."

Most plaintiff work is contingent, which the opinion does not address directly. My reading, offered as analysis and not as a rule: Rule 1.5's reasonableness requirement still applies to a contingency fee, but none of the sources cited here says AI use reduces a contingency percentage. The sharper risk is case costs. If your fee agreement passes costs to the client, check whether a Claude subscription or usage charge counts as a case cost under the agreement or as overhead under the ABA and California guidance. In my view, a flat monthly seat almost always looks like overhead.

### State and local opinions to read

The table lists the opinions most often cited on lawyers' use of generative AI, with what each adds. Dates and points come from each source or, where marked, from ABA 512's citation.

| Jurisdiction | Opinion | Date | Key point |
|---|---|---|---|
| ABA | Formal Op. 512 | July 29, 2024 | Existing duties apply to GAI |
| Florida | Op. 24-1 | 2024 | Chatbots under advertising Rule 4-7.13 |
| Pennsylvania and Philadelphia | Joint Formal Op. 2024-200 | 2024 | Cited by ABA 512 |
| District of Columbia | Op. 388 | April 2024 | Understand GAI; client file duties |
| New York City | Formal Op. 2024-5 | Aug. 7, 2024 | Rules 1.1 to 7.3 analyzed |
| Texas | Op. 705 | Feb. 2025 | Texas rules applied to GAI |
| California | 2026 Practical Guidance | 2026 | Agentic AI; no autonomous filing |

[D.C. Ethics Opinion 388](https://www.dcbar.org/for-lawyers/legal-ethics/ethics-opinions-210-present/ethics-opinion-388) says lawyers "need to understand enough about how GAI works, what it does, and its risks and limitations" to be comfortable it will be accurate and will not breach confidentiality, and it flags duties regarding the client file. [NYC Bar Formal Opinion 2024-5](https://www.nycbar.org/reports/formal-opinion-2024-5-generative-ai-in-the-practice-of-law/) works through Rules 1.1, 1.4, 1.5, 1.6, 3.3, 5.1, 5.3, 7.1 and 7.3. Texas Opinion 705 was issued at the request of the State Bar of Texas Taskforce on Responsible AI in the Law.

California's 2026 guidance replaces its 2023 version and, at the California Supreme Court's request, addresses agentic AI. It names "autonomously facilitating client intake" as an agentic example and says lawyers "must not permit AI systems to autonomously file documents, communicate with the court, or make representations." That line matters for anyone wiring Claude agents into intake or filing workflows.

One limit on this table: the Florida and Pennsylvania/Philadelphia opinions were not read directly for this guide; they appear here as ABA 512 cites them. Read your own state's text, and check whether your state has issued newer guidance since these dates.

### Court standing orders and disclosure

Some judges require lawyers to disclose AI use in filings or to certify that AI-generated content was checked. This guide did not verify current standing orders, and they change often. Before filing, check the court's local rules and the assigned judge's standing orders, which California's guidance also advises. In an MDL, check the transferee court's case management orders as well.

## Privilege, confidentiality, retention and HIPAA

Whether Claude use stays privileged and confidential depends on the plan's terms, how the firm configures it and whether a lawyer directs the work. Commercial plans do not train on your content by default; consumer plans let the user choose, and only Enterprise can enable HIPAA.

Anthropic's Help Center has a dedicated article, [Using Claude for Legal Work: Privilege, Confidentiality, and How to Think About Configuration](https://support.claude.com/en/articles/15707726-using-claude-for-legal-work-privilege-confidentiality-and-how-to-think-about-configuration), dated August 26, 2026. It calls itself "Anthropic's perspective on our own products and the public record; it isn't legal advice," current "as of June 2026."

### Consumer versus commercial terms

The biggest single decision is which terms govern the account. Anthropic's August 28, 2025 [consumer terms update](https://www.anthropic.com/news/updates-to-our-consumer-terms) lets Free, Pro and Max users choose whether chats train models; opting in extends retention to five years for new or resumed chats, and otherwise the 30-day retention period continues. The same page says the update does not apply to services under Commercial Terms, including Team, Enterprise, the API, Amazon Bedrock and Google Cloud Vertex.

Under commercial terms, Anthropic says customer content "by default" is "not used to train our models," administrators can set an organization-wide retention period (the pricing page lists custom retention controls on Enterprise only), and API customers can use 30-day retention or, on eligible endpoints and models, Zero Data Retention. On human review, it says "Only content flagged by the automated safety system can be reached by a human reviewer," that each human read is logged, and that customer-managed encryption keys are "available for qualifying customers."

The table summarizes data handling by plan as of October 2026, from Anthropic's [pricing page](https://claude.com/pricing) and Help Center. Confirm current terms with Anthropic before you rely on it.

| Question | Free/Pro/Max | Team | Enterprise | API |
|---|---|---|---|---|
| Model training default | User opt-out | None by default | None by default | Not used by default |
| Retention control | 30 days, or 5 years if opted in | Not customizable (Enterprise-only per pricing page) | Custom, min. 30 days | 30 days or ZDR |
| Audit logs | No | No | Yes | See Anthropic |
| HIPAA-ready | No | No | Yes | See Anthropic |
| Human review | Flagged content only | Flagged content only | Flagged content only | Flagged content only |

Read the table as a floor, not a recommendation. The consumer column is the one Opinion 512's "self-learning" warning was written for: a setting a user can change is not a firm control. Anthropic's legal solutions page promises "full audit trails," but the pricing page lists audit logs on Enterprise only, so a Team firm cannot reconstruct who put what into Claude the way an Enterprise firm can.

### Heppner and the later work-product rulings

Courts have only begun to address whether AI use waives privilege or work product, and every account here is attributed because this guide did not read the opinions themselves. Anthropic's Help Center describes *United States v. Heppner* (S.D.N.Y. Feb. 2026) as having "held that a criminal defendant's own use of a consumer AI account—without his lawyers' involvement—wasn't privileged work product." Anthropic says "The court's reasoning turned on the consumer privacy terms and the absence of attorney direction."

Anthropic adds that the ruling "said nothing about lawyers using enterprise tools," and that its citations to *United States v. Adlman* and *United States v. Kovel* "left open that an enterprise-level AI used at an attorney's direction may be privileged."

The same article cites four later trial-level decisions it says went the other way: *Warner v. Gilbarco* (E.D. Mich.), *Morgan v. V2X* (D. Colo.), *Tate Group Automotive v. Legacy Automotive Capital* (Tex. Bus. Ct.) and *Assini v. Hayward* (N.Y. Sup. Ct.). Anthropic says each "held that AI-assisted litigation preparation is protected work product and that disclosing inputs to an AI tool does not, by itself, waive that protection." It also notes: "None of these are appellate decisions." A single district court ruling binds no one outside its case, and neither do these.

The practical takeaways, in my view, are modest and cheap. Use an account on commercial terms for anything touching a client matter. Document that a lawyer directed the AI work, in the file and in the prompt itself. And read the protective order: Anthropic says *Morgan* approved language permitting parties to upload confidential information into an AI tool "where the provider is contractually barred from training on, retaining, or disclosing confidential data." If you are negotiating a protective order in a mass tort, consider whether you need similar language before uploading defendant documents.

### Retention, ZDR and in-tenancy options

Retention settings decide how long your prompts and outputs exist on Anthropic's side, and the defaults are not always intuitive. On Enterprise, the [custom retention controls](https://support.claude.com/en/articles/10440198-configure-custom-data-retention-controls-for-enterprise-plans) have a "minimum retention period" of 30 days, and by default "projects are retained indefinitely, so chats inside projects are not deleted by your chat retention period." Setting retention requires the Primary Owner or Owner role.

[Incognito chats](https://support.claude.com/en/articles/12260368-use-incognito-chats) are not used for training but are still retained for 30 days by default, or longer under an organization's setting, and on Team and Enterprise they are included in organizational data exports available to Owners. For [covered models](https://support.claude.com/en/articles/15425996-data-retention-practices-for-covered-models), which include Fable 5 and Fable 5.1, prompts and outputs are retained for 30 days on every platform, and these models cannot be used from zero-data-retention organizations unless Anthropic authorizes it.

Some firms want data to stay in their own cloud. Anthropic says Claude via AWS Bedrock or Google Cloud Vertex AI "changes where your data resides and whom you contract with; it does not by itself switch off the safety monitoring." Its privilege article lists four things to weigh:

1. Whether you are on commercial terms, with a DPA and the no-training commitment.
2. Whether the matter's sensitivity calls for ZDR or an in-tenancy deployment.
3. How you will keep a lawyer in the loop, verify output against primary sources and document AI use.
4. Your own client-consent and engagement-letter practices, which several bar opinions address.

### HIPAA, PHI and what the BAA covers

Personal injury files are full of medical records, so HIPAA questions come up early. Anthropic's [HIPAA-ready Enterprise](https://support.claude.com/en/articles/13296973-hipaa-ready-enterprise-plans) page says the feature "is available for Enterprise plans only," and that "Team plans and individual plans (Free, Pro, and Max) can't enable HIPAA." Only the Primary Owner can accept the business associate agreement, a click-through in organization settings. Anthropic calls it "a one-way decision," and the self-serve BAA "is a standard agreement and can't be modified."

Turning on HIPAA does not cover every product. Anthropic says "Enabling HIPAA alone doesn't bring Claude Code or Cowork under your BAA." Those need an extra HIPAA configuration that is "currently only available to organizations with an Anthropic account team," and Anthropic's instruction is direct: "Don't use PHI with a product that isn't covered."

The [BAA coverage article](https://support.claude.com/en/articles/8114513-business-associate-agreements-baa-for-commercial-customers) sorts features into three groups as of October 2026:

- **Covered:** chat, Projects, Artifacts, file creation and code execution, Voice, Web Search and Research.
- **Available but not covered:** Skills (off by default once HIPAA is enabled), MCP connectors and Enterprise Search.
- **Excluded:** Claude Console, Cowork in the cloud, and beta features including Claude in Office, Claude Design, Claude Slides and Claude Docs.

That list has teeth for a plaintiff workflow. Connectors that send data to third parties are not covered, skills are off by default, and Claude in Office is excluded as a beta feature. A firm cannot assume the full Claude for Legal stack works with PHI just because the account is HIPAA-ready.

Whether HIPAA applies to your firm at all is a separate question. Under [45 CFR 160.103](https://www.ecfr.gov/current/title-45/subtitle-A/subchapter-C/part-160/subpart-A/section-160.103), a business associate includes a person who provides legal services "to or for" a covered entity where the service involves disclosing protected health information. Many plaintiff firms receive records from or on behalf of their own clients rather than as a service to a hospital or insurer, but arrangements vary, and some firms do work that changes the analysis. Ask counsel. And no Claude plan is "HIPAA certified"; there is no such government certification to claim.

### Client consent and engagement letters

Opinion 512 requires informed consent before client information goes into a self-learning tool and says boilerplate engagement-letter language is not enough. California's 2026 guidance adds that reasonable confidentiality efforts "require more than reliance on generalized marketing assurances." Consent has to be specific.

Here is a template of what an AI disclosure in a plaintiff engagement letter could explain. It is my starting point, not legal advice; adapt it to your state's rules.

- Which AI tools the firm uses, and for which tasks (for example, record summaries and first drafts).
- Which client information goes into them, and which never does.
- The plan's terms: commercial account, no training by default, retention period.
- That a lawyer reviews every AI-assisted output before it is relied on or sent.
- How the client can ask questions or opt out, and what opting out changes.

## AI intake and marketing compliance for plaintiff firms

When Claude touches leads, intake calls, ads or website content, plaintiff firms take on marketing and consumer-protection rules on top of ethics rules: TCPA consent for AI voices, Rules 7.1 to 7.3 for advertising, Rule 5.5 for answers to legal questions, and Google's rules for AI-written pages.

This is the part of the guide closest to Mass Tort Marketing Agency's daily work, which is the path from ad to lead to signed retainer, so the focus here is on where AI meets that path.

### AI voice and chat at intake: TCPA and Rule 5.5

The FCC settled the voice question in [Declaratory Ruling FCC 24-17](https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf), adopted February 2, 2024 and released February 8, 2024. It holds that the TCPA's restrictions on "artificial or prerecorded voice" encompass "current AI technologies that generate human voices," so such calls require "the prior express consent of the called party … absent an emergency purpose or exemption." An AI voice agent calling a claimant back is covered.

Separately, the FCC's 2023 "one-to-one" consent rule is not current law. In *Insurance Marketing Coalition Ltd. v. FCC*, the [Eleventh Circuit](https://media.ca11.uscourts.gov/opinions/pub/files/202410277.pdf) vacated that part of the order on January 24, 2025.

Bar rules add two open issues. [Model Rule 7.3, Comment 2](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_7_3_direct_contact_with_prospective_clients/comment_on_rule_7_3/) defines live person-to-person contact as real-time visual or auditory communication "where the person is subject to a direct personal encounter without time for reflection," but it does not say whether an AI voice agent qualifies. Treat that as unresolved until your state bar speaks. And under [Model Rule 5.5](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_5_5_unauthorized_practice_of_law_multijurisdictional_practice_of_law/), a lawyer must not practice in violation of a jurisdiction's rules "or assist another in doing so," which is the risk when an AI agent answers a claimant's legal question. California's guidance names autonomous intake as agentic AI that needs care.

My design rules for any AI intake tool: disclose that the caller or chat is AI, never let it give legal advice or predict outcomes, hand off to a human on any legal question, and log every transcript. For the full consent and advertising picture, see our guide to [TCPA and bar-advertising compliance for mass tort marketing](/mass-tort-marketing-compliance).

### Reviewing ads and landing pages against Rules 7.1 to 7.3

[Model Rule 7.1](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_7_1_communication_concerning_a_lawyer_s_services/) is short: "A lawyer shall not make a false or misleading communication about the lawyer or the lawyer's services." [Rule 7.2, Comment 5](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_7_2_advertising/comment_on_rule_7_2/) allows paying lead generators only if the generator does not recommend the lawyer, payments fit Rules 1.5(e) and 5.4, and its communications satisfy Rule 7.1. Opinion 512 also cites Florida's Op. 24-1 as discussing AI chatbots under Florida's advertising rules.

Claude makes a reasonable first-pass reviewer. Paste in the rule text and your state's advertising and filing requirements, then ask it to flag phrases that could mislead: guarantees, unverifiable superlatives such as "best" or "top-rated," past results without the required disclaimer, and implied specialization. Ask for each flag with the rule it may breach. Then a lawyer reviews the flags and the copy and signs off. Claude narrows the search; it does not approve the ad.

The same review applies to new channels. If you are weighing [advertising inside ChatGPT](/services/chatgpt-ads-for-lawyers) or other AI platforms, the ad copy and the landing page behind it are still lawyer advertising under Rule 7.1 and your state's version of it.

### AI-written website content and Google's rules

Google does not ban AI content. Its [spam policies](https://developers.google.com/search/docs/essentials/spam-policies) target scaled content abuse, pages generated "for the primary purpose of manipulating search rankings and not helping users … no matter how it's created," and list mass-producing pages with generative AI "without adding value for users" as an example. Its [guidance on generative AI content](https://developers.google.com/search/docs/fundamentals/using-gen-ai-content) adds that "It is critical to manually factcheck and review all AI-generated content for accuracy and trustworthiness before publishing."

For a law firm, that means the Five Gates apply to website pages too, plus attorney review for advertising rules. One small schema note: Google retired the FAQ rich result as of [May 7, 2026](https://developers.google.com/search/updates#removing-faq-rich-result), so FAQ markup no longer earns that display. For the broader strategy, see our guide to [mass tort law firm SEO](/blogs/mass-tort-law-firm-seo).

### How Claude finds and cites law firms

This guide is about using Claude inside a firm. Getting a firm named when a claimant asks Claude, ChatGPT or Google's AI Overviews for help is a separate marketing job with its own playbook. Claude's Research feature, for example, "runs multiple searches that build on each other" ([Research on Claude](https://support.claude.com/en/articles/11088861-use-research-on-claude)), so what it finds depends on what the web says about you across many sources, not only your own site.

If that side matters to your intake numbers, read our guide on [how personal injury firms get cited in Claude, ChatGPT and AI Overviews](/blogs/aeo-for-personal-injury-lawyers), and see our [AI search optimization for law firms](/services/ai-search-optimization) service. With the rules in place, the remaining question is how Claude stacks up against the other tools firms are weighing.

## Claude vs ChatGPT, Harvey and CoCounsel

Claude is a general assistant with legal plugins; Harvey and CoCounsel are legal-specific platforms that now also connect into Claude; plaintiff platforms such as EvenUp and Eve are built on Claude. The right choice depends on whether you need a general tool, a research database or a PI-specific product.

Most plaintiff firms will not pick just one. A firm can run Claude on a Team plan for drafting and summaries, keep its research subscription, and still buy a demand platform for high-volume auto cases. The comparison below sorts the options by what they are, not by which is "best", because the honest answer to "best" depends on the task and on who in the firm will maintain the setup. The table uses only what each vendor's own pages said when they were checked, and prices are as of October 2026.

| Tool | Type | Legal-specific content | Plaintiff focus | Pricing as published |
|---|---|---|---|---|
| Claude with Claude for Legal | General model plus open-source legal plugins | Via connectors | None built in | From $17/mo (Pro, annual) |
| ChatGPT | General model | Per OpenAI | None | See OpenAI |
| Harvey | Legal AI platform | Harvey's knowledge sources | Not described as PI-focused on pages checked | Quote-based |
| CoCounsel Legal | Thomson Reuters legal agent | Westlaw content | Not PI-specific | Requires a CoCounsel license |
| EvenUp | PI demand platform built on Claude | PI demand and records data | PI | Quote-based |
| Eve | Plaintiff-side platform built on Claude | Plaintiff case workflows | PI and plaintiff | Quote-based |

Read the table as a map of layers. Claude is the model and the workspace. Harvey and CoCounsel bring their own legal content and can be reached from inside Claude through connectors. EvenUp and Eve sit on top of Claude and add plaintiff-specific logic. The Claude price shown is the Pro plan on [Anthropic's pricing page](https://claude.com/pricing), which is $17 per month billed annually or $20 monthly as of October 2026; for client files, the plan picker earlier in this guide points most firms to Team or Enterprise instead. Every other price cell says quote-based or points to the vendor because none of those vendors published a list price on the pages checked.

### Claude vs ChatGPT for lawyers

Neither is better across the board, and the only fair test is your own redacted file run through both. On long documents, Claude's current models have a 1M-token context window as of October 2026, according to [Anthropic's models overview](https://platform.claude.com/docs/en/about-claude/models/overview), and chat accepts PDFs up to 1,000 pages. That matters for medical records. OpenAI publishes its own limits, and they change often enough that I have not repeated them here.

On research, Claude's plugins flag any citation that came from model knowledge alone with `[verify]`, as the [Claude for Legal README](https://github.com/anthropics/claude-for-legal) describes. ChatGPT has its own research and browsing modes; check how each one labels unsourced citations before you trust either.

On Office work, Claude for Word, Excel and PowerPoint reached general availability on May 7, 2026, and Outlook is in public beta for paid plans, per [Anthropic's Microsoft 365 announcement](https://claude.com/resources/articles/collaborate-with-claude-across-excel-powerpoint-word-and-outlook). Compare that against whatever Office integration your ChatGPT plan offers today.

On data defaults, the split that matters is consumer versus business. Claude's Free, Pro and Max plans let users opt in or out of training, while Team and Enterprise content is not used for training by default (the privilege section covers this). OpenAI also separates consumer and business terms; read its current terms before client data goes in, as ABA Formal Opinion 512 expects. A practical test: take one closed, redacted file, run the same three tasks in both tools, and push both outputs through the Five Gates described above. Count the errors caught at each gate. If your firm is also weighing ChatGPT as an advertising channel rather than a drafting tool, that is a separate decision covered in our page on [ChatGPT ads for personal injury firms](/services/chatgpt-ads-for-personal-injury-lawyers).

### Claude for Legal vs Harvey

Harvey and Claude for Legal are not strict alternatives, because Harvey also plugs into Claude. The [Harvey connector](https://claude.com/marketplace/connectors/harvey) was added in February 2026 and exposes tools such as Ask Harvey, Ask Harvey About Vault Project and List Knowledge Sources, with US, EU and AU endpoints. Anthropic announced it among new connectors in its [February 24, 2026 enterprise post](https://claude.com/resources/articles/cowork-plugins-across-enterprise).

Harvey's CEO is quoted on [Anthropic's legal solutions page](https://claude.com/solutions/legal) saying Claude Opus 4.7 scored 90.9% on Harvey's BigLaw Bench. That is a vendor-reported result on a partner's own benchmark, and it predates Opus 5.5. [Artificial Lawyer](https://www.artificiallawyer.com/2026/06/01/claude-for-legal-has-over-90-ai-agents/) has pointed to the trade-off: a platform like Harvey can mix models, while Claude for Legal ties you to Claude. Harvey's own marketing is aimed at large firms and in-house teams, so a plaintiff firm should ask its sales team directly how it handles PI records and demands.

### Claude vs CoCounsel and Westlaw

CoCounsel now runs on Claude's foundations: Anthropic says Thomson Reuters rebuilt CoCounsel Legal "from the ground up using the Claude Agent SDK" in its [enterprise agents post](https://claude.com/resources/articles/cowork-plugins-across-enterprise). You can reach it from Claude through the [CoCounsel Legal connector](https://claude.com/marketplace/connectors/cocounsel-legal), added in May 2026, and through the partner-built `cocounsel-legal` plugin in the GitHub repo.

Both need a paid CoCounsel Legal license and a Thomson Reuters account, and the connector "is offered as a pilot and may be modified, suspended, or withdrawn at any time." Its terms also bar routing Thomson Reuters content into another professional AI tool and warn against "always allow" or "auto-approve" settings. For a firm that already pays for Westlaw, the connector is the cleanest way to get sourced citations into Claude drafts. For a firm that does not, it is not a reason to start.

### Plaintiff platforms built on Claude: EvenUp and Eve

EvenUp and Eve are the two plaintiff-side platforms Anthropic features as Claude customers, and every number they publish is their own claim. On its [Claude customer story](https://claude.com/customers/evenup), EvenUp says its customers have resolved more than 200,000 cases and secured more than $10 billion in damages, and it reports cutting demand drafting from "8 to 15 hours of skilled work per case" to minutes. It also claims a 300% rise in settlement offers for firms presenting the full damages picture. No methodology is disclosed, so treat these as marketing figures.

[Eve's customer story](https://claude.com/customers/eve-legal) describes "an AI-native platform for plaintiff-side law firms" with over 1,700 firms, and says firms settle cases up to 60 days faster and increase settlement values by up to 30%. Eve tests models on its own benchmark, "PlaintiffBench." Its co-founder credits Claude's temporal reasoning for injury timelines. Litify also says on [its homepage](https://www.litify.com/) that it is built on Anthropic among other infrastructure.

Build or buy? In my view, buy a PI platform when demand volume and medical-billing logic are the bottleneck and nobody in the firm wants to maintain prompts. Build with Claude plugins and custom skills when your workflows are unusual, you run mass tort dockets the platforms do not model, and you have one person who will own the skills. Many firms will end up doing both.

### How to read legal AI benchmarks

Treat every published legal AI benchmark as a vendor document until someone independent repeats it. Harvey's BigLaw Bench is Harvey's. Eve's PlaintiffBench is Eve's. [GC AI's post on Claude for Legal](https://gc.ai/blog/claude-legal-ai) reports its own In-House Legal Bench from May 2026, scoring GC AI at 86.8%, ChatGPT (GPT-5.5) at 79.8% and Claude (Opus 4.7) at 68.4%; GC AI competes with Claude for Legal and notes the test predates Opus 5.5. This guide found no independent, current benchmark of Opus 5.5 on legal work as of October 2026.

Adoption data is thin too. The [ABA's 2024 Cloud Computing TechReport](https://www.americanbar.org/groups/law_practice/resources/tech-report/2024/2024-cloud-computing-techreport/) found 30% of respondents using AI-based tools, up from 11% in 2023, with accuracy (75%), reliability (56%) and data privacy (47%) as top concerns. Anthropic's [deployment guide](https://claude.com/resources/articles/deploying-claude-across-the-legal-industry) cites a 2026 FTI Consulting and Relativity report saying 87% of general counsel report generative AI use in their teams; that figure covers in-house lawyers, not law firms. For other products and how they compare, see our [full legal AI tools roundup](/blogs/legal-ai-tools).

## A 90-day Claude rollout plan for plaintiff firms

Roll Claude out in three 30-day stages: one pilot task on a commercial plan with a written policy, then skills and connectors with training, then measurement and a go or no-go decision. The structure keeps risk small early and makes the final decision depend on your own numbers rather than vendor claims.

![Three-stage 90-day rollout roadmap for Claude at a plaintiff law firm with pilot, skills and measurement phases](/images/blog-content/claude-for-lawyers/07-rollout-roadmap.jpg "A 90-day rollout keeps the first pilot small and measurable, then expands only after the numbers and the review log hold up.")

The table summarizes the three stages. Owners are suggestions for a 5 to 100 person plaintiff firm; a solo or small firm will collapse several roles into one person.

| Stage | Days | Goals | Owner | Exit test |
|---|---|---|---|---|
| Pilot | 1 to 30 | One task, one commercial plan, signed policy, baseline timing | Managing partner and ops lead | Policy signed, baseline logged, no client data on consumer plans |
| Expand | 31 to 60 | Cold-start interviews, research connector, one custom skill, training | Practice lead and IT | Skill passes the Five Gates on 10 closed files |
| Decide | 61 to 90 | Time study, error log review, go or no-go | Managing partner | Net time saved after review, error rate stable or falling |

The exit tests are deliberately boring. A pilot that cannot show a signed policy and a baseline at day 30 is not ready to expand, however impressive the drafts look. And the day-90 test is net time after review, because a fast draft that takes an hour to fix has saved nothing.

### Days 1 to 30: one pilot task, one plan, one policy

Pick one task with high volume and easy verification. Intake call summaries and first-draft medical chronologies both fit: they recur daily, and a paralegal can check them against the source in minutes. Avoid demand letters and briefs as a first pilot, since errors there are costly and slow to spot.

Put the pilot on Team or Enterprise, not a personal Pro account, so that training is off by default and the firm controls the workspace. Before anyone uploads a client document, have every pilot user sign the AI use policy below and walk through the Five Gates described above on a sample file.

Start the baseline time study in week one. Time the chosen task done the current way on 10 to 20 matters before Claude touches it. Without that baseline, day 90 becomes an argument about impressions. Anthropic's [deployment guide](https://claude.com/resources/articles/deploying-claude-across-the-legal-industry) includes its own three-phase adoption plan, which is worth reading alongside this one as further reading.

### Days 31 to 60: skills, connectors and training

Run the cold-start interview for each plugin you will use. The README says it takes 10 to 20 minutes per plugin, and that skipping it "is the single most common reason a skill produces generic output." Connect a research tool next, so citations arrive tagged with their source. If the firm runs Filevine or MyCase, test the [Filevine connector](https://claude.com/marketplace/connectors/filevine-api) or the [8am MyCase connector](https://claude.com/marketplace/connectors/8am-mycase) read-only on a few matters first.

Build one custom skill from the pilot task: the instructions, the output template and the verification checklist your reviewers already use. Test it on 10 closed files before live use.

Training is a duty, not a perk. [Model Rule 1.1, Comment 8](https://www.americanbar.org/groups/professional_responsibility/publications/model_rules_of_professional_conduct/rule_1_1_competence/comment_on_rule_1_1/) asks lawyers to keep up with "the benefits and risks associated with relevant technology," and ABA Opinion 512 says a lawyer may not bill a client for learning a tool used regularly. Beyond Anthropic's own resources, independent options include [claudeforlawyers.com](https://claudeforlawyers.com/), which describes itself as an independent guide with "105 tested prompts" and live team training, and Ernie Svenson's [Attorney at Work column](https://www.attorneyatwork.com/claude-ai-for-lawyers/) on switching to Claude. Those are their own descriptions, not endorsements.

### Days 61 to 90: measure and decide

Decide with a time study, not a feeling. Log every pilot task in a shared sheet with these fields:

- Task type (for example, intake summary or chronology draft)
- Matter ID
- Minutes without Claude (from the baseline sample)
- Minutes with Claude, including review time
- Errors caught at each gate (Source, Law, Jurisdiction, Numbers, Send)
- Rework minutes after the first review
- Reviewer name

For plaintiff firms, three operational metrics matter more than raw minutes. Track the days from records-complete to demand sent, the time from first intake contact to signed retainer, and the error rate at Gate 2 (law) and Gate 4 (numbers). If error rates at those two gates are not stable or falling by day 90, slow down rather than expand.

ROI has an acquisition side too. Faster file work only pays if the cases arriving are worth working, so compare the pilot's savings with your [cost per signed retainer benchmarks](/blogs/cost-per-signed-retainer-2026) and your [lead-to-retainer conversion rates](/mass-tort-lead-conversion-rates) by lead source.

One readiness check before scaling: clean the case-management data. AI speeds up whatever system it touches, including a messy one. If matter fields, document names and deadlines are inconsistent, connectors will pull inconsistent context into every draft.

### AI use policy template

This is the author's template, not legal advice; adapt it with your ethics counsel and your state's rules. It rests on the supervision duty in [ABA Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf): managerial lawyers "must create effective measures" so everyone in the firm follows the rules.

1. **Approved tools and plans.** List each approved tool and plan by name (for example, Claude Team or Enterprise), with the date it was approved.
2. **Prohibited plans for client data.** No client information in Free, Pro or Max accounts, or in any personal account.
3. **Data classes.** Define public, internal, client-confidential and PHI, and state which approved tool may receive each class. PHI only where a signed BAA covers the specific feature.
4. **Review gates.** Every output passes the Five Gates before it is filed, sent or relied on, and the reviewer is recorded.
5. **Citation rule.** No citation leaves the firm unless a person has pulled and read the cited source; `[verify]` flags must be cleared.
6. **Client consent and disclosure.** Get informed consent where Opinion 512 requires it, never by boilerplate alone, and consider Rule 1.4 disclosure case by case.
7. **Billing.** Bill actual time only; never bill clients for learning the tool; treat general subscriptions as overhead unless your jurisdiction says otherwise.
8. **Connectors and write permissions.** Only admin-approved connectors; read-only by default; no "always allow" on write tools; no sensitive local folders in Cowork.
9. **Retention.** Retention settings are set by the firm's admin and documented, including how project chats are handled.
10. **Incidents and review.** Report any suspected data exposure or bad citation to the named owner within one business day; review this policy at least annually and whenever plans or models change.

Clause 8 follows Anthropic's [Cowork safety guidance](https://support.claude.com/en/articles/13364135-use-claude-cowork-safely), which warns that write tools "inherently carry more risk" and advises against granting access to local files with sensitive information. Clause 9 matters because on Enterprise, per the [retention controls article](https://support.claude.com/en/articles/10440198-configure-custom-data-retention-controls-for-enterprise-plans), projects are kept indefinitely by default and are not deleted by the chat retention period.

### Security checklist for the managing partner and IT

Work through this list before the pilot opens to client data. Plan features are as of October 2026.

- Turn on SSO and domain capture, available from the [Team plan](https://support.claude.com/en/articles/9266767-what-is-the-team-plan).
- On Enterprise, set up role-based access, SCIM and audit logs.
- Set and document the retention period; the minimum on Enterprise is 30 days.
- On Enterprise, use IP allowlisting if the firm restricts access by network.
- Review every connector's scopes; never rely on auto-approve, as the CoCounsel terms also warn.
- Make the HIPAA decision before any PHI goes in. Only Enterprise can enable the [HIPAA-ready configuration](https://support.claude.com/en/articles/13296973-hipaa-ready-enterprise-plans), and accepting the BAA is a one-way decision.
- Restrict Cowork's local file access to non-sensitive folders.
- Approve the Microsoft 365 add-ins centrally, and decide whether to route traffic through an existing LLM gateway.

Claude can make a plaintiff firm faster at working the cases it already has. It does nothing about whether good cases arrive in the first place, which still depends on qualified, compliant intake; Mass Tort Marketing Agency's [24/7 live legal intake](/services/intake-services) is one way firms cover that side, and our piece on [where AI-routed intake is taking plaintiff acquisition](/blogs/future-of-plaintiff-acquisition) looks at what comes next.

## Sources and methodology

This guide was researched in October 2026 from Anthropic's own documentation, court and agency records, bar opinions and named secondary sources; every price, plan, model and connector fact was last verified on 22 October 2026.

Primary sources were read directly: claude.com product, pricing, marketplace and customer pages; support.claude.com Help Center articles; platform.claude.com model documentation; anthropic.com announcements; the GitHub README; ABA, D.C., New York City, Texas and California bar texts; FCC and Eleventh Circuit documents; the JPML pending-dockets report; Cornell LII rule text; and the eCFR. Secondary sources, including Artificial Lawyer and GC AI, are attributed by name wherever they are used.

Some sources could not be read and were handled cautiously. The Florida and Pennsylvania opinions are described only through ABA Opinion 512's citations. *United States v. Heppner* and the later work-product decisions are described only through Anthropic's Help Center summary, and none of them is presented as binding. OpenAI's current plans and defaults were not verified for this guide, so ChatGPT is discussed qualitatively. Vendor results (EvenUp, Eve, Harvey, GC AI) are reported as their claims.

Key dated sources:

- Anthropic, [Claude for the legal industry](https://claude.com/blog/claude-for-the-legal-industry), May 12, 2026
- Anthropic, [claude-for-legal GitHub repository](https://github.com/anthropics/claude-for-legal), accessed Oct 22, 2026
- Anthropic, [Claude pricing](https://claude.com/pricing), accessed Oct 22, 2026
- Anthropic Help Center, [Using Claude for Legal Work: Privilege, Confidentiality, and How to Think About Configuration](https://support.claude.com/en/articles/15707726-using-claude-for-legal-work-privilege-confidentiality-and-how-to-think-about-configuration), Aug 26, 2026
- Anthropic, [Models overview](https://platform.claude.com/docs/en/about-claude/models/overview), accessed Oct 22, 2026
- Anthropic, [Updates to our consumer terms](https://www.anthropic.com/news/updates-to-our-consumer-terms), Aug 28, 2025
- Anthropic, [Deploying Claude across the legal industry](https://claude.com/resources/articles/deploying-claude-across-the-legal-industry), May 15, 2026
- American Bar Association, [Formal Opinion 512](https://www.americanbar.org/content/dam/aba/administrative/professional_responsibility/ethics-opinions/aba-formal-opinion-512.pdf), July 29, 2024
- State Bar of California, [Practical Guidance for the Use of Generative AI](https://www.calbar.ca.gov/Portals/0/documents/ethics/Generative-AI-Practical-Guidance.pdf), 2026
- FCC, [Declaratory Ruling FCC 24-17](https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf), Feb 8, 2024
- U.S. Court of Appeals for the Eleventh Circuit, [Insurance Marketing Coalition v. FCC](https://media.ca11.uscourts.gov/opinions/pub/files/202410277.pdf), Jan 24, 2025
- JPML, [Pending MDL Dockets by Actions Pending](https://www.jpml.uscourts.gov/sites/jpml/files/Pending_MDL_Dockets_By_Actions_Pending-October-1-2026.pdf), Oct 1, 2026
- ABA, [2024 Cloud Computing TechReport](https://www.americanbar.org/groups/law_practice/resources/tech-report/2024/2024-cloud-computing-techreport/), 2024
- Artificial Lawyer, [Claude For Legal Has Over 90 AI Agents](https://www.artificiallawyer.com/2026/06/01/claude-for-legal-has-over-90-ai-agents/), June 1, 2026

Update policy: pricing, models and connectors are due for a monthly re-check, and this page should show the new verification date when anything changes. Anthropic's pricing page itself notes that prices and plans are "subject to change at Anthropic's discretion."

Disclosure: the author's agency does lead generation and intake for plaintiff firms. It does not sell Claude or any AI software, is not a law firm, and nothing here is legal advice.

## Frequently asked questions

### What is Claude for Legal?

Claude for Legal is Anthropic's set of open-source legal plugins and connectors for Claude, launched on [May 12, 2026](https://claude.com/blog/claude-for-the-legal-industry). The launch added 12 practice-area plugins, such as Litigation Legal and Legal Clinic, and 20+ MCP connectors that link Claude to legal research, document and practice-management software. It runs inside Claude (Cowork, chat, Claude Code and the Microsoft 365 add-ins). It is not a separate legal model; it is Claude configured for legal work.

### Is Claude for Legal free?

Partly. The plugins and connectors are open source under Apache-2.0, so the code costs nothing. Using them as documented needs a paid Claude plan, because Cowork, Claude Code and the Microsoft 365 add-ins are not on the Free plan as of October 2026, per [Anthropic's pricing page](https://claude.com/pricing). Some connectors, such as CoCounsel Legal and Harvey, also need the vendor's own paid subscription. Budget for a plan, not for the plugins.

### How much does Claude cost for lawyers?

As of October 2026, Pro costs $17 per month billed annually or $20 monthly, and Max starts at $100 per month. Team runs $20 per standard seat per month billed annually ($25 monthly) for 2 to 150 people, with premium seats at $100 billed annually ($125 monthly). Enterprise is $20 per seat per month billed annually with a 20-seat minimum, plus usage at API rates. Confirm current prices on [Anthropic's pricing page](https://claude.com/pricing) before you budget.

### Is Claude AI good for legal work?

It is useful for drafting, summarizing and organizing legal material, provided a lawyer verifies everything. Anthropic's own README calls every plugin output "a draft for attorney review," and flags citations from model knowledge alone with `[verify]`. ABA Opinion 512 cites a Stanford study finding legal-research AI tools hallucinate between 17% and 33% of the time. Use Claude for first drafts and structure, and keep a written review step before anything is filed or sent.

### Which Claude is best for legal documents?

Opus 5.5 is the sensible default as of October 2026. [Anthropic's models overview](https://platform.claude.com/docs/en/about-claude/models/overview) says to "start with Claude Opus 5.5 for most workloads" and to use Fable 5.1 for "demanding reasoning and long-horizon agentic work." All four current models have a 1M-token context window. On Pro, Fable is available only through usage credits, so most firms will draft on Opus 5.5 and reserve Fable for the hardest analysis.

### What version of Claude should a lawyer get?

For client work, a commercial plan: Team for most small and mid-size plaintiff firms, Enterprise if you need audit logs, custom retention or a HIPAA-ready setup. Free, Pro and Max are consumer plans where training is opt-out, which makes them a poor default for client files. A solo lawyer testing on public material can start on Pro. The plan picker earlier in this guide walks through the choice by firm size and data type.

### How do I install Claude for Legal?

In Cowork, open the Cowork tab, click Customize, then Browse plugins and install the ones you want, as the [GitHub README](https://github.com/anthropics/claude-for-legal) describes. In Claude Code, run `/plugin marketplace add` with the repo, then `/plugin install litigation-legal@claude-for-legal`, restart, and run the cold-start interview. Install the Word, Excel and PowerPoint add-in from Microsoft AppSource; Outlook is a separate listing. Enterprise admins can enable plugins in workspace settings.

### Is the Claude legal plugin free?

Yes, the plugin code is free. Anthropic's original "Legal" plugin in the [Claude marketplace](https://claude.com/marketplace/plugins/legal) and the 12 newer practice-area plugins are open source. What costs money is the Claude plan you need to run them, and any third-party connector subscription. The Legal plugin page also states that "All outputs should be reviewed by licensed attorneys," which applies whatever plan you use.

### Can attorneys use Claude ethically?

Yes, if they meet the duties ABA Formal Opinion 512 lists: competence, confidentiality, communication, supervision, meritorious claims, candor and reasonable fees. In practice that means reading the tool's terms, verifying outputs, getting informed consent where the opinion requires it, and supervising staff who use it. State rules vary, so check your own bar's guidance. This is general information, not legal advice for your situation.

### Is Claude or ChatGPT better for lawyers?

Neither wins on every task, and plans change quickly. Claude's strengths for plaintiff work include a 1M-token context window on current models, legal plugins with built-in review gates, and Microsoft 365 add-ins. ChatGPT has its own research, Office and business-plan features, which you should check on OpenAI's site. The fair test is one redacted closed file, the same tasks in both, and a count of errors each produces under the Five Gates.

### Claude for Legal vs Harvey: which should a firm choose?

They overlap less than it seems. Harvey is a legal AI platform with its own knowledge sources, and since February 2026 it also connects into Claude through the [Harvey connector](https://claude.com/marketplace/connectors/harvey). Harvey's CEO reports that Opus 4.7 scored 90.9% on Harvey's own BigLaw Bench, a vendor benchmark. A plaintiff firm without a Harvey license can start with Claude for Legal, then ask Harvey directly how it handles PI records before buying.

### Is using Claude covered by attorney-client privilege?

It depends on jurisdiction, configuration and attorney direction, and no appellate court has settled it. [Anthropic's Help Center](https://support.claude.com/en/articles/15707726-using-claude-for-legal-work-privilege-confidentiality-and-how-to-think-about-configuration) describes *United States v. Heppner* as holding that a defendant's own consumer-account use, without his lawyers, was not protected. It also cites later decisions it says went the other way for attorney-directed work. Anthropic adds that none is appellate. Ask your own ethics counsel.

### Does Claude train on my client data?

On commercial plans, not by default. Anthropic says content from Claude for Work, the Claude Platform and the API is not used to train its models by default. On Free, Pro and Max, users choose whether chats are used for training, and opting in extends retention to five years under the [August 2025 consumer terms update](https://www.anthropic.com/news/updates-to-our-consumer-terms). Keep client files on Team or Enterprise.

### Is Claude HIPAA compliant?

Only a specific Enterprise configuration is HIPAA-ready. Per [Anthropic's Help Center](https://support.claude.com/en/articles/13296973-hipaa-ready-enterprise-plans), Team and individual plans cannot enable HIPAA, and accepting the BAA is a one-way decision. The BAA covers chat, Projects, Research and some other features, but not connectors, Skills, Cowork in the cloud or beta features. Whether a PI firm is a business associate at all depends on the facts.

### Can Claude summarize medical records for a personal injury case?

Yes, with limits you should know. Chat accepts PDFs up to 1,000 pages and 20 files per chat, but for PDFs over 100 pages Claude reads text only, so scanned pages without a text layer may be skipped. The Litigation Legal plugin's Chronology Builder can draft a chronology, though none of the agents is built specifically for personal injury. A paralegal should check every entry against the page it cites.

### Can Claude write a demand letter?

It can draft one. The Litigation Legal plugin includes a Demand Letter Drafter that, per the README, works "with FRE 408 awareness and a send gate," and outputs a .docx draft. [Federal Rule of Evidence 408](https://www.law.cornell.edu/rules/fre/rule_408) limits using compromise offers to prove or disprove a claim. Do not ask Claude to invent verdict or settlement values; supply your own verified damages and specials, and have a lawyer approve the final letter.

### Can Claude screen mass tort leads?

It can help organize screening, but a person should make every qualification decision. A practical setup gives Claude the litigation's written criteria as a grid (product, exposure window, injury, diagnosis timing, proof available) and asks it to mark each intake answer as met, not met or missing, with the transcript line it relied on. An intake specialist or attorney then reviews the flags before any claimant is accepted or declined.

### Does Claude make up case citations?

It can, which is why verification matters. The plugins flag citations from model knowledge with `[verify]` and recommend connecting a research tool first. In *Mata v. Avianca* (S.D.N.Y. 2023), lawyers were sanctioned $5,000 for filing a brief citing opinions invented by ChatGPT, as [Texas Opinion 705](https://www.legalethicstexas.com/resources/opinions/opinion-705/) recounts. Pull and read every cited case yourself before anything is filed.

### Can I bill clients for time spent using Claude?

You can bill the actual time spent, but not more. ABA Opinion 512 says hourly lawyers "must bill for their actual time," including time inputting information and reviewing the draft, and may not charge clients to learn a tool used regularly. California's 2026 guidance treats general AI subscriptions as overhead. On contingency matters the fee question is about reasonableness overall, so check your state's rules.

### Do I have to tell clients I use AI?

Sometimes. ABA Opinion 512 says informed consent is required before putting client information into a self-learning AI tool, and that boilerplate engagement-letter language is not enough. Where consent is not required, lawyers must still consider whether Rule 1.4 calls for disclosure. Many firms now describe their AI use in the engagement letter in plain terms and get specific consent for higher-risk uses.

### Can an AI voice agent call or answer leads for my firm?

Outbound AI voice calls need care. The FCC's [Declaratory Ruling 24-17](https://docs.fcc.gov/public/attachments/FCC-24-17A1.pdf) treats AI-generated voices as "artificial or prerecorded voice" under the TCPA, so such calls need prior express consent. The separate one-to-one consent rule was vacated by the Eleventh Circuit in January 2025. Whether an AI voice is "live person-to-person contact" under Model Rule 7.3 is an open question for state bars.

### What AI do most lawyers use?

No reliable figure shows which tool most lawyers use. The [ABA's 2024 survey](https://www.americanbar.org/groups/law_practice/resources/tech-report/2024/2024-cloud-computing-techreport/) found 30% of respondents using AI-based tools, up from 11% in 2023, without a breakdown that settles the Claude versus ChatGPT question. Anthropic cites a 2026 FTI Consulting and Relativity report that 87% of general counsel report AI use in their teams, which covers in-house lawyers only.

### Where is the Claude for Legal GitHub repo?

It is at [github.com/anthropics/claude-for-legal](https://github.com/anthropics/claude-for-legal), published by Anthropic under the Apache-2.0 license with the description "A suite of plugins for legal workflows." As of October 2026 it has nearly 10,000 stars. The README lists the plugins, the 90+ named agents, install steps for Cowork and Claude Code, and the guardrail notice that every output is a draft for attorney review.

### Does Claude for Legal work with Clio, Filevine or MyCase?

Filevine and MyCase, yes; Clio, not officially as of October 2026. The [Filevine connector](https://claude.com/marketplace/connectors/filevine-api) was added in September 2026 and the [8am MyCase connector](https://claude.com/marketplace/connectors/8am-mycase) in August 2026; both respect the connected user's existing permissions. No official Clio, Smokeball or Litify connector appeared in Claude's directory when checked on October 22, 2026. Recheck the directory before you plan around one.