Claude for Lawyers: What 20,000 Legal Pros Asked Anthropic
Anthropic's legal webinar drew 20,000 registrants and 51 upvoted questions. What lawyers asked about privilege, hallucinations and Claude, answered.
Key facts
- The Claude for Legal Teams webinar, presented by Mark Pike (legal counsel and legal product lead at Anthropic) and Maggie Russo (Applied AI), drew 20,000 registrants, 51 submitted questions and 2,470 upvotes (session figures, our count of the live Q&A, May 2026; recording on demand at claude.com).
- Four privilege and security questions took 1,050 upvotes, 42% of the total; the top question had 372 (our count, May 2026).
- On 12 May 2026 Anthropic released 20+ new MCP connectors and 12 legal plugins for Claude Cowork, open source (Anthropic release, 12 May 2026; Artificial Lawyer, 16 May 2026).
- The public claude-for-legal repository held 151 skill files at commit 4a6c651 (23 July 2026); "EU" appears in 28 of them, "UK" in 16 and "Delaware" in 14 (HAQQ count, 1 Oct 2026).
- In United States v. Heppner (S.D.N.Y., opinion of 17 Feb 2026), documents a defendant made with a consumer AI tool on his own initiative were held protected by neither privilege nor work product (Proskauer, 2026).
- Charlotin's AI Hallucination Cases database lists 2,202 court and tribunal decisions involving AI-fabricated material; Mark Pike cited "over 2,000" filings during the webinar (database update of 19 Oct 2026).
- In our 23 September 2026 run, Claude Opus 5.5 matched Claude Fable 5.1 on legal drafting at $0.3647 per answer against $0.7249 (HAQQ run, OpenRouter list prices).
What did Anthropic's Claude for Legal Teams webinar cover?
Mark Pike framed the session around four pillars of how Claude does legal work, and almost every audience question maps onto one of them.
- Live data through MCP. Model Context Protocol connectors let Claude read the systems a team already uses (document management, CLM, Drive, Outlook, Slack) live, instead of from uploaded snapshots. Mark called MCP the "USB-C of AI".
- Legal skills. A skill is a markdown file that writes down a workflow the team runs every week: NDA triage, redlining, privilege logs, matter intake.
- Document comprehension. Tracking defined terms, cross-references, exceptions and schedules the way a lawyer reads an agreement.
- Context across apps. A redline in Word becomes a slide, an email and a calendar invite without restating the matter. Mark's phrase: "The work moves with you."
The opening story set the tone. A paralegal named Andrew, on a four-person pro bono team defending an elder-abuse case, built a tool on Anthropic's API that suggested cross-examination angles at counsel's table during trial. The team won a large jury verdict. The architectural bet behind the session is that one general model, given the same tools lawyers use, does not need a separate legal model: the tailoring lives in skills, plugins, connectors and matter context.
What did 20,000 legal professionals ask about Claude?
| Theme | Questions | Upvotes | Share |
|---|---|---|---|
| Security and privilege | 4 | 1,050 | 42% |
| Skills and plugins | 13 | 511 | 21% |
| Integrations | 3 | 269 | 11% |
| Product surfaces and models | 7 | 213 | 9% |
| Adoption and use cases | 7 | 191 | 8% |
| Document handling | 6 | 175 | 7% |
| Logistics | 5 | 165 | 7% |
| Accuracy and verification | 4 | 161 | 7% |
Nobody asked whether AI would replace lawyers. Every question was operational: how to keep privilege, how to verify, how to connect the systems we already use, how to roll out. Privilege drew 6.5 times the votes of hallucinations, which tells you the room had already decided to use AI and wanted the policy to do it safely.
The top questions, Anthropic's answers and ours
Ranked by upvotes. Where the session ran out of time, the middle column says so.
| Question (upvotes) | What Anthropic said | Our answer |
|---|---|---|
| How is your team dealing with attorney-client privilege when using Claude? (372) | Not your lawyer. Heppner involved a non-lawyer on a consumer plan acting on his own. On Team or Enterprise, configured correctly, Mark was "quite confident" privilege can be kept. ABA rules require lawyers to keep up with technology. | Right as far as it goes. Privilege has to hold at four layers (contract, infrastructure, deployment, matter); the vendor covers the first two and the firm owns the rest. See below. |
| How do we explain to clients that their data is safely used? (200) | Pointed to Anthropic's trust centre. | The firm has to make its own representation: a completed vendor security questionnaire, an attestation for sensitive matters and a one-page client explainer. |
| What security considerations apply when linking an inbox or using MCP? (150) | Connectors inherit your source permissions; each action can be set to always allow, needs approval or blocked. | A connector is standing access until revoked. Govern it like an OAuth grant: log retrievals, audit, revoke, and keep personal accounts off firm tools. Set send, delete and modify to needs approval. |
| With 70+ page documents, how do you avoid context rot? (101) | Cowork's agentic harness plans, writes notes to disk and works in stages; ask Claude to cite sources. | Correct. When a skill splits a long document into sections, add a final synthesis step that reconciles the per-section findings. |
| Our work lives in a DMS or case management system. Can Claude read it? (84) | MCP connectors for iManage, NetDocuments, CLMs, Drive and Microsoft 365. | Depends on your vendor. Check whether yours ships an installable connector or only an API someone must wire. |
| I want to build agents on a Team account. How is that different from a skill? (79) | Not answered live. | A skill is the procedure (how to triage an NDA). An agent is the role that runs skills and connectors end to end, often on a schedule. |
| How does Anthropic reduce hallucinations? (70) | Ground Claude in case law through connectors, ask for sources, keep a human in the loop. | Necessary, not enough. A real citation can be summarised wrongly. Check that quoted language appears in the cited source. |
| Cowork or Claude Code? And which model? (68) | Code is to Cowork what a terminal is to an IDE: same engine, same skills. | Lawyers should use Cowork. Use a small, fast model for triage steps and an Opus-class model for drafting and analysis. |
| What methodology should we use to verify outputs before senior review? (60) | Cite sources; review. | Require a page or paragraph reference per claim, check every quotation against the source, and have the reviewer check the flags first. |
| Can Claude track which version reflects which party's changes? (57) | Redlining was demoed; multi-party comparison was not. | Yes, if you give it every version. Ask for a clause-level changelog by version and party, and save the prompt as a skill. |
| Can we get a glossary? (50) | Not given. | Ask the assistant to define any term in your chat. |
| How do I start from zero? I have only used ChatGPT. (49) | Not answered live. | See the four-hour setup below. |
| How accurate is Claude on old scanned deeds and JPEGs? (44) | Not answered live. | Digital PDFs work like Word files. Degraded scans depend on the OCR in front of the model; treat any transcription as a draft and verify against the original. |
| We are a small firm on Team. Do we get Enterprise features? (20) | Not answered live. | Most features, yes. Enterprise adds SCIM, audit logs, custom retention and org-wide controls; buy it when you need information barriers. |
Is Claude safe for privileged legal work?
Privilege is a property of a workflow, and it has to hold at four layers at once.
- Contract. The vendor must not train on your inputs, in writing. On Anthropic's Team, Enterprise and API plans that is the default; on consumer plans (Free, Pro, Max) training is a setting the user controls under the August 2025 terms.
- Infrastructure. Audited security controls, retention options for sensitive matters and hosting in a region your data-residency duties allow.
- Deployment. Who in your firm can install which connectors and read whose chats. The common break we see is a sensitive memo pasted into a workspace a non-lawyer administrator can read. Only the firm can configure this.
- Matter. Privilege belongs to this client, this lawyer, this purpose. General-purpose tools organise work in projects and folders, which a firm has to map to matters itself.
Heppner is often cited as a warning against AI in legal work. The holding is narrower: a criminal defendant used a public AI tool on his own initiative, before his lawyers directed anything, and the court found neither privilege nor work product. The safer pattern is a commercial plan, use at counsel's direction, and connector actions set to needs approval. More on the ruling in our privilege analysis.
What is the difference between a skill, a plugin, a connector and Cowork?
- Skill. A markdown file describing one procedure, written as you would brief a new associate. The one Mark showed was about 200 lines.
- Plugin. A bundle of skills. Anthropic's legal plugins start with a setup interview that records your playbook, escalation chain and house style.
- Connector (MCP server). The live link to a system: iManage, NetDocuments, Outlook, Drive, Slack, a CLM.
- Cowork. Anthropic's desktop agent that opens folders, writes files and runs tasks in parallel inside a sandboxed environment. Anthropic merged Cowork and chat into one Claude on 16 September 2026.
- Claude Code. The same engine in a terminal, for engineers. Skills and connectors written for Code work in Cowork.
- Projects and artifacts. Shared workspaces with common instructions and files; and live outputs (a document, dashboard or small app) generated inside a conversation.
Mark's advice on the legal plugin was the most useful line of the session: do not use it off the rack. Remix it with your own risk matrices, playbooks and fallback language. Claude will draft the first version if you give it 15 of your past NDA reviews and ask for the playbook; expect to edit it. Anthropic ships a skill-creator skill for exactly this. Our count of the public plugins explains why the remix matters: they are written around US, UK and EU practice, so a firm working under Gulf law customises more than one in Delaware. We set that out in Claude vs legal AI for lawyers.
Can Claude handle 70-page legal documents?
In a single chat turn, accuracy on detail from the middle of a long input degrades even when it fits the context window. Cowork works around that the way an associate would: it opens the document, splits the work into stages, writes intermediate notes to disk and reads back the parts it needs. Two practices make it reliable. Ask for a page or paragraph reference for each finding, which gives you a check and keeps the model grounded. And when a skill analyses sections separately, add a final step that reads every section's output and reconciles conflicts, or you get fifteen partial answers instead of one.
How do you verify Claude's legal output?
Mark pointed to more than 2,000 court filings with hallucinated content; Charlotin's database stood at 2,202 decisions on 19 October 2026, 872 of them involving lawyers. Asking for sources is the start. The failure that reaches court is usually a real citation summarised wrongly, which a busy reviewer pastes without opening. A workable method:
- Every claim carries the paragraph or page it came from.
- Every quotation is checked against the cited document and flagged if it does not appear verbatim.
- Every legal proposition without a source is marked model-only and escalated.
- The reviewer checks the flags first, then reads the rest.
- Errors go into a running log that feeds back into the skill as things to avoid.
Review still happens. What changes is its length: on a well-written skill, Mark's team reported redlines in about 20 minutes.
Which Claude plan and model should a law firm use?
Use a commercial plan for client work. On Anthropic's pricing page, project sharing and no training by default sit on Team, while SCIM, audit logs and custom retention sit on Enterprise. A small firm with one office rarely needs Enterprise; a firm with practice groups that need information barriers does.
On models, the rule of thumb from the webinar Q&A still holds: a small, fast model for classification and triage steps inside skills, a frontier model for drafting and analysis. When we tested three Claude models on ten lawyer-grade prompts in September 2026, we could not separate them on quality, and Opus 5.5 cost half as much per answer as Fable 5.1.
How long does it take to set Claude up for legal work?
- Hour 1: account and install. A Team or Enterprise plan, Cowork installed, the legal plugin added from the plugin directory.
- Hour 2: one connector. Pick the system where most of your work lives. In its permission grid, set every send, delete and modify action to needs approval; read-only actions can stay on always allow.
- Hour 3: remix one skill. Open the plugin skill closest to work you do every week and edit it: your fallback positions, the clauses you fight over, the output format your team wants.
- Hour 4: run it on real work. Read the output, fix the skill, run it again until the result is something you would hand a partner.
After that the second skill is cheaper than the first. Mark's examples show where to aim: a regulatory colleague's two-hour daily news synthesis turned into a scheduled 8 a.m. brief the whole team reads; privacy impact assessments compressed to 20 to 30 minutes; the Friday team newsletter drafted from Slack and ticket activity. For a firm, someone has to own the internal plugin library, or the skills stay in one partner's laptop.
What has changed since the webinar?
- Claude for Legal shipped (12 May 2026): 20+ MCP connectors and 12 open-source legal plugins for Cowork, including an M&A plugin with tabular review.
- Cowork and chat merged (16 September 2026). The surface question from the Q&A is mostly settled: there is one Claude, with Code for engineers.
- The plugins are public and countable. 151 skill files at commit 4a6c651, written around US, UK and EU law.
- Models moved on. Claude Opus 5.5 and Fable 5.1 replaced the models discussed in the session; in our September run they were close on legal drafting quality.
- The hallucination record kept growing, from the "over 2,000" Mark cited to 2,202 decisions by 19 October 2026.
The webinar's gap is the same one five months later: what the platform does in a demo against what a twelve-lawyer firm with one IT person can wire up. If you have the hours, plain Claude with your own skills is a serious setup. If you do not, the work you are buying from a legal AI vendor is the connectors, the matter layer and the verification, already built. Start with one workflow that hurts this week, write the skill and give it to one colleague.
Sources
- Anthropic: Claude for Legal Teams webinar (on demand)
- Artificial Lawyer: Claude for Legal moves centre stage (16 May 2026)
- anthropics/claude-for-legal on GitHub
- Claude pricing and plans
- Proskauer: SDNY on privilege and public AI tools (Heppner)
- Damien Charlotin: AI Hallucination Cases database
- Claude vs legal AI: which one a lawyer needs
- Claude Opus 5.5 for legal work: our benchmark
- The Claude for Word plugin and legal AI
- AI hallucination cases: the 2,202-case tracker
- What Claude replaces in legal tech, and what it does not
FAQ
What did Anthropic's Claude for Legal Teams webinar cover?
Mark Pike and Maggie Russo of Anthropic showed Claude Cowork on legal work (NDA triage, redlining in Word, meeting briefs) and framed it around four pillars: live data through MCP connectors, legal skills, document comprehension and context across apps. The session drew 20,000 registrants and 51 audience questions with 2,470 upvotes.
Does using Claude waive attorney-client privilege?
Not by itself. In United States v. Heppner (S.D.N.Y., February 2026), documents a defendant made with a consumer AI tool on his own initiative were held unprotected. Anthropic's position is that on Team or Enterprise, configured correctly, privilege can be maintained. Use a commercial plan, use the tool at counsel's direction, control who in the firm can see what, and take advice for your jurisdiction.
What is the difference between a Claude skill and a plugin?
A skill is a markdown file that describes one procedure, such as how to triage an NDA. A plugin is a bundle of skills. Anthropic released 12 open-source legal plugins for Cowork on 12 May 2026, and each starts with a setup interview to capture the firm's playbook and house style.
Can Claude handle long legal documents?
Yes, with the right setup. For documents of 50 pages or more, use Cowork, which works in stages and writes intermediate notes to disk instead of reading everything in one turn. Ask for a page or paragraph reference for each finding, and add a final step that reconciles section-by-section results.
How should a law firm verify Claude's legal output?
Require a source reference for every claim, check every quotation against the cited document, mark unsourced legal propositions as model-only, and have the reviewer check those flags first. Charlotin's database listed 2,202 court decisions involving AI-fabricated material on 19 October 2026, so citation checks are not optional.
Does a small law firm need Claude Enterprise?
Usually not. On Anthropic's pricing page, Team includes project sharing and no training on your data by default; Enterprise adds SCIM, audit logs, custom retention and org-wide controls. Enterprise makes sense when practice groups need information barriers or central administration.