Claude for Lawyers: Is It Legal to Use Claude in 2026?

Key Takeaways

‍
It is legal for lawyers to use Claude in every United States jurisdiction, and the rules that govern how lawyers use Claude are the ones they already know. ABA Formal Opinion 512 applies competence (Model Rule 1.1), confidentiality (1.6), candor toward the tribunal (3.3), and supervision (5.3) to generative AI rather than creating new standards.

The account tier decides the confidentiality answer: consumer Claude plans can be used to train Anthropic’s models, commercial plans are not trained on by default, and zero data retention is a separate agreement covering the Anthropic API and Claude Code for Enterprise rather than Enterprise chat. Three federal decisions in 2026 split on whether AI-assisted material keeps privilege or work-product protection, and each turned on how the lawyer used the tool.

That said, legality at the jurisdiction level does not override an individual judge’s standing order. A growing number of federal and state judges now require lawyers to disclose or limit generative-AI use in court filings, and a standing order’s restriction controls regardless of what the jurisdiction’s own ethics rules allow. The duties fall the same way on in-house counsel and on commercial lawyers in law firms, though the controls each one can switch on depend on who owns the account.

[cta-1]

What Is the Ethical Framework for Using Claude in Legal Practice?

ABA Formal Opinion 512 is explicit that generative AI does not need a new ethics regime. It maps four existing duties onto AI use and leaves enforcement to the rule that already covers the conduct.

Why these four duties apply without a new rule

The opinion treats Claude the way earlier opinions treated legal research databases and document assembly software: a tool a lawyer must understand well enough to use competently, not a separate category of risk requiring its own rule. A lawyer who would not submit an unverified cite from a junior associate should not submit one from Claude either, and the rule that catches both is 3.3, not a Claude-specific standard.

Where the guidance comes from

A handful of state bars have issued their own opinions layering onto ABA 512 rather than contradicting it, and state bar AI guidance tracks the jurisdictions that have weighed in. None has gone further than requiring disclosure or client consent in specific circumstances; none has banned the category.

Related Article: Perplexity for Lawyers: What It Gets Right (and Where It Stops)

The Four Ethical Obligations Lawyers Must Meet When Using Claude

Four obligations recur across every state bar opinion on generative AI, and each maps to a specific Model Rule.

Competence: Model Rule 1.1

A lawyer using Claude must understand enough about the tool to supervise its output responsibly.

• Know the plan’s training and retention defaults before a matter touches the tool, not after.
• Understand what Claude cannot do (it has no access to a matter’s privileged file unless a human puts it there) as well as what it can.

Confidentiality: Model Rule 1.6

The confidentiality analysis turns on the account, not on the brand. A free or Pro account can train Anthropic’s models on a conversation unless the model improvement setting is off; a Claude for Work or Enterprise account is not trained on by default. United States v. Heppner is the case most often cited against using Claude at all, but it is narrower than that reading. The court denied privilege to 31 documents a criminal defendant generated in a consumer account on his own initiative, without a lawyer directing the use or reviewing the output, and the opinion did not address a lawyer using a commercial account under firm policy.

• Match the account tier to the matter’s sensitivity before any client fact enters the tool.
• Treat a consumer account as unable to support a reasonable expectation of confidentiality until the model improvement setting is confirmed off.

Candor Toward the Tribunal: Model Rule 3.3

Model Rule 3.3 requires candor regardless of which tool produced the draft, including verifying every citation an AI tool generates before it reaches a court.

• Verify every case citation Claude produces against a primary source before filing, every time, with no exception for routine motions.
• Keep a record of the verification step for any filing that later comes into question.

Supervision: Model Rule 5.3

Model Rule 5.3 extends a lawyer’s supervisory duty over nonlawyer assistance to technology, which means firm policy, not individual discretion, decides which Claude plan, connectors, and MCP servers are approved for client work.

• Put the approved plan and configuration in writing, not in an email thread, so new hires inherit the policy rather than reinventing it.
• Review connector and Skill permissions on a cadence, since a connector approved for one matter can expose a different matter’s data if left on by default.

Never enter confidential client data without enterprise controls

The obligation, stated plainly: do not put a client’s confidential information into any Claude account that has not been configured for it, and do not treat “the firm has a subscription” as proof that a given plan is configured correctly. Obtaining informed client consent before using AI on a matter, per ABA Formal Opinion 512, is a separate step from picking the right plan.

How Can Lawyers Use Claude Legally?

A lawyer uses Claude legally by matching the account tier to the matter, directing the use with specific, well-scoped prompts rather than delegating it open-endedly, and verifying every output before it reaches a client or a court.

• Pick a commercial plan (Claude for Work or Claude Enterprise) for anything touching client information; reserve consumer accounts for non-client research.
• Direct the use: ask Claude to draft or analyze a specific document rather than handing it open-ended access to a matter.
• Verify outputs against a primary source before relying on them in advice, a filing, or a client deliverable.
• Disclose AI use to the client where the engagement letter, the jurisdiction’s guidance, or a standing order requires it.
• Document the account tier and configuration used on a matter, so a later privilege or work-product dispute has an answer rather than a guess.
• Check the specific matter’s standing order before assuming the jurisdiction’s general rule is the only constraint.

The same discipline applies whether the task is drafting a contract from scratch or reviewing paper someone else wrote.

Claude for Legal Document Drafting and Review

Claude drafts and reviews legal documents through Claude Projects (persistent context for a matter), Claude Skills (reusable instructions for a recurring task type), and, as of this writing, a Microsoft Word add-in that lets a lawyer work with Claude without leaving the drafting environment. None of that changes the governance analysis above: the output still needs the same verification and the same account-tier discipline, whether the task is drafting or reviewing an existing file.

Firms building reusable Claude Skills for a recurring task type, such as a standard intake memo, should still run every output through the same verification step as a one-off draft. That discipline covers document analysis as much as it covers original drafting, and it applies equally to work done inside Claude for lawyers generally.

What Claude does not currently do, compared with a contract-specific tool, is ground its output in a firm’s own playbooks, benchmark a clause against a market-standard data set, or carry preference learning across matters. A general-purpose assistant and a contract-specialist tool therefore serve different parts of the same workflow.

• General legal research and first-draft memoranda, with citations verified against a primary source before use.
• Document summarization and issue-spotting on a single file a lawyer uploads directly.
• Correspondence and filing drafts, always reviewed under the Rule 3.3 candor duty above.
• Cross-matter pattern work inside a Project, limited to what a human has deliberately scoped into it.

Which Claude Plan Can Lawyers Use with Client Data?

Zero data retention is a separate agreement rather than a setting inside Enterprise chat. It currently covers eligible Anthropic APIs and Claude Code for Enterprise plans. A flagged conversation, on any plan, is currently retained for up to two years regardless of the plan’s ordinary retention period, and a Business Associate Agreement excludes several products still in beta along with Cowork in the cloud, which matters for any firm handling protected health information under a medical malpractice or health-industry engagement.

• Zero data retention does not currently extend to Enterprise chat; treat it as API and Claude Code for Enterprise only until Anthropic changes that.
• The BAA question and the training question are separate: a plan can be not-trained-on-by-default and still lack BAA coverage.
• GDPR runs on a different trigger than HIPAA: it attaches to EU personal data regardless of plan, so check it independently rather than assuming the BAA analysis covers it.

The Word add-in exception

Working inside Microsoft Word carries one plan-level exception worth flagging on its own: Claude for Word does not currently inherit an organization’s custom data retention settings, and activity through the add-in is not included in Enterprise audit logs. A firm that has configured a strict retention period for Claude Enterprise chat should not assume that configuration extends to the Word add-in; it does not, and a firm relying on audit logs to document AI use on a matter will not find Word add-in activity in them.

• The Word add-in operates outside whatever retention and audit-log configuration the rest of an Enterprise account has in place.
• A firm using the Word add-in on a BAA-covered Enterprise account should not assume health-information-safe handling without checking the add-in specifically, since the Word add-in is itself on the BAA exclusion list, which Anthropic lists as Claude in Office.

[cta-2]

What Courts Have Held About AI and Privilege Since Heppner

Three 2026 federal district court decisions now shape how attorney-client privilege and work-product protection apply to AI use in litigation, and the court rulings do not all point the same way. United States v. Heppner (S.D.N.Y., February 17, 2026) denied privilege protection to 31 documents a criminal defendant generated in a consumer AI account on his own initiative. Warner v. Gilbarco (E.D. Mich., February 10, 2026) and Morgan v. V2X (D. Colo., March 30, 2026) each held that AI-assisted litigation preparation is protected attorney work product, and that routing material through an AI tool does not by itself waive privilege.

United States v. Heppner (S.D.N.Y., February 2026)

Judge Jed S. Rakoff issued a bench ruling on February 10 and a written opinion on February 17, 2026 WL 436479, No. 25-cr-00503-JSR. Privilege failed on three grounds: the tool is not a lawyer, the consumer privacy policy defeated any reasonable expectation of confidentiality, and the defendant was not seeking legal advice when he used it. Work product failed separately, because the defendant acted on his own initiative rather than at counsel’s direction.

• The opinion says nothing about a lawyer directing AI use under a commercial account; it noted the result “may have been different” had counsel directed it.
• Read narrowly, Heppner is a case about unsupervised, self-directed consumer use, not a general rule against lawyers using AI tools.

Warner v. Gilbarco (E.D. Mich., February 2026)

Decided a week before Heppner’s written opinion and reaching the opposite result, 2026 WL 373043, No. 2:24-CV-12333. Work product is waived by disclosure to an adversary, the court held, not by disclosure to a third-party tool operating under a duty of confidentiality; it described AI chatbots as “tools, not persons” for that purpose.

• The holding turns on who the material was disclosed to, not on whether AI was involved in producing it.
• It directly contradicts a reading of Heppner as a categorical bar on AI-assisted work product.

Morgan v. V2X (D. Colo., March 2026)

Decided by Magistrate Judge Maritza Dominguez Braswell, No. 25-cv-01991. The court held that Rule 26(b)(3) work product does not depend on attorney involvement in creating the material, citing Carpenter v. United States, 585 U.S. 296 (2018), for the proposition that routing information through a third-party system does not forfeit all privacy in it. The identity of the AI tool used was not itself protected as work product.

The order also approved AI-specific protective-order language: no party may input confidential information into an AI platform unless the provider is contractually barred from storing or training on the inputs and from disclosing them to third parties except where essential to service delivery, the provider must allow deletion on request, and the party must keep written documentation of those protections.

• The clause ties protection to the provider’s contractual commitments, not to the fact of using AI at all.
• A firm that wants this protection on a matter needs the contractual terms in place before the AI tool is used, not after a discovery dispute raises the question.

None of these three decisions is appellate, so the attorney-client privilege analysis for AI-assisted work is still developing and courts in other circuits may reach different conclusions.

What separates the protected cases from the denied one

Three factors separate the two protected cases from the denied one: counsel direction, a contractual no-training commitment from the provider, and documentation of both.

• Counsel direction and a contractual no-training commitment were present in both Warner v. Gilbarco and Morgan v. V2X.
• Both were absent in Heppner, where the defendant acted on his own initiative in a consumer account with no provider commitment limiting further use of the material.

How to Configure Claude for Legal Work

Configuring Claude for legal work means settling five things before any client matter reaches the tool: the model improvement setting, the account tier, the retention period, the Business Associate Agreement, and the connectors an administrator leaves enabled. Anthropic documents each control, and the defaults are not the configuration a lawyer wants.

1. Turn off the model improvement setting

• On any consumer plan, confirm this setting is off before a single client fact touches the account; it is on by default for free and Pro accounts.
• Incognito chats are currently excluded from training regardless of the setting, but they are not a substitute for a commercial plan on an ongoing matter.

2. Move client work onto a commercial plan

• Claude for Work and Claude Enterprise are not trained on by default, which removes the training question entirely rather than relying on a setting a user could toggle.
• A firm-wide policy, not an individual lawyer’s preference, should decide which plan is approved for which matter type.

3. Set a retention period, and know its floor

Anthropic documents the Enterprise retention floor directly: a 30-day minimum, with indefinite retention if no period is set and Project-level settings overriding chat-level ones.

• Claude Enterprise retention is configurable with a 30-day minimum; left unset, it defaults to indefinite.
• Project-level retention settings currently take precedence over chat-level settings, so a matter’s Project needs its own check, not the account default alone.

4. Activate HIPAA settings and accept the BAA if you hold protected health information

• The Business Associate Agreement is not automatic on Claude Enterprise; a Primary Owner currently has to take a separate action to accept it.
• The BAA currently excludes several products still in beta along with Cowork in the cloud, Claude Console, and Enterprise Search, so a health-adjacent matter using one of those surfaces is not covered even on a BAA account.

5. Decide which connectors, MCP servers, and Skills stay on

• A connector that reaches a third-party system is not covered by the BAA for data it sends to that third party, regardless of the Claude plan underneath it.
• Review the connector list on a fixed schedule rather than only at onboarding, since a connector approved for one matter can expose a different matter if nobody turns it off.

What to Do If Client Data Has Already Gone into Claude

If confidential client data has already gone into a consumer Claude account, remediation runs in four steps: turn off the model improvement setting, delete the affected conversations, document what was entered and when, and assess the notification duty under Model Rule 1.4. Anthropic states that a deleted chat is not used to train future models, though data already inside a training run that is underway remains in it.

1. Switch off model improvement

Do this first, before anything else, so no further exposure accumulates while the rest of the remediation happens.

2. Delete the conversations

Deletion stops future training use, but a flagged conversation can currently be retained for up to two years regardless of the plan’s ordinary retention period, which affects what deletion accomplishes.

3. Record what was disclosed

• Note what was entered, when, into which account, and whether the conversation was flagged, since the retention consequences differ for a flagged conversation.
• Keep this record outside the Claude account itself, in the matter file, where it survives account changes.

4. Work the Model Rule 1.4 and 1.6 analysis

• Determine whether the disclosure went to a vendor operating under a duty of confidentiality or to an open consumer platform; that distinction drives both the Rule 1.6 analysis and any state breach-notification statute that might apply.
• This is the analysis a lawyer runs, not a conclusion that applies the same way to every matter; the right answer depends on the account tier, the data involved, and the jurisdiction’s breach-notification rule.

Spellbook vs. Claude: Which Tool Fits Transactional Legal Work

Spellbook and Claude solve adjacent but different problems. Claude is a general-purpose assistant that currently reaches Word through its own add-in, generally available to Pro, Max, Team, and Enterprise plans, and produces tracked-change redlines reviewable in Word’s native review pane. It does not ground those redlines in a firm’s own playbooks, does not benchmark a clause against a market-standard data set the way the Compare to Market feature does, and does not carry preference learning across matters. The Word add-in also sits outside the configuration a firm may have set up for the rest of its Claude Enterprise account: it does not currently inherit custom retention settings, and its activity does not appear in Enterprise audit logs, which matters for any firm relying on those logs to document AI use on a matter.

Spellbook is a commercial-contract specialist, not a general-purpose research or litigation tool, and it does not try to be one. The choice on a given matter is whether the work needs a general assistant’s flexibility or a specialist tool’s playbook grounding and benchmark data; many firms use both for different parts of the same matter.

When each tool fits

• Choose Claude when the work is general research, drafting from scratch, or analysis outside a recurring contract workflow.
• Choose Spellbook when the task is reviewing paper against a firm’s own standards, benchmarking a clause, or carrying preference learning across matters.

Try Spellbook

Lawyers weighing whether a general-purpose assistant or a contract-specialist tool fits a given matter can try Spellbook free to compare the playbook-grounded review directly against their current workflow.

Frequently Asked Questions About Lawyers Using Claude

Can lawyers use Claude without disclosing it to clients?

Disclosure obligations currently depend on the jurisdiction, the engagement letter, and whether the matter is in litigation under a standing order requiring AI disclosure. ABA Formal Opinion 512 recommends disclosure where the use is material to the representation, but it does not impose a universal disclosure rule. The disclosure requirements that attach to AI-assisted work product differ by jurisdiction.

Does using Claude waive attorney-client privilege?

Not automatically. Heppner denied privilege to unsupervised, self-directed consumer use, but two later 2026 decisions, Warner v. Gilbarco and Morgan v. V2X, protected AI-assisted litigation work as work product where counsel directed the use and the provider’s contract limited further disclosure. None of the three is appellate, and the outcome currently turns on how the tool was used, not on whether it was used at all.

Is Claude HIPAA or GDPR compliant for legal work?

HIPAA coverage on Claude Enterprise is not automatic. A Primary Owner currently has to separately activate HIPAA settings and accept the Business Associate Agreement, and that BAA excludes several products still in beta along with Cowork in the cloud. A firm handling protected health information should confirm the BAA is active on the specific product surface being used, not assume Enterprise coverage extends everywhere. GDPR works differently: it attaches to any processing of EU personal data regardless of plan or BAA status, so a matter touching an EU data subject needs a data processing agreement and a lawful basis for the transfer, not a HIPAA-style activation step.

Is a general-purpose AI model enough for contract review?

It depends on what the review needs to catch. A general-purpose model can flag obvious issues and draft a first pass, but it does not currently ground its analysis in a firm’s own playbooks or benchmark a clause against a market-standard data set the way a contract-specialist tool does. For a one-off review with no firm standard to apply, a general model may be enough; for recurring contract types where consistency against a playbook matters, a specialist tool currently adds something a general model does not.

Does the Claude Team Plan have zero data retention?

No. Zero data retention is currently an approved agreement covering eligible Anthropic APIs and Claude Code for Enterprise plans, not a feature of the Team plan or of Enterprise chat. Claude for Work (Team) retains chats in the product with no separate zero-retention option.

Is Claude confidential for lawyers?

Confidentiality currently depends on the plan and the contract terms, not on the Claude brand name generally. A commercial plan (Claude for Work or Claude Enterprise) is not trained on by default, which covers the training question; the retention and audit-log questions still depend on the specific product surface, including the Word add-in exception covered above, and on counsel directing the use rather than delegating it unsupervised.

What Claude plan should a small law firm choose?

Small law firms handling confidential client information should choose a commercial plan, such as Claude for Work (Team) or Claude Enterprise, rather than a consumer subscription. Both commercial plans exclude customer data from model training by default. Claude Enterprise adds configurable retention and additional administrative controls, making it better suited for firms with stricter confidentiality and compliance requirements.

How do lawyers use Claude for legal research?

Lawyers use Claude for legal research by summarizing statutes, analyzing legal documents, identifying relevant issues, and drafting initial research memoranda. Claude can also help compare legal arguments and organize complex information. However, lawyers must independently verify every citation, legal claim, and interpretation against authoritative sources such as Westlaw or LexisNexis before relying on the results.