AI Agents for Lawyers

Regulation & Ethics

How to Write a Law Firm AI Policy (With a Template Outline)

A law firm AI policy lists approved tools by tier, classifies data, makes verification personal, covers disclosure, billing and training, and limits agents.

By Dr. Niklas SchmidtUpdated 13 November 20269 min read
On this page
  1. Why "no ChatGPT" is not a policy
  2. What a working policy covers
  3. The outline, section by section, with sample language
  4. Data classes: the traffic-light table
  5. Verification: the section the sanctions cases wrote for you
  6. Agent permissions: the section most policies still lack
  7. Disclosure, billing and the Article 4 training record
  8. Rolling it out without killing adoption: what to do this week

A law firm AI policy names the tools and account tiers staff may use, classifies what data may go into each, makes verifying every citation a personal duty of the signing lawyer, sets rules on client disclosure and billing, keeps the training record Article 4 of the EU AI Act expects, and tells people how to report a mistake without penalty for candour. The outline and sample clauses follow.

Why "no ChatGPT" is not a policy

Clio's Legal Trends Report 2025 found that 79% of legal professionals use AI while 53% say their firm has no AI policy or do not know whether one exists; legal-specific tool use fell from 58% to 40% as lawyers defaulted to general chatbots. Thomson Reuters' Future of Professionals survey (June 2026, 1,816 respondents) put unsanctioned "shadow AI" use at 34%, and found that 66% of professionals at organisations with a named AI strategy say AI meets expectations, against 22% where there is none. The North Carolina Bar's January 2026 note, Beyond the Ban, draws the conclusion: prohibitions move AI use onto personal phones and personal accounts, where the firm has no visibility, no logs and no contract.

Regulators now address management rather than individuals. In Ayinde v Haringey [2025] EWHC 1383 (Admin) the Divisional Court spoke to "heads of chambers and managing partners" about their responsibility for AI literacy. The SRA's Warning Notice of 17 August 2026, issued after 42 AI-misuse reports in twelve months, expects a solicitor to understand a tool "well enough to review its output critically". And while no major US professional-liability carrier had filed an explicit AI exclusion as of 5 May 2026, application forms now probe governance; the carrier tracker at legalaigovernance.com notes that AmTrust's form asks whether the firm lets AI "draft documents". A written policy is the answer to that question.

What a working policy covers

SectionWhat it decidesWhat drives it (as of September 2026)
1. Scope and approved toolsWhich products, on which account tier, for which tasksConsumer tiers train on or retain inputs unless opted out; business, enterprise and API tiers do not by default
2. Data classesWhat may go into which toolUK v SSHD [2026] UKUT 81 (IAC): client letters pasted into public AI tools are "in the public domain"
3. VerificationWho checks what before anything leaves the firmFletcher v. Experian (5th Cir. 2026): $2,500 for 16 fabricated quotations from vLex and CoCounsel
4. Client disclosureWhen clients are told and may objectCCBE Guide (2 Oct 2025); ÖRAK practice rules; 78% of clients want disclosure, 35% of lawyers rarely give it (Clio 2025)
5. BillingWhat may be charged for AI-assisted workABA Formal Opinion 512; Texas Opinion 705; California guidance (May 2026)
6. Training and recordsWho is trained, on what, whenEU AI Act Article 4; Commission FAQ updated 27 July 2026
7. Incident reportingHow an error is reported and correctedCouvrette v. Wisnovsky (D. Or. 2025): $110,204.38, aggravated by quiet refiling
8. Agent permissionsWhat autonomous tools may do without a humanCalifornia Practical Guidance on agentic AI (14 May 2026); the Zenity "PleaseFix" hijack

The outline, section by section, with sample language

1. Scope and approved tools. State who is covered (partners, fee-earners, business services, contractors, secondees) and list tools by product name and account tier. ChatGPT Plus and ChatGPT Enterprise look identical on screen and are different products for confidentiality purposes; the same applies to Claude and to Copilot. Give the list an owner and a review date, and route new tools through vendor due diligence rather than a partner's credit card.

1.2 Approved tools. Only tools listed in Schedule A may be used for firm or client work. Schedule A records, for each tool, the account tier, the data class permitted under section 2, the contractual no-training and retention terms, and the partner responsible. Personal accounts, free tiers and browser extensions not listed in Schedule A may not be used with any firm or client information.

2. Data classification. Three classes are enough; the traffic-light table below supplies them. Add that "anonymised" client facts remain client information if a counterparty could recognise the matter.

3. Verification and output review. The signing lawyer verifies personally, verification has three checks, and nothing is filed without reading the judge's standing order.

3.1 Every authority, quotation, statutory reference and factual assertion produced with AI assistance must be verified by the lawyer responsible for the document against the primary source before the document is filed, sent or relied upon. Verification means confirming that the source exists, that it says what the document says it says, and that it remains good law. Verification may not be delegated to an AI tool.
3.2 Before filing in any court or tribunal, the responsible lawyer must check the standing orders of the court and of the assigned judge on AI disclosure or certification and comply with them.

4. Client disclosure. Adopt a default engagement-letter paragraph, and say when a matter-specific conversation is required: whenever client data will enter a tool the client has not approved, and whenever AI output will go to a court or counterparty. The disclosure rules differ by jurisdiction, but the CCBE's position that clients should be informed and allowed to object is a safe floor across Europe.

5. Billing. One paragraph settles it: bill actual time, including verification time; treat general AI subscriptions as overhead; never mark up a pass-through AI cost without informed written consent; and record AI-assisted work in narratives so the client sees the method.

6. Training and records. Name the modules, the cadence and where the attendance register lives. The Article 4 duty is met with a record, not a certificate.

7. Incident reporting. The section that decides whether a $2,500 mistake becomes a $110,000 one.

7.1 Anyone who discovers that AI-assisted material containing an error has been filed, sent or relied upon must report it to the supervising partner and the risk team the same day. Prompt self-reporting is treated as compliance with this policy, not a breach of it. Corrected documents may not be filed or sent without the risk team's approval and, where required, candid notice to the court or the counterparty.

8. Agents and automation. A permissions schedule, logging and a human-controlled stop; see below.

Data classes: the traffic-light table

ClassExamplesWhere it may goAnchor
Red: privileged or identifiable client materialClient correspondence, draft advice, deal documents, personal data of clients or opponentsOnly Schedule A tools with contractual no-training, defined retention and, for EU firms, EU processingUK v SSHD; BRAK Hinweise (Dec 2024): tools that store or forward client data for training breach § 43a Abs. 2 BRAO; ÖRAK: client data stays on EU servers
Amber: firm-internal, non-clientPrecedents with names removed, know-how, pitch drafts, internal memosSchedule A tools; approved general tools on business tiersABA Opinion 512: informed consent where a self-learning tool may expose client data
Green: public and non-confidentialStatutes, published judgments, marketing copy, CLE materialAny approved tool, including business-tier general chatbotsConsumer Gemini's own privacy page asks users not to enter confidential information a human reviewer should not see

One case explains why the red class is drawn so tightly. In United States v. Heppner (S.D.N.Y., 17 Feb 2026) Judge Rakoff held that a defendant's research memos produced in consumer Claude were not privileged because "Claude is not an attorney" and the terms allowed logging and disclosure. The confidentiality checklist tracks which tiers of which tools train on inputs.

Verification: the section the sanctions cases wrote for you

Damien Charlotin's database counted 2,041 decisions involving AI hallucinations worldwide on 14 September 2026, 811 of them involving lawyers rather than litigants in person, per damiencharlotin.com. Three cases should shape the wording.

Wadsworth v. Walmart (D. Wyo., 24 Feb 2025) involved Morgan & Morgan lawyers, an internal AI platform and eight fake cases out of nine cited; the result was $5,000 in sanctions and a revoked pro hac vice admission. "Use the approved tool" is therefore not the rule; "verify regardless of tool" is. Fletcher v. Experian (5th Cir., 18 Feb 2026) traced 16 fabricated quotations to vLex and CoCounsel, paid products with retrieval behind them. Couvrette v. Wisnovsky (D. Or., 12 Dec 2025) produced the largest US sanction, $110,204.38, and the ABA Journal's report makes clear that the lawyer's decision to quietly refile amended briefs instead of confessing drove the figure.

California's May 2026 guidance supplies the standard in one sentence, quoted by Wilson Elser: "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified." Put the three-check method (existence, support, treatment) in the policy and point to the firm's citation-checking protocol for the mechanics. Add the standing-order check: over 30 of the 94 US federal district courts and, separately, 300+ individual federal judges had AI disclosure or certification requirements as of April 2026.

Agent permissions: the section most policies still lack

Chatbot-era policies assume a human reads every output. Agents act: they send, file, move files and run code. Three 2026 incidents show what the section has to prevent. In February 2026 a venture investor let Claude Cowork "organise" a desktop and an rm -rf command removed fifteen years of photos. In July 2026 xAI's Grok Build uploaded whole repositories including an unredacted .env file that users had told it to ignore; instructions to an agent are not access controls. Zenity's "PleaseFix" research (March 2026) showed zero-click hijacks of browser agents through emails and calendar invites; its advice belongs in the policy verbatim: "put hard limits around the agent that the agent cannot override".

The vendors' own designs show what good looks like: Claude for Legal's plugins carry explicit gates before filing or sending, Harvey's agents "flag moments requiring user confirmation", and the iManage MCP server is permission-bound, logged and respects ethical walls. The agent security guide explains the attack patterns; the policy only has to set the permissions.

8.1 Agent permissions. An AI agent may draft, summarise, extract, classify and propose. It may not, without a named individual approving the specific action: file with any court or authority; send any communication outside the firm; delete, move or overwrite documents; sign, pay or commit the firm; or change matter, time or conflict records.
8.2 Agents operate only under the user's own credentials and permissions, never a shared or administrator account. All agent actions are logged. A human-controlled stop must be available at all times.
8.3 Browser agents may not run on a profile logged into the DMS, email or banking unless IT has approved the site allowlist.

Disclosure, billing and the Article 4 training record

The CCBE Guide of 2 October 2025 says clients should be informed and allowed to object. ÖRAK's four practice rules, presented at its first KI-Tag on 19 March 2026, put "disclose AI use to clients" alongside EU servers, final lawyer responsibility and no training on client data. Illinois, by contrast, requires no disclosure in pleadings; one engagement-letter paragraph plus a matter-level judgment covers the spread.

On billing, ABA Formal Opinion 512 allows billing actual time and treats subscriptions as overhead; Texas Opinion 705 forbids charging for time "saved"; California adds that AI costs may not be marked up without informed written consent. The billing page has the detail; the policy needs the rules.

Training is where EU firms have a legal hook. Article 4 of the AI Act has applied since 2 February 2025; the Digital Omnibus (Regulation (EU) 2026/1744, in force 27 July 2026) reworded it to "take measures to support the development of AI literacy", and national market-surveillance authorities have supervised it since 2 August 2026. The Commission's FAQ, updated 27 July 2026, says no certificates or tests are required and recommends an internal training record. The policy should therefore name the modules (tool operation, data classes, verification, agent limits), a cadence, and a register with names and dates. A literacy curriculum built on the firm's actual tools satisfies the regulator and the SRA at once.

One further rule earns its place: meeting recording. NYC Bar Formal Opinions 2025-6 (22 Dec 2025) and 2026-2 (5 Aug 2026) require consent of all parties and set the default at not recording, so the policy should keep AI transcription off unless consent is documented.

Rolling it out without killing adoption: what to do this week

Freshfields saw roughly 500% growth in Claude usage in six weeks after its April 2026 rollout, and Linklaters' internal chatbot Laila is used more than 200,000 times a month. The firms with the most use also have the most rules. A sequence that works:

  1. Publish Schedule A with at least one approved general tool on a business tier, so nobody has to choose between the policy and getting work done.
  2. Put the three sample clauses above (tools, verification, incidents) into force first; they cover the sanctions risk. Perfect the rest later.
  3. Log usage centrally where the platform allows it, and tell people the log exists; that is what turns shadow AI into visible AI.
  4. Run the first training session on the firm's own tools and record attendance the same day.
  5. Announce the amnesty in section 7 out loud. Candour lowered the outcome in every 2026 sanction case where it appeared.
  6. Set a quarterly review, because the tool list will be wrong within three months.

The one-day working template in the policy playbook turns this outline into a document, and Module 6 of the AI Lab for Lawyers course works through the risk and regulation behind each section.

Frequently asked questions

Does a law firm legally need an AI policy?

No statute in the US, UK or EU says a law firm must have one, but the pressures converge. Article 4 of the EU AI Act expects deployers to take measures on AI literacy and the Commission recommends an internal training record. Ayinde v Haringey addressed managing partners directly, the SRA's August 2026 notice expects firm-level supervision, and US malpractice carriers now ask about AI governance on application forms.

Should a law firm AI policy ban ChatGPT?

Ban consumer accounts for client information, not the product. Consumer ChatGPT trains on conversations unless the user opts out; ChatGPT Business and Enterprise do not train by default and sit under a data-processing agreement. A blanket ban pushes lawyers onto personal phones, which is why the North Carolina Bar's 2026 guidance argues against bans and for a traffic-light structure with approved tools.

What should the verification section of an AI policy say?

That the lawyer responsible for a document must personally confirm every AI-assisted authority, quotation and factual assertion against the primary source before filing or sending, checking existence, support and current treatment. It should add that verification cannot be delegated to another AI tool, and that the assigned judge's standing order on AI must be checked before any court filing.

How does an AI policy deal with AI agents rather than chatbots?

With a permissions schedule. The policy should list the actions an agent may take autonomously, such as drafting and summarising, and the actions that always need a named human approval, such as filing, sending to a counterparty, deleting or moving files, signing, paying or changing records. It should also require logging of agent actions and a kill switch that a human, not the agent, controls.

Put this into practice

AI Lab for Lawyers: a live, hands-on course for practising lawyers

4 live sessions (8 hours), recordings included, taught by Dr. Niklas Schmidt of Wolf Theiss. Rated 4.7/5 from 381 ratings; alumni from Norton Rose Fulbright, Paul Hastings, Pillsbury and Taylor Wessing.

Reserve a seat · €200 Maven money-back guarantee · certificate of completion

Playbooks

Write a Law Firm AI Policy in a Day: A Working Template

A working law firm AI policy template with ten clauses, commentary on each, EU and US variants and an hour-by-hour rollout plan that gets it signed in one day.

12 min read

Guides

Which AI Tools Train on Your Data? A Confidentiality Checklist for Lawyers

Which AI tools train on your data, plan by plan as of September 2026: ChatGPT, Claude, Gemini, Copilot, legal vendors, plus bar rules and a 10-point checklist.

11 min read

Regulation & Ethics

EU AI Act Article 4: The AI Literacy Duty for Law Firms, Explained

Article 4 of the AI Act requires every firm whose staff use AI to support their AI literacy. In force since 2 February 2025, supervised nationally since August 2026.

7 min read

Regulation & Ethics

AI Vendor Due Diligence for Law Firms: The Questions to Ask Before You Sign

AI vendor due diligence for law firms starts with the subprocessor chain, then training and retention terms, ZDR eligibility, data residency, audit logs and exit rights.

11 min read

Guides

Agent Security for Lawyers: Prompt Injection, Data Leaks and Guardrails

AI agent security for law firms: prompt injection via hostile documents, the Grok Build and Cowork incidents, and the allowlists and gates that contain them.

11 min read

Regulation & Ethics

Do You Have to Tell Clients You Use AI? Disclosure Rules and Practice

When lawyers must disclose AI use to clients under ABA 512, California, CCBE and ÖRAK rules, when consent is required, and a sample engagement-letter clause.

10 min read