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.
On this page
A workable law firm AI policy fits on two pages: ten clauses covering scope, approved tools, data classes, verification, disclosure, billing, training, agents, incidents and review. The template below gives the clause wording and the reason behind each, with EU and US variants, and a rollout plan that gets it signed and circulated in a single working day.
Why a day, and why two pages
Clio's 2025 Legal Trends Report found 79% of legal professionals using AI while 53% said their firm had no AI policy or did not know. Thomson Reuters' Future of Professionals report (June 2026, 1,816 respondents) found 34% using unsanctioned "shadow AI", and AI meeting expectations for 66% of organisations with a named strategy against 22% without. The gap between use and rules is where the sanctions live.
Length is not the cure. In Johnson v. Dunn (N.D. Ala., 23 July 2025) three Butler Snow lawyers were removed from a prison-safety case and reprimanded over ChatGPT citations; the firm had an AI policy dating from June 2023. In Wadsworth v. Walmart (D. Wyo., 24 February 2025) the tool was the firm's own internal platform. A policy nobody opens under deadline is a document, not a control. The North Carolina Bar's January 2026 guidance adds that blanket bans drive "shadow AI" and recommends a traffic-light structure with human review and mandatory training (NC Bar, 13 January 2026).
Two more audiences read it. The Divisional Court in Ayinde v Haringey [2025] EWHC 1383 (Admin) addressed "heads of chambers and managing partners" directly (Burges Salmon), and the SRA's Warning Notice of 17 August 2026 followed 42 misuse reports in a year. Indemnity carriers, none of which had filed an AI exclusion as of 5 May 2026, now ask about governance; one application form asks whether the firm lets AI "draft documents" (carrier tracker). The regulatory outline is in how to write a law firm AI policy; this page is the working text.
The ten clauses at a glance
| # | Clause | What it settles | Anchor |
|---|---|---|---|
| 1 | Scope and definitions | Who and what is covered, including agents and self-built tools | Supervision rules; SRA notice |
| 2 | Approved tools and accounts | Named tools and plans; no personal accounts | Vendor data terms; § 43a BRAO |
| 3 | Data classes | Red, yellow, green | ABA 512; California Rule 1.6 comment; ÖRAK |
| 4 | Verification | Personal, non-delegable check | California guidance, 14 May 2026 |
| 5 | Disclosure | Engagement letter; court certifications | CCBE Guide; 30+ federal districts |
| 6 | Billing | Actual time only; no mark-up | ABA 512; Texas 705; Virginia LEO 1901 |
| 7 | Training and record | Mandatory sessions, logged | AI Act Article 4; CLE rules |
| 8 | Agents and self-built tools | Confirmation gates; credentials; synthetic data | Claude for Legal repo; Zenity |
| 9 | Incidents | Hallucination after filing; data in the wrong tool | Couvrette; UK judiciary guidance |
| 10 | Ownership and review | Named owner; monthly and quarterly checks | Vendor setting changes |
The template, clause by clause
Replace the square brackets. Keep the clause text; the commentary can become internal guidance if a longer document is wanted.
1. Scope and definitions
Clause. This policy applies to every partner, employee, contractor and secondee of [Firm] and to any AI tool used in connection with the firm's work, whether or not the firm pays for it. "AI tool" includes chat assistants, AI features inside other software, agents that take actions, and tools built by staff. The supervising lawyer on a matter is responsible for AI use on that matter by anyone.
Commentary. The last sentence is the one that matters. ABA Formal Opinion 512 (29 July 2024) grounds AI duties in the existing supervision rules (ABA Opinion 512); the SRA expects solicitors to understand a tool "well enough to review its output critically". Naming the supervising lawyer removes the argument that nobody was responsible.
2. Approved tools and accounts
Clause. Client-related work may be done only in the tools and plans listed in Schedule A [for example: ChatGPT Enterprise; Claude Team; Microsoft 365 Copilot; Harvey]. Personal or consumer accounts, including paid personal subscriptions, may not be used for client-related work. Accounts are created and configured by [IT/the AI owner] using the settings in Schedule B, and are re-checked monthly.
Commentary. Name the plan, not the brand. Consumer tiers of ChatGPT, Claude, Gemini and Perplexity train on chats unless opted out; Business, Team, Enterprise and API tiers do not by default. The German BRAK treats tools that store or forward client data for training as a breach of § 43a Abs. 2 BRAO. Schedule B is the privacy settings playbook, applied to your accounts.
3. Data classes: the traffic light
Clause. Red data may not be entered into any AI tool without the written approval of [the AI owner]: privileged communications, client identities linked to matter facts, personal data of third parties, material under a confidentiality undertaking, and anything a court has sealed. Yellow data may be used in Schedule A tools only: matter documents, drafts, research questions and anonymised fact patterns. Green data may be used in any Schedule A tool: public documents, marketing, internal administration and training material. When in doubt, the class is red.
Commentary. The proposed California comment to Rule 1.6 treats exposing client data to AI as "revealing" it where there is "a material risk of unauthorized access, retention, or use" (Wilson Elser on the 2026 California guidance); ABA 512 requires informed consent where a self-learning tool may expose client data. The Austrian ÖRAK adds a location test, client data stays on EU servers (Anwaltsblatt on the ÖRAK KI-Tag), which EU firms should write into the red definition.
4. Verification
Clause. No document leaves the firm, and no paper is filed with any court or tribunal, containing a citation, quotation, factual assertion or figure generated with AI that the responsible lawyer has not personally read and verified against the primary source. Verification is recorded in the matter file using the firm's checking table. Asking an AI tool to verify its own output is not verification.
Commentary. The first sentence adapts California's May 2026 guidance: "No paper filed in court should contain citations that the attorney responsible for the filing has not personally read and verified." Charlotin's database stood at 2,041 decisions on 14 September 2026, 811 involving lawyers. The checking table is in how to verify AI legal research.
5. Disclosure to clients and courts
Clause. The firm's engagement letter states that the firm uses AI tools under this policy, names the categories of tool, and gives the client the right to object. Before any filing, the responsible lawyer checks the court's and the judge's current AI rules and complies with any certification or disclosure requirement. Client-facing automated systems identify themselves as AI.
Commentary. The CCBE's Guide of 2 October 2025 asks lawyers to inform clients and let them object; Clio found 78% of clients want disclosure while 35% of lawyers rarely disclose. Over 30 of the 94 US federal district courts had disclosure or certification rules by April 2026 and 300+ judges have standing orders; Illinois requires none. The last sentence implements Article 50 of the EU AI Act, applicable since 2 August 2026. See AI disclosure to clients and court standing orders on AI.
6. Billing
Clause. Time spent using, prompting and checking AI tools is billed as time actually spent. Time saved by AI is not billed. AI subscriptions are firm overhead and are not charged to clients unless the engagement letter says otherwise. Per-use AI costs are passed through at cost, without mark-up, and only with the client's informed written consent.
Commentary. ABA 512, Texas Opinion 705 (February 2025) and California's 2026 guidance agree on the first two sentences (Texas Opinion 705). Virginia LEO 1901 (November 2025) adds that a reasonable value-based fee need not be cut because AI saved time, which is why the clause leaves fixed and value fees alone. Detail in billing clients for AI.
7. Training and the literacy record
Clause. Every person covered by this policy completes the firm's AI training within [30] days of joining and at least annually thereafter, and before using any tool in Schedule A for the first time. [The AI owner] keeps a record of who attended which session, on what date, covering what content, and how understanding was checked.
Commentary. Article 4 of the EU AI Act has applied since 2 February 2025 and was not deferred by the Digital Omnibus, Regulation (EU) 2026/1744, which reworded it as a duty to "take measures to support the development of AI literacy" (EUR-Lex). The Commission's FAQ of 27 July 2026 says no certificates or tests are required but an internal training record should be kept (Commission AI literacy FAQ); national authorities have supervised the duty since 2 August 2026. US firms arrive at the same clause via Comment 8, adopted in 40 states, and via courts: Beus Gilbert PLLC v. BYU (D. Utah, 9 September 2026) added two AI-ethics CLE courses to a $3,000 sanction. Content is in AI literacy for law firms.
8. Agents, automation and self-built tools
Clause. An AI agent may not send, file, sign, pay, delete or change a record outside its own workspace without a named person confirming that specific action. Agents run with minimum permissions, never with a lawyer's personal credentials, and their limits are set in the platform, not in the prompt. Tools built by staff, including automations and vibe-coded apps, are registered with [the AI owner] before use and built on synthetic data only. Meeting recording is off by default and used only with the documented consent of everyone on the call.
Commentary. This clause did not exist in 2024 policies and is now the most important one. Anthropic's open-source Claude for Legal plugins treat every output as "a draft for attorney review" with explicit gates before filing or sending (claude-for-legal repository); Harvey's agents "flag moments requiring user confirmation". The lesson of 2026 is that instructions are not controls: in July 2026 Grok Build uploaded whole repositories including secrets users had told it to ignore, and in February 2026 a desktop agent asked to "organise" a folder deleted 15 years of photos. Zenity's PleaseFix research: "put hard limits around the agent that the agent cannot override". On recording, NYC Bar Formal Opinions 2025-6 and 2026-2 make not recording the default (Duane Morris). See agent security for lawyers and AI meeting notes for lawyers.
9. Incidents
Clause. Anyone who discovers that a filed or sent document contains AI-generated material that fails verification reports it to the responsible partner the same day; the firm corrects the record with the court or recipient promptly and candidly. Anyone who discovers red data in a non-approved tool reports it to [the data protection lead] the same day, and the event is handled as a data incident. Nobody is disciplined for reporting.
Commentary. The record US sanction, $110,204.38 in Couvrette v. Wisnovsky (D. Or., 12 December 2025), was driven by the cover-up: the lawyer quietly refiled amended briefs rather than confessing (ABA Journal). The UK judiciary's guidance of 31 October 2025 treats leaks into public AI tools as data incidents; in the EU that starts the GDPR Article 33 clock. The no-retaliation sentence makes the other two work.
10. Ownership and review
Clause. [Name] owns this policy. The account settings in Schedule B are re-checked on the first working day of each month and the check is logged. The policy is reviewed every quarter and whenever a Schedule A vendor changes its terms, models or products. Version and date appear at the top of the document.
Commentary. Vendors changed defaults or products at least four times in thirteen months, from Anthropic's consumer training change of 28 August 2025 to the merger of Cowork into the Claude app on 17 September 2026. A policy with no owner is out of date the first time that happens.
EU and US variants: what to swap
| Topic | EU and UK firms | US firms |
|---|---|---|
| Training record (clause 7) | Article 4 record: who, when, what, how assessed; keep it for the national market-surveillance authority | CLE log; New Jersey requires one technology credit every two years from 1 January 2027, Florida three technology hours per cycle |
| Data location (clause 3) | Add "client data stays on EU servers" (ÖRAK) and an Art. 28 GDPR processing agreement per tool | Informed consent language from ABA 512; check your state's opinion (17 jurisdictions plus the ABA as of June 2026) |
| Disclosure (clause 5) | CCBE: inform and allow objection; Article 50 self-identification for client-facing systems | Engagement-letter disclosure; per-court certification check; no duty in Illinois |
| Recording (clause 8) | Consent plus a data-protection basis | NYC Bar 2025-6 and 2026-2; twelve all-party-consent states including California, Florida and Illinois |
The one-day rollout
- 09:00 Inventory. Ask every fee-earner which AI tools they used last week and on which account. Expect personal accounts; do not punish the answers. The list becomes Schedule A, and the first incident-log entry if red data is already out.
- 10:00 Clauses 1 to 3. Fill in tools, plans and data classes. Decide the two or three tools the firm will pay for; a six-week pilot can follow, but the policy should not wait for it.
- 11:00 Clauses 4 to 6. Adopt the verification table, draft the engagement-letter sentence, pick the billing rule. These are decisions, not drafting; the wording above is usable as is.
- 12:00 Clauses 7 to 10. Name the owner and the data protection lead, book the first training session, set the incident route.
- 14:00 Partner review. Thirty minutes, decisions only: "can we live with this", not "can we improve the prose".
- 15:00 Circulate. Two pages with a one-line acknowledgement request. The acknowledgement list is the first page of the Article 4 record.
- 16:00 Settings. Run the settings check on every firm account and file the log as Schedule B.
- 17:00 Diary. Book the 30-day review and the first quarterly review.
Where this leaves you
By the end of the day the firm has a written answer to the question a court, regulator, insurer or client will eventually ask. It converts "we told people to be careful" into named tools, named data classes, a personal verification duty, a confirmation gate on agents and a training record with dates in it. Firms with a formal AI strategy were 3.9 times more likely to report measurable benefit in Thomson Reuters' 2026 research; the document that keeps you out of Charlotin's database is the one that makes the tools pay.
Module 6 of the AI Lab for Lawyers course, on risks and regulation, covers the professional rules and AI Act duties behind clauses 3 to 7.
Frequently asked questions
Does a law firm need a written AI policy?
No rule in the US, UK or EU mandates a document called an AI policy, but the duties it implements are mandatory: confidentiality, competence, supervision, candour to courts, and in the EU the Article 4 AI-literacy duty since 2 February 2025. Regulators, insurers and courts now ask what the firm's rules were. A firm without a written answer is arguing from memory.
How long should a law firm AI policy be?
Two pages, or about ten short clauses. The policy is read by busy lawyers under deadline pressure, which is exactly when hallucinated citations get filed. A longer document can sit behind it as guidance, but the rules themselves should fit on a screen. Butler Snow's 2023 policy did not prevent sanctions in 2025; length was not the problem, but reading was.
Should the policy ban ChatGPT?
Ban consumer accounts for client data, not the tool. Personal ChatGPT, Claude, Gemini and Perplexity accounts train on chats unless opted out and are discoverable; Business, Team, Enterprise and API tiers do not train by default. A blanket ban drives use to personal phones, which the North Carolina Bar calls shadow AI. Approve a firm-paid tier, then forbid everything else for client work.
What is the EU AI Act Article 4 training record?
Article 4 obliges deployers, including law firms whose staff use tools like ChatGPT, to take measures to support AI literacy. The Commission's FAQ, updated 27 July 2026, says no certificates or tests are required but the organisation should keep an internal training record. In practice that means a list of who attended which session on which date, what it covered and how understanding was checked.
Do we need a separate policy for AI agents?
Not a separate document, but a separate clause. Agents that draft, send, file, sign or change records need rules that chat assistants do not: a named human confirmation before any external action, credentials the agent cannot read, and limits set in the platform rather than in the prompt. The July 2026 Grok Build incident showed that telling an agent to ignore a file is not an access control.
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.
Related reading
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.
9 min read
Regulation & Ethics
ABA Formal Opinion 512 Explained: What US Lawyers Must Do When Using AI
ABA Formal Opinion 512 applies six Model Rules to generative AI: competence, confidentiality, communication, candour, supervision and fees. What each duty requires in 2026.
8 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
Playbooks
AI Privacy Settings for Lawyers: Step-by-Step in ChatGPT, Claude, Gemini, Copilot and Perplexity
Exact steps to turn off training and check retention in ChatGPT, Claude, Gemini, Copilot and Perplexity as of September 2026, plus admin settings and a log.
10 min read
Guides
AI Literacy for Law Firms: What to Teach and What Article 4 Expects
AI literacy for law firms under Article 4 of the EU AI Act: an obligation of effort, no certificates, keep a record. A curriculum by role and a training record.
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