Most law firm AI policies are either one sentence (“do not use it”) or ten pages of principles nobody can apply at 6pm with a filing due. Neither works. AI is already inside the research platform, the office suite and the phone system, so a ban only stops people reporting what they use. A page of principles leaves every real question to whoever is in a hurry.
The template below is a working policy. It says which information may go into which tools, how a tool gets approved, what has to be checked before anything leaves the firm, what the client is told and billed, and what happens when something goes wrong. The schedules are filled in with examples, so you edit rather than draft.
Download the policy as a Word document, or copy it from the page. It is the same text.
What is in it
- Purpose and scope: who and what it covers, including tools a lawyer pays for personally.
- Definitions: six terms, so “AI tool” and “verified” mean one thing.
- Roles: an AI Policy Owner, the responsible lawyer on each matter, and everyone else. In a two-lawyer firm one person can hold both lawyer roles.
- Information classes: three classes of information and which tools each may enter. This is the section people will use most.
- Approving a tool: nine questions a vendor must answer before any client information goes in, and the register that records the answers.
- Red, yellow and green uses: prohibited, allowed with review, and allowed by default.
- Verification: what “checked” means for citations, quotations, facts and figures, and who signs for it.
- Court filings: local rules, standing orders, and what to do the day an error is found after filing.
- Clients: the engagement-letter paragraph, when specific consent is needed, and opt-outs.
- Billing: hourly time, flat fees, learning time and tool costs, as the ABA and Texas opinions treat them.
- Staff, contractors and vendors: supervision beyond the lawyers.
- Incidents: what counts, the first 24 hours, and the log.
- Training, acknowledgment and review.
Four schedules follow: the tool register, the engagement-letter clause, the acknowledgment form and the incident log.
Three decisions to make before you fill it in
Who owns it. Name one person as AI Policy Owner. In a small firm that is usually the managing partner; in a larger one, whoever already owns technology or risk. The policy fails without a name on it, because “can I use this?” needs someone to answer it.
Which tools you already have. Before approving anything, list what people use now, including AI features switched on inside software the firm already pays for and personal accounts used for work. That list is the first draft of Schedule A.
What your default is for clients. Decide whether the engagement letter discloses AI use for every client (the template assumes it does) and whether you will offer an opt-out. Deciding once saves every lawyer improvising it.
The template
[Firm name] Policy on the Use of Artificial Intelligence
Adopted [date]. AI Policy Owner: [name]. Next review: [date].
1. Purpose and scope
1.1 This policy governs the use of artificial intelligence tools in the firm’s work. Its purpose is to protect client information, to ensure that nothing relied on by a client, a court or another party is sent without a lawyer having verified it, and to keep the firm’s use of these tools consistent, recorded and reviewable.
1.2 It applies to every partner, lawyer, paralegal, staff member, contractor and temporary worker, and to every AI tool used for firm work, whether the firm licenses it or an individual does, and whether it is a standalone product or a feature inside other software.
1.3 Where this policy is stricter than a court’s rules or a client’s instructions, follow this policy. Where a court’s rules or a client’s instructions are stricter, follow them.
2. Definitions
AI tool means any software that generates or transforms text, images, audio or data in response to instructions, including chat assistants, drafting and research features, transcription, summarization and AI phone or intake systems.
Approved tool means an AI tool listed in Schedule A for the information class concerned.
Client information means any information relating to the representation of a client or prospective client, whatever its source, including the fact of the representation.
Output means anything an AI tool produces.
Verify means to check an output against its source as set out in section 7, by a person, before it is relied on.
Responsible lawyer means the lawyer responsible for the matter on which the output is used.
3. Roles
3.1 The AI Policy Owner maintains this policy and Schedule A, decides requests to approve tools, receives incident reports, and reports to the partners at least twice a year.
3.2 The responsible lawyer decides whether and how AI tools are used on a matter, ensures outputs are verified, and decides what the client is told.
3.3 Every person covered by this policy uses only approved tools for the information they handle, verifies or has verified what they rely on, and reports incidents under section 12 the same day.
4. Information classes
4.1 Class 1: no client information. Public material, the firm’s own non-confidential material, and work containing no client information. May be used in any AI tool, including public ones, subject to section 6.
4.2 Class 2: client information. Any client information not in Class 3. May be used only in a tool approved for Class 2 in Schedule A.
4.3 Class 3: highly sensitive client information. Health, financial account and government identification numbers; information about minors; criminal matters; immigration status; trade secrets; material under a protective order or seal; and any material a client has asked to be handled with added restriction. May be used only in a tool approved for Class 3 in Schedule A, and only with the responsible lawyer’s approval for that matter.
4.4 Removing names does not move information to a lower class if the client or matter could still be identified from what remains.
4.5 Material under a protective order, a sealing order or a confidentiality agreement may not enter any AI tool unless the order or agreement permits it.
5. Approving a tool
5.1 No client information may enter an AI tool until the AI Policy Owner has approved it for that class and recorded the approval in Schedule A.
5.2 Before approving a tool for Class 2 or Class 3, the AI Policy Owner must have written answers, from the contract or the vendor’s documentation rather than a sales call, to these questions:
- Does the vendor use firm content to train or improve any model? (It must not.)
- Where is content stored, and for how long? Can the firm delete it?
- Who at the vendor can access firm content, and in what circumstances?
- Which sub-processors handle it?
- Is access controlled by firm accounts, with the ability to remove a user?
- Will the vendor notify the firm of a security breach, and within what time?
- Is content from different users or clients kept separate?
- Can the firm export its content if it leaves?
- Do the terms allow the vendor to change these answers without notice?
5.3 A tool that retains or learns from what is entered is approved only for the classes and uses Schedule A records, and client consent under section 9.3 is required before client information is entered into it.
5.4 Each approval is reviewed at least once a year and whenever the vendor’s terms change.
5.5 Personal accounts on consumer AI products are not approved tools, even if the individual pays for a professional tier.
6. Red, yellow and green uses
6.1 Red: prohibited.
- Entering client information into a tool not approved for its class.
- Relying on, filing, sending or quoting any citation, quotation or factual statement from an output that has not been verified.
- Using an AI tool to make a decision the rules of professional conduct require a lawyer to make, including accepting a matter, giving legal advice to a client or setting a fee.
- Allowing an AI tool to communicate with a client, a court or another party in the firm’s name without a person approving each message, except for automated acknowledgments approved in Schedule A.
- Presenting an output as a person’s own work where that would mislead.
6.2 Yellow: permitted in approved tools, with review. Drafting from client material; legal research; summarizing records, transcripts or discovery; translation; client-facing letters and emails. The output must be verified under section 7 by the person relying on it, or by a lawyer they name, before it is used.
6.3 Green: permitted by default. Work with Class 1 information: internal drafts containing no client information, formatting, proofreading, brainstorming, summarizing public material, and learning to use approved tools on non-client material.
7. Verification
7.1 The person who relies on an output is responsible for it, whoever produced it.
7.2 Before an output is relied on:
- every citation is looked up in a primary source and read, and confirmed to exist, to say what the output says it says, and to remain good law;
- every quotation is checked word for word against its source;
- every factual statement about the matter is checked against the record;
- every calculation or figure is recalculated; and
- anything that cannot be verified is removed or clearly marked as unverified.
7.3 For any document filed with a court or sent to another party, the person who verified it records their initials and the date in the matter file.
8. Court filings
8.1 Before filing, the responsible lawyer checks the court’s local rules and the presiding judge’s standing orders for any requirement to disclose or certify AI use, and complies with it. These requirements vary by court and judge and change often.
8.2 Nothing is filed with an authority, quotation or factual statement that has not been verified under section 7.
8.3 If an error in a filed document is found, the person who finds it tells the responsible lawyer the same day. The responsible lawyer corrects the record promptly and candidly, and the incident is logged under section 12.
9. Clients
9.1 The firm’s engagement letter includes the paragraph in Schedule B.
9.2 A client’s instruction that AI tools not be used on their matter, or be used only in particular ways, is recorded on the matter file and followed.
9.3 Before client information is entered into a tool that retains or learns from it (section 5.3), the responsible lawyer obtains the client’s informed consent for that use. The engagement-letter paragraph is not enough on its own. The client is told which tool, why, what information would go in, and the risks.
9.4 If a client asks whether AI tools are used on their matter, they receive an accurate answer.
9.5 The responsible lawyer tells the client about AI use whenever it matters to a decision the client is making, or when the output will be relied on in a way the client would reasonably expect to be told about.
10. Billing
10.1 Hourly work is billed for the time actually spent, including time spent reviewing and verifying outputs. Time saved by an AI tool is not billed.
10.2 Time spent learning a tool the firm uses generally is not billed to clients, unless a client asks for a particular tool on their matter.
10.3 The cost of AI tools is treated as overhead unless the client agrees in advance, in writing, to be charged for a specific tool’s use on their matter, in which case it is charged at actual cost.
10.4 Flat and contingent fees must remain reasonable in light of how the work was actually done.
11. Staff, contractors and vendors
11.1 Lawyers with supervisory authority make sure that paralegals, staff and contractors are trained in this policy before they use AI tools on firm work, and that their use complies with it.
11.2 Contractors and outside providers who handle client information are bound by this policy, or by written terms at least as protective, before they use AI tools on firm work.
11.3 Vendors providing AI tools are approved and reviewed under section 5.
12. Incidents
12.1 An incident is any of the following:
- client information entered into a tool not approved for its class;
- an output relied on that turns out to be materially wrong;
- a security event reported by, or affecting, an AI vendor; and
- any breach of section 6.1.
12.2 The person who discovers an incident reports it to the AI Policy Owner and the responsible lawyer the same day. Within 24 hours they:
- stop further use of the tool on the matter;
- where possible, delete the information from the tool and keep a record of the deletion;
- decide with the responsible lawyer whether the client must be told, and whether any correction must be made to a court or another party; and
- record the incident in Schedule D.
12.3 The AI Policy Owner reviews Schedule D at least every quarter and updates this policy and Schedule A where the log shows a gap. People who report incidents promptly are not disciplined for the report.
13. Training, acknowledgment and review
13.1 Everyone covered by this policy completes the firm’s AI training before using any approved tool on client information, and at least once a year after that.
13.2 Everyone signs the acknowledgment in Schedule C when they join and after each revision of this policy.
13.3 The AI Policy Owner reviews this policy at least every six months, and whenever a tool is added to or removed from Schedule A, a relevant ethics opinion or court rule changes, or the incident log shows a need.
Schedule A: Approved AI tools
| Tool | Approved for | Trains on firm content | Retention | Approved by, date | Next review |
|---|---|---|---|---|---|
| [Example: research platform AI feature, firm license] | Class 1, 2 | No (contract) | Deleted after 30 days | [name, date] | [date] |
| [Example: office suite assistant, firm tenant] | Class 1, 2, 3 | No (contract) | Firm controls | [name, date] | [date] |
| [Example: public chat assistant, personal account] | Class 1 only | Yes, by default | Unknown | [name, date] | [date] |
| [Example: after-hours AI intake line] | Class 2; automated acknowledgment approved | No (contract) | 90 days, firm can delete | [name, date] | [date] |
Schedule B: Engagement letter paragraph
The firm may use technology, including artificial intelligence tools, to help with research, drafting, review and administration. We use only tools the firm has approved, under terms that keep your information confidential and do not allow it to be used to train the provider’s systems. A lawyer reviews and is responsible for all work we provide to you. We do not bill you for time saved by these tools or for learning to use them. If we propose to use a tool that would retain your information in a way these terms do not cover, we will ask for your consent first. You may ask us at any time how these tools are used on your matter, or instruct us not to use them.
Schedule C: Acknowledgment
I have read the [Firm name] Policy on the Use of Artificial Intelligence, version dated [date]. I understand which information I may enter into which tools, that I must verify any output before relying on it, and that I must report an incident the same day. I will comply with the policy.
Name: ______________________ Signature: ______________________ Date: __________
Schedule D: Incident log (one entry per incident)
- Date and time discovered, and by whom
- Matter and client
- Tool involved
- What happened: what was entered or relied on, and what was wrong
- Class of information involved
- Actions taken in the first 24 hours
- Client told? When, by whom, and what was said
- Court or other party told or corrected? When and how
- Change made to the policy or Schedule A as a result
How to adopt it in a week
Day 1. Inventory. Ask every person which AI tools and features they use for work, including personal accounts and features inside existing software. Do not penalize answers; you need the true list.
Day 2. Vendor answers. For each tool anyone wants to keep, get written answers to the nine questions in section 5.2. Most come from the vendor’s data processing terms. Anything that trains on firm content by default goes in Class 1 only.
Day 3. Fill it in. Name the AI Policy Owner, complete Schedule A, adjust the Class 3 list for your practice areas, and delete anything that does not apply.
Day 4. Clients and courts. Put Schedule B into your engagement letter template. List the courts you file in and check each one’s local rules and the standing orders of the judges you appear before for AI disclosure requirements.
Day 5. Train and sign. Walk everyone through sections 4, 6 and 7 using the firm’s own examples, then collect Schedule C. An hour with real material does more than any memo.
Why each section is there
| Section | Authority | What it answers |
|---|---|---|
| 4, 5 Information classes and approval | Model Rule 1.6; Opinion 512; Texas Opinion 705 | Lawyers must understand how a tool uses and stores data and safeguard client information. Texas says to review, and consider renegotiating, the terms of service. |
| 5.3, 9.3 Tools that retain or learn | Opinion 512 | Informed consent is recommended before client confidences go into such a tool, and boilerplate consent in an engagement letter is not adequate. |
| 6, 7 Uses and verification | Model Rules 1.1, 3.1; Texas Opinion 705 | Competence includes understanding the tool’s limits, and lawyers may not rely blindly on its output. |
| 8 Court filings | Model Rules 3.3, 8.4(c); court rules | Output must be reviewed before it reaches a tribunal. Some Texas federal courts now require disclosure, and a Dallas lawyer has been sanctioned for not making it. |
| 9 Clients | Model Rule 1.4; Opinion 512 | When and how much the client must be told depends on the circumstances; the policy sets the firm’s default. |
| 10 Billing | Model Rule 1.5; Opinion 512; Texas Opinion 705 | No charge for learning a tool used generally unless the client asks for it; costs are overhead unless agreed in advance; no billing for hours not actually worked. |
| 11 Staff and vendors | Model Rules 5.1, 5.3; Opinion 512 | Managers must set clear policies and make sure nonlawyers, inside and outside the firm, are trained. |
| 12, 13 Incidents, training, review | NIST generative AI profile | Governance continues after a tool is deployed: record failures, review them, change the controls. |
The sanctions orders since 2023 show what sections 7 and 8 prevent. Section 7.2 is the cite-check routine written as a rule.
What this is not
It is not a security policy. The approval questions in section 5 need input from whoever runs the firm’s IT, and Schedule A should point to the vendor’s actual terms rather than paraphrase them. It is not legal advice, and it is not finished when it is adopted. It is working when the incident log has entries and the policy has changed because of them. If you want help putting it into practice with your own matters, that is what AI training for lawyers on this site is for.
Sources and methodology
- Scope
- Written for US firms of one to fifty lawyers, with the ethics authorities cited from the ABA Model Rules, ABA Formal Opinion 512 and Texas Opinion 705. Your state's rules govern and some are stricter. The template is a starting draft, not legal advice; a lawyer at the firm should check it against your state's rules and the local rules of the courts you appear in before adoption.
- How this was produced
- Built from the author's work designing governance for AI systems in legal workflows and from the cited guidance. Each section maps to a Model Rule or opinion, set out in the table at the end. The red, yellow and green structure follows the North Carolina Bar Association's recommendation; the register and incident controls follow the NIST generative AI profile. The template has not been adopted by a named firm.
- Formal Opinion 512: Generative Artificial Intelligence ToolsAmerican Bar Association, Standing Committee on Ethics and Professional Responsibility
- ABA Ethics Opinion on Generative AI Offers Useful FrameworkAmerican Bar Association, Business Law Today
- Opinion 705: Lawyers' use of generative artificial intelligenceProfessional Ethics Committee for the State Bar of Texas
- Beyond the Ban: Why Your Law Firm Needs a Realistic AI Policy in 2026North Carolina Bar Association, Center for Practice Management
- Dallas Attorney Sanctioned For Not Disclosing Generative AI UseBloomberg Law
- Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence ProfileNational Institute of Standards and Technology
Read the editorial standards, corrections policy and AI-use disclosure.