How to Write a Law Firm AI Policy Attorneys Will Actually Follow

Why do attorneys ignore law firm AI policies?

Attorneys skip law firm AI policies that are long, vague or ban everything. A ten-page policy written in compliance language does not get read between hearings. A policy that bans AI outright, with no approved tool, pushes people to personal accounts on their phones, where the firm has no visibility and no contract. A policy attorneys follow is short, names the tools they can use, and makes the approved path the easy one.

The stakes are professional. Business and Professions Code section 6068(e)(1) requires California attorneys to maintain client confidences inviolate and to preserve client secrets at every peril to themselves. In United States v. Heppner, a federal court in New York held in February 2026 that documents a defendant generated with a public AI platform on his own initiative were not protected by privilege or the work product doctrine, pointing to the platform's privacy policy. The case involved a client acting alone, and it still shows how a court reads public AI terms.

What do the ethics rules expect a firm AI policy to cover?

They expect the firm to have one, to train on it, and to apply it to staff as well as lawyers. California Rule of Professional Conduct 5.1(a) requires lawyers with managerial authority to make reasonable efforts to ensure the firm has measures giving reasonable assurance that all its lawyers comply with the rules and the State Bar Act. Rule 5.3(a) applies the same duty to the conduct of nonlawyers. The State Bar of California's generative AI guidance says managerial and supervisory lawyers should establish clear policies on the permissible uses of generative AI and provide training.

That guidance and ABA Formal Opinion 512 supply most of the content:

• Confidentiality: do not input confidential client information into a generative AI tool that lacks adequate confidentiality and security protections, and review the tool's terms of use.
• Consent: ABA Opinion 512 says informed client consent is required before inputting information relating to the representation into a self-learning tool, and that boilerplate engagement letter language is not adequate. The opinion is persuasive authority in California.
• Competence: the comment to Rule 1.1 includes keeping abreast of the benefits and risks of relevant technology.
• Fees: the State Bar guidance says a lawyer may charge for actual time spent using generative AI and must not charge hourly fees for the time saved.
• Client instructions: review any client guidelines that restrict or limit AI use.

What does California SB 574 add to a law firm AI policy?

It puts several of these expectations into statute, effective January 1, 2027. Governor Newsom signed SB 574 on September 30, 2026. The bill adds Business and Professions Code section 6068.1 and amends Code of Civil Procedure section 128.7:

• No delegation: an attorney may not delegate the practice of law to generative AI.
• Verification: attorneys must take reasonable steps to verify AI output, including case and statutory citations, and correct erroneous or hallucinated content. Under amended section 128.7, a filed paper may not contain a citation the responsible attorney has not personally verified, and the firm is jointly responsible absent exceptional circumstances.
• Confidentiality: confidential, personally identifying or other nonpublic information may not go into a generative AI system unless access is restricted to the attorney and people the attorney authorizes who are obligated to keep it confidential.
• Disclosure: attorneys who use generative AI must disclose that use to the court for documents submitted.

Each of these maps to a line in the one-page policy below. Have counsel confirm how the firm will apply the disclosure and confidentiality provisions before the effective date.

What should the policy actually say?

Keep it to one page, organized around what attorneys do every day:

• Approved tools. Name each approved AI tool, the tasks it is approved for, and who to ask about anything else.
• Data rules by category. Public information: any approved tool. Firm internal information: approved tools only. Client confidential information: only tools under a firm contract that restricts use of inputs, and only on matters without client restrictions. Privileged or screened matter material: only where the matter team is cleared, and never in a public tool.
• Verification. The responsible attorney checks every citation, quotation and factual statement in AI output against the source before it leaves the firm. Under SB 574 this becomes a statutory duty on January 1, 2027.
• Client consent and instructions. Check engagement terms and outside counsel guidelines before using AI on a matter.
• Billing. Bill time actually spent. Do not bill hours the tool saved.
• Staff. Paralegals, assistants and contractors follow the same rules, and the supervising attorney is responsible for their use.
• Incidents. If client information goes into an unapproved tool, report it the same day to a named person.
• Exceptions. Requests go in writing to the AI policy owner.

Put examples, the approved tool list and vendor details in an appendix that can change without reopening the policy.

How do you make the approved path easier than the workaround?

Give attorneys an approved tool that works, signed in with their firm account, before blocking anything else. People reach for personal accounts when the firm option is missing or slower. Once the approved tool is in place:

• Block personal accounts on firm devices. Tenant restrictions block personal Microsoft accounts and outside tenants for Microsoft services such as Copilot; for other AI services, use web filtering and the vendor's workspace controls so only the firm account works. See how to block personal AI accounts on firm devices.
• Apply sensitivity labels to privileged and screened matter material, and use DLP to warn or block when labeled content is pasted or uploaded to unapproved AI sites.
• If Microsoft 365 Copilot is approved, enforce ethical screens inside Microsoft 365 before rollout, since Copilot shows each user whatever that user can already open.
• Turn on single sign-on so the approved tool needs no separate password.

Each control should match a sentence in the policy. When an attorney hits a block, the message should name the approved tool and the person to contact.

How do you get attorneys to follow it?

Make a named partner the sponsor, train in 30 minutes with the firm's own examples, and collect a signed acknowledgment. Training lands when it uses familiar work: a deposition summary, a demand letter, a research memo, each shown once in the approved tool and once as the mistake to avoid. Use the Heppner decision as a teaching example of how public AI terms can play out in court.

Repeat the acknowledgment each year and whenever the approved tool list changes. Review DLP warnings and blocked sign-ins monthly for the first quarter. If they cluster in one practice group, that group likely needs a better tool or clearer rules, so fix that first.

What should your firm do this quarter?

1. Name an AI policy owner and a partner sponsor.
2. Choose at least one approved AI tool with contract terms on data use the firm has reviewed.
3. Cut the policy to one page using the sections above.
4. Update the policy for SB 574 before January 1, 2027, with counsel confirming how the firm applies the disclosure and confidentiality provisions.
5. Train every attorney and staff member and collect acknowledgments.
6. Turn on tenant restrictions, labels and DLP for privileged material.
7. Review DLP and sign-in reports monthly for the first three months.

How s90 helps

s90 helps law firms of 5 to 100 attorneys stand up the approved AI tool, tenant restrictions, sensitivity labels and DLP rules behind a firm AI policy, plus the reports that show the controls are working. Your counsel owns the ethics analysis, and s90 makes the technical controls match the policy you adopt. If you want to see where your firm's AI use stands today, ask for a short review.