California wants lawyers to verify every AI output: what the State Bar's proposed rule requires

Key Takeaways

What happened: at its March 13, 2026 meeting, the State Bar of California's Committee on Professional Responsibility and Conduct approved proposed amendments to six Rules of Professional Conduct that write AI duties into the black-letter rules, and opened a 45-day public comment window. The headline language: a lawyer must independently review, verify, and exercise professional judgment regarding any output generated by the technology.
The new client-disclosure trigger: a proposed comment says a lawyer must inform the client when AI materially affects the scope, cost, manner, or decision-making of the representation. That moves AI use from a private workflow choice into a client-communication duty.
Agentic AI is now in scope: the updated 2026 Practical Guidance, which replaces the 2023 version, addresses agentic AI at the California Supreme Court's request, meaning tools that act and not just draft are squarely on the radar.
What to do now: the rule is a competence mandate dressed as compliance. Senior attorneys already exercise judgment; the gap is proving it. Stand up a verification-of-record workflow, add a client-disclosure paragraph to your engagement letter, and map which agentic actions need a human gate before the rule lands.

The Leveraged Years Briefing. Permalink

What the California Bar amendments require

For most of the AI era, the rules a lawyer answered to about technology were inferred. You read the competence duty, the confidentiality duty, the supervision duty, and you reasoned your way to how they applied to a chatbot. California is now proposing to stop making you infer. At its March 13, 2026 meeting, the State Bar's Committee on Professional Responsibility and Conduct approved a package of amendments to six Rules of Professional Conduct that name the technology directly, and it opened a 45-day public comment period that puts the profession on notice.

The center of gravity is a single requirement. Under the proposed language, a lawyer who uses generative tools must independently review, verify, and exercise professional judgment regarding any output generated by the technology. Read that slowly, because the word doing the work is every. Not the suspicious output, not the output you happen to doubt, but any output. The rule does not care whether the draft looks clean. It assumes the draft might be wrong in a way you cannot see by reading it, and it puts the burden of catching that on you.

The six rules touched are the ones you would expect once you think about where a tool can hurt a client: competence, communication, confidentiality, candor toward the tribunal, and the supervision rules that make a partner answerable for what the people and now the tools under them produce. The amendments do not invent new duties out of thin air. They take duties California lawyers already carry and make explicit that they survive contact with AI. That is the quiet thing to understand. This is not a new burden so much as a refusal to let an old one be outsourced.

The package is a proposal, not yet law. The comment window is the point of the news. The Bar is asking the profession to weigh in before the language is final, which means the smart move is not to wait for the rule to land and then scramble. It is to build the workflow now, while you can still shape it and while you have runway to make it routine.

The verify-every-output and client-disclosure duties

Two duties in this package change daily practice the most, and they pull in different directions, which is why they are worth separating.

The first is verification. The verify-every-output language is a direct answer to the year of headlines about fabricated citations and invented holdings. But notice what it does not say. It does not say do not use AI. It says you own the output. The model is allowed to be wrong; you are not allowed to pass its work along as if you checked it when you did not. In litigation that means cite-checking every authority and confirming every quote and holding before the brief is signed. In transactional work it means reading the clause the model drafted against the deal you actually negotiated, not the deal it assumed. The duty is not heavy for a lawyer who already reads their own work. It is fatal for one who has quietly started trusting the draft.

The second duty is newer and easier to miss. A proposed comment says a lawyer must inform the client when AI materially affects the scope, cost, manner, or decision-making of the representation. This is the communication rule reaching into your tooling. If AI is meaningfully changing how the matter is staffed, what it costs, or how decisions get made, the client has a right to know. The trigger is materiality, not mere use, so a model that helps you organize records is different from one that reshapes how you bill or how the strategy gets set. The practical effect is that AI use becomes a thing you disclose on the front end, in writing, rather than a thing a client discovers later and resents. The lawyers who handle this well will treat it the way they already treat any material term of the engagement: name it, explain it, and put it in the letter.

Agentic AI guidance

The part of this package that points furthest forward is the smallest in word count. The updated 2026 Practical Guidance, which replaces the 2023 version, now addresses agentic AI, and it does so at the California Supreme Court's request. That last detail matters. When the state's highest court asks the Bar to speak to a technology, the technology has stopped being a curiosity.

The reason agentic AI gets its own treatment is that it breaks the assumption underneath the verify-every-output rule. Verification assumes there is an output sitting in front of you to check before anything happens. An agent does not wait. It acts: it files, it sends, it accepts terms, it moves through a sequence of steps and only stops when it is done or when it hits a wall. By the time you see the output, the action may already be taken. A drafting tool that hallucinates a case gives you a chance to catch it before you sign. An agent that files the brief does not.

That is why the supervision and verification duties have to migrate upstream for agentic tools. You cannot review every output after the fact when the tool produces consequences as it goes. You have to decide in advance which actions an agent is allowed to take on its own and which require a lawyer to approve before they fire. The Bar addressing agentic AI now is a signal that this checkpoint thinking is becoming part of the competence baseline, not an optional sophistication. The senior lawyer's edge here is exactly the instinct that experience builds: knowing which moves are irreversible and gating those before anything automated touches them.

A compliance workflow you can run now

None of this requires a committee or a new platform. It requires a record. Here is the workflow that turns the proposed rule into a routine you can defend.

Build a verification-of-record habit. For any work product where AI touched the substance, keep a short note of what you checked: citations confirmed, quotes and holdings verified, key facts traced to the source. The note does not need to be long. It needs to exist, so that your verification is a documented fact rather than a claim. The rule asks you to verify; the record proves you did.
Add a client-disclosure paragraph to the engagement letter. Draft one clean paragraph that tells the client you may use AI tools to assist the work, that a lawyer reviews and remains responsible for everything produced, and that you will inform them if AI materially affects scope, cost, or decisions. Get it in the letter at intake so the disclosure duty is satisfied on the front end rather than improvised later.
Map your agentic checkpoints. List the actions any agentic tool could take in your practice, then mark each one as auto-allowed or human-gated. Filing, sending to opposing counsel, accepting terms, and anything that touches a deadline get gated. Summarizing, organizing, and drafting do not. Write the map down so the rule is enforced by your setup and not by memory.
Lock confidentiality at the tool level. Use workspace controls and a no-training configuration so client material is not used to train models and does not leave the protected environment. Confidentiality is one of the six rules in the package, and it is the one most easily waived by a default setting nobody checked.
Set the supervision rule for your team. Make explicit that the lawyer who signs the work owns the AI output in it, the same way a partner owns a junior's draft. The supervision rules in the package mean the buck does not stop at the tool.

Run those five and you are not reacting to the rule. You are ahead of it.

Competence and candor guardrails

Two of the six rules deserve a second look, because they are where a quiet workflow choice becomes an ethics problem.

Competence is the rule the verify-every-output language lives inside, and it is doing more than telling you to check the model. It is telling you that delegating the judgment is itself the breach. A lawyer who lets the draft stand because it reads well has not been incompetent at using AI; they have been incompetent at lawyering, which is the thing the rule actually protects. The competence duty does not ask you to be a prompt engineer. It asks you to stay the professional in the room, which means the output is raw material and the judgment is yours.

Candor toward the tribunal is the rule with teeth. The year of sanctions stories was never really about hallucinated cases; it was about lawyers who let unverified AI output reach a judge and then, too often, compounded it. California writing AI into the candor rule closes the gap between I did not know it was wrong and I did not check. Under the proposed language, not checking is not a defense. Verify before you file, disclose AI use where the rules require it, and never let a tool put something in front of a court that you would not personally vouch for. The guardrail is simple to state and unforgiving when ignored.

What to do now

You do not need to wait for the comment period to close. The duties are foreseeable enough to act on today.

Adopt one firm rule: AI output is raw material, and the signing lawyer verifies and owns it. If it was not verified, it does not go out.
Put a verification-of-record note into your matter file whenever AI touched the substance of the work product.
Add the client-disclosure paragraph to your engagement letter now, so the communication duty is handled at intake.
Map every agentic action your tools can take and gate the irreversible ones behind a human approval.
Confirm your tool configuration is no-training and workspace-controlled, and make supervision of AI output an explicit expectation for your team.

The firms that treat this as a checkbox after the rule lands will be improvising under pressure. The firms that build the record now will find the rule describes what they already do.

What AI does not replace

The proposed rule is, in the end, a long way of saying the judgment is still the job. AI does not decide whether an output is right, whether a disclosure is material, or whether an action is too consequential to automate. Those are the calls the competence rule reserves to a lawyer, and the whole package is built around protecting them. The model can carry the volume underneath the work, but reading the authority, weighing the client's interest, and standing behind the result in front of a court are not delegable, and California is proposing to say so out loud. The leverage is in doing that judgment work cleanly while the tool handles the load beneath it.

If you want to build that discipline into a full operating system for your practice, the role-specific program is below. It is the same skill set the proposed rule assumes you already have: knowing exactly where the tool helps and where the lawyer has to stay in the chair.

Frequently Asked Questions

Is the California rule requiring lawyers to verify AI output final?

Not yet. The State Bar of California's Committee on Professional Responsibility and Conduct approved the proposed amendments to six Rules of Professional Conduct at its March 13, 2026 meeting and opened a 45-day public comment period. The package is a proposal in the comment stage, not yet adopted law. That is why the practical move is to build the verification and disclosure workflow now, while you can still comment on the language and have time to make compliance routine before the rule is finalized.

When do I have to tell a client I used AI on their matter?

Under the proposed comment, the trigger is materiality. A lawyer must inform the client when AI materially affects the scope, cost, manner, or decision-making of the representation. Routine assistance that does not change how the matter is staffed, billed, or decided is different from AI that meaningfully reshapes those things. The cleanest way to handle it is a short disclosure paragraph in the engagement letter at intake, stating that AI tools may assist the work, that a lawyer reviews and remains responsible, and that you will inform the client when AI materially affects scope, cost, or decisions.

Does the rule mean my firm should stop using AI?

No. The proposed rule is a competence and supervision mandate, not a ban. It requires you to independently review, verify, and exercise professional judgment over any AI output, to disclose material AI use to clients, and to supervise agentic tools that act on their own. Lawyers who already read and own their work product, control confidentiality at the tool level, and gate irreversible automated actions are doing exactly what the rule asks. The response is to build a verification-of-record workflow, not to abandon tools that do real work.