The California Bar Just Mapped Agentic AI to Your Supervision Duty
Bindingness: Regulatory Guidance · Scope: California · Legal
On May 14, 2026 the State Bar of California gave lawyers their first formal guidance on AI that acts on its own. The principle is simple and uncomfortable: the more an AI tool can do without you, the more supervision your existing duty asks of you. Here is how that applies to Rules 1.1, 1.6, and 5.3, and a supervision protocol you can run on Monday.
Part of AI Regulation Tracker, our running brief on the rules reshaping AI in professional practice.
For most of the AI era, the question a lawyer asked about a tool was whether its output was right. You read the draft, you checked the citations, you owned what you signed. That model assumes there is an output sitting in front of you to check before anything happens. Agentic AI breaks that assumption. An agent does not wait for you to read it. It plans, it executes, it moves through a sequence of steps and stops only when the task is done or it hits a wall. By the time you see the result, the action may already be taken. A drafting tool that invents a case gives you a chance to catch it before you file. An agent that files the brief does not.
The State Bar of California is among the first to write guidance for that gap. On May 14, 2026 its Board of Trustees approved an updated Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, replacing the version first published on November 16, 2023. The update reflects a directive from the California Supreme Court to consider whether lawyers need specific guidance on agentic tools, defined as systems that can plan and execute tasks and work toward a goal with minimal or no human intervention. The answer the Bar gave is that they do, and the guidance threads agentic considerations through the duties lawyers already carry rather than inventing a new one.
Key takeaways
- Autonomy scales your duty. The core principle of the 2026 guidance is that the more independence you grant an AI system, the more supervisory control and verification you owe. Autonomy is not a way to delegate responsibility. It is a reason to tighten it.
- Three rules carry the weight. Competence (Rule 1.1) keeps the judgment yours, confidentiality (Rule 1.6) makes you responsible for how an agent handles client data across matters, and supervision of nonlawyer assistants (Rule 5.3) reaches the AI tools your team runs.
- Some actions need a human gate. Substantive legal determinations, legal advice, preparing and filing pleadings, and acting in a representative capacity must stay under meaningful lawyer supervision. They need lawyer review before an agent acts on them.
- California is out front, others may follow. This is interpretive guidance, not black-letter rule, and it is issued for California lawyers. But it is among the most detailed treatments any bar has given on agentic AI, and it is a likely reference point as other states and disciplinary bodies confront the same questions.
What changed on May 14, and what did not
It helps to separate two things the State Bar did in 2026, because they are easy to blur and they carry different weight. The first is a set of proposed amendments to six Rules of Professional Conduct, which its Committee on Professional Responsibility and Conduct approved on March 13, 2026 and put out for public comment that closed in early May. Those proposed rules are still in the pipeline; they are not yet adopted. The second is the Practical Guidance, a separate, non-binding interpretive document, and that is what the Board of Trustees approved on May 14, 2026. The proposed rules write AI duties into the black letter. The guidance explains how those duties already apply, and it is the instrument that takes on agentic AI head-on.
That distinction matters for how you act. You do not have to wait for the proposed rules to be adopted to be governed by the underlying duties, because the duties already exist. Competence, confidentiality, and supervision have always applied to whatever tools a lawyer uses. The guidance simply makes explicit how they reach a tool that acts on its own. Treating it as optional because it is not a rule yet is the mistake that produces a disciplinary file later.
The guidance does not ask whether you use agentic AI. It asks whether you can show how you control it.
How agentic AI maps to Rules 1.1, 1.6, and 5.3
The 2026 guidance does not create new duties. It takes three duties California lawyers already carry and shows where an autonomous tool stresses each one. Understanding the map is the difference between a defensible workflow and a guess.
Rule 1.1, competence: the judgment stays yours
The competence duty is where the verify-the-output principle lives, and agentic AI sharpens it rather than softening it. The proposed competence language asks a lawyer to independently review, verify, and exercise professional judgment regarding any output the technology generates. With a drafting tool, you satisfy that after the fact: you read the draft and check it before it goes out. With an agent, after the fact can be too late, because the action is the output. So competence for agentic tools moves upstream. It becomes a duty to decide, before you turn the agent loose, which of its possible actions you are competent to let it take unsupervised and which you are not. Competence is no longer only about checking work. It is about scoping autonomy.
Rule 1.6, confidentiality: you own the agent's data handling
Confidentiality is a key duty the guidance highlights for agentic systems, and for a specific reason. An unrestricted or poorly configured agent can reach across matters. It can pull a document from one client's file into another client's task, or surface privileged material in a context where it does not belong, without anyone intending it. Under Rule 1.6 that is not the tool's failure. It is yours. The lawyer remains responsible for configuring an agentic system so it cannot unintentionally disclose confidential information, and for putting the access controls and monitoring in place that keep one matter walled off from another. Feeding identifiable client data into a tool that is not covered by a confidentiality and no-training configuration is the kind of default nobody checks until it is a problem.
Rule 5.3, supervision: treat the agent as a nonlawyer assistant
This is the conceptual move that makes the whole framework click. Rule 5.3 governs a lawyer's duty to instruct and supervise nonlawyer assistants, the paralegals and staff who do work under a lawyer's authority. The proposed amendment and the guidance extend that duty to the technology those people use, and the practical reading is that the same logic reaches the agent itself. An agentic tool that drafts, organizes, and acts is functionally equivalent to a nonlawyer assistant that never sleeps and never asks. You would not let a new paralegal file a motion without review or email opposing counsel unsupervised. The supervision duty says the same standard applies when the assistant is software. The buck does not stop at the tool, and it does not stop at the vendor. It stops at the supervising lawyer.
| Rule | The duty | What it requires for agentic AI |
|---|---|---|
| 1.1 Competence | Independently review, verify, and apply judgment to any output | Scope which actions an agent may take unsupervised before you deploy it; keep substantive judgment human |
| 1.6 Confidentiality | Protect client information from unauthorized disclosure | Configure the agent to prevent cross-matter leakage; set access controls and monitoring; use a no-training, walled environment |
| 5.3 Supervision | Instruct and supervise nonlawyer assistants and their technology | Treat the agent as a supervised assistant; gate filings, advice, and representative acts behind lawyer review |
The generic competence duty for lawyers using AI is the foundation here, and we cover it as a standalone in the AI competence-duty checklist. This briefing is the layer above it: what changes when the tool stops drafting and starts acting.
The agentic-AI supervision protocol, paste-ready
The guidance gives you a principle. This is one way to turn that principle into a protocol you can drop into a practice memo and follow without a compliance department. The Bar did not publish this checklist; it is our plain-reading implementation of the guidance. It has four parts, each tied to a rule, and the example lines show what a defensible record looks like.
- 1. Classify every agent action by autonomy tier (Rule 1.1). Before any agentic tool runs in your practice, list the actions it can take and sort each into one of three tiers. Auto-allowed: summarizing, organizing files, drafting internal notes, building a research list. Notify-and-proceed: drafting client-facing documents or external correspondence, where the agent prepares but a lawyer reviews before sending. Hard-gated: filing pleadings, communicating legal advice, accepting terms, anything touching a deadline or a court, anything irreversible. Hard-gated actions require an affirmative lawyer approval before they fire, every time.
- 2. Wall off client data and configure for confidentiality (Rule 1.6). Run agentic tools in a workspace that is no-training and matter-scoped, so the agent cannot pull material from one client's file into another's task. Confirm in writing that the configuration prevents cross-matter access, that monitoring is on, and that no identifiable client data reaches any tool not covered by a data agreement. Example record: "Agent operates in matter-scoped workspace X; cross-matter retrieval disabled; no-training confirmed; access logged."
- 3. Name the supervising lawyer for every agent (Rule 5.3). Each agentic workflow gets a named lawyer who owns it the way a partner owns a junior's work. That lawyer is responsible for the agent's instructions, its guardrails, and its output. Make explicit to the team that the same supervision standard applies whether the assistant is a person or a program. Example record: "Supervising attorney for intake agent: [name]. Reviewed configuration and tier map on [date]."
- 4. Keep a supervision-of-record note (Rules 1.1 and 5.3). For any matter where an agent took a notify-and-proceed or hard-gated action, keep one short line: what the agent did, what the lawyer reviewed, and what the lawyer approved. A line that reads "intake agent drafted the engagement letter and assembled the conflict check; partner reviewed both, confirmed the conflict screen, and approved sending" is a defensible record. A blank where that line should be is the gap that turns AI use into a supervision failure.
Our methodology, honestly
This briefing is built on the State Bar of California's public announcement that its Board of Trustees approved the updated Practical Guidance on May 14, 2026, on the California Supreme Court's request that the Bar consider guidance for agentic AI, on the text of the six proposed rule amendments approved March 13, 2026, and on contemporaneous reporting from legal-press outlets covering both. We distinguish the proposed-rule track from the non-binding guidance on purpose, because they are separate instruments at different stages. We do not quote unpublished internal language or cite a precise count of firms affected, because we do not have clean sources for those and will not invent them. The supervision protocol is a plain-reading application of the guidance principle, not a claim that the Bar published this exact checklist.
Why this matters outside California
It is tempting for a lawyer in another state to file this under not-my-jurisdiction. That would be a mistake on two counts. First, the duties underneath the guidance are not California inventions. Competence, confidentiality, and supervision of nonlawyer assistants are near-universal across the Model Rules and their state adaptations, and the ABA's own AI guidance reaches the same place: a lawyer's existing duties apply to whatever the tool is. The California guidance is among the most developed articulations of how those shared duties handle an agent, and it is likely to be cited well beyond California. Second, the technology does not respect state lines. The agent your firm runs behaves the same in Texas as in California. When a disciplinary body in another state confronts an agentic-AI failure for the first time, the California guidance is a document it may well reach for, because it is one of the first to do the thinking.
What AI does not replace
The whole framework comes back to one idea: autonomy is a property of the tool, not a transfer of responsibility. An agent can carry enormous volume, run overnight, and chain together steps a person would find tedious. It cannot hold a bar license, owe a client a duty of loyalty, or stand behind a filing in front of a judge. It cannot decide whether an action is too consequential to take without a human in the loop, because it has no concept of consequence. Those judgments are exactly what Rules 1.1, 1.6, and 5.3 reserve to a lawyer, and the 2026 guidance is a long way of saying so out loud. Used well, with the tiers set and the supervising lawyer named, agentic AI gives a small team the reach of a large one. Used as an excuse to stop supervising, it just produces consequences faster, with your bar number on them.
- California's verify-every-output proposed rule. The proposed-amendment track and the client-disclosure duty that sit alongside this guidance.
- The AI competence-duty checklist. The general Rule 1.1 competence baseline this briefing builds on.
- The confidentiality guide for attorneys. How Rule 1.6 applies to AI tools before you ever reach agentic systems.
Frequently asked questions
What did the State Bar of California approve on May 14, 2026?
The Board of Trustees approved an updated Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, replacing the version published November 16, 2023. It is the first version to address agentic AI, included at the California Supreme Court's request. It is interpretive guidance, not a black-letter rule, but it explains how a lawyer's existing duties of competence, confidentiality, communication, and candor apply to AI tools that act on their own.
Does the guidance ban agentic AI in legal practice?
No. The guidance does not prohibit agentic tools. Its principle is that the greater the autonomy you give an AI system, the more supervisory control and verification you must apply. It directs that substantive legal determinations, legal advice, preparing and filing pleadings, and acting in a representative capacity stay under meaningful lawyer supervision and review, while lower-stakes tasks can run with less. You can use agentic AI as long as you can show how you control it.
How does Rule 5.3 apply to an AI agent?
Rule 5.3 governs a lawyer's duty to instruct and supervise nonlawyer assistants. The 2026 guidance and the related proposed amendment extend that duty to the technology those assistants use, and the logic reaches the agent itself. An agentic tool that drafts and acts is functionally a nonlawyer assistant working under your authority. You are responsible for instructing it, configuring its guardrails, and reviewing its work, the same way you would supervise a junior paralegal. The responsibility stays with the supervising lawyer, not the tool or the vendor.
Is this guidance binding on lawyers outside California?
Not directly. The Practical Guidance is issued for California lawyers and is interpretive rather than a binding disciplinary rule even there. But the duties underneath it, competence, confidentiality, and supervision of nonlawyer assistants, are near-universal across the Model Rules and their state adaptations, and the ABA reaches the same conclusion that existing duties apply to AI. The California guidance is among the most developed treatments of how those shared duties handle an autonomous tool, so it is likely to be cited and followed well beyond California.
What is the minimum a small firm needs to do now?
Four steps cover the core. Classify every agent action into auto-allowed, notify-and-proceed, or hard-gated tiers and keep filings and advice in the gated tier. Run agents in a no-training, matter-scoped workspace so client data cannot leak across matters. Name a supervising lawyer for each agentic workflow. And keep a one-line supervision-of-record note whenever an agent takes a reviewed action. Together they apply Rules 1.1, 1.6, and 5.3 to agentic AI on a few pages.
Sources: State Bar of California, "Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law" (2026 update, approved by the Board of Trustees May 14, 2026; originally published November 16, 2023), and Ethics and Technology Resources (https://www.calbar.ca.gov/legal-professionals/ethics-compliance-practice-resources/ethics/ethics-technology-resources). State Bar of California, "Proposed Amendments to the Rules of Professional Conduct Related to Artificial Intelligence" (COPRAC, approved March 13, 2026; public comment closed May 2026). LawSites (Robert Ambrogi), "California Bar Proposes Rule Requiring Lawyers to Verify Every AI Output," May 2026. ABA Journal, "State Bar of California proposes first AI-specific changes to ethics rules," May 2026. Guidance is interpretive and evolving; confirm against current State Bar publications and your own counsel before relying on this page.