Supervising Agentic AI in Legal Practice: A Professional-Responsibility Framework
7/29/20265 min read
Supervising Agentic AI in Legal Practice: A Professional-Responsibility Framework
Generative AI answers a prompt. Agentic AI does more — it plans and executes multi-step tasks toward a goal with limited human prompting: pulling authorities, drafting a section, revising it, and moving to the next step on its own. That autonomy is what makes it useful, and it's also what creates a supervision problem most firms haven't worked through.
Here's the key point, and it's easy to get wrong: an agentic tool is not a "colleague" or an "assistant" you can hold responsible. It's technology you use — and California's professional-responsibility duties run to you, the lawyer, not to the software. The autonomy raises the stakes on duties you already have; it doesn't shift them onto the machine.
This is general information, not legal advice.
The duty runs to the lawyer, not the tool
It's tempting to think of an agent that researches and drafts as a fast, confident junior associate — but that analogy can mislead. Under California's Rules of Professional Conduct, the obligations attach to the humans who deploy the technology. Three rules do the real work here:
Rule 1.1 (competence) — you must understand a tool's capabilities, data sources, limitations, and risks before using it in client work, and you must independently review and verify its output. This is the primary anchor for agentic AI.
Rule 5.1 (managerial and supervisory lawyers) — lawyers in a managerial role are responsible for establishing internal policies and procedures governing how AI is used in the firm.
Rule 5.3 (nonlawyer assistants) — you must give your staff appropriate instruction and supervision on the ethical use of technology, including AI, in providing legal services.
Note the structure: Rule 5.3 governs supervising the people who use AI, not treating the AI itself as a supervised "assistant." The tool is a tool. The accountability is yours.
What California is doing about it
This isn't speculation — the State Bar is actively moving to make these duties explicit for AI.
On August 22, 2025, the California Supreme Court directed the State Bar to consider folding the principles of its 2023 Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law into the binding Rules of Professional Conduct, and specifically to address agentic AI tools that can plan and execute tasks with little or no human intervention.
In response, COPRAC (the State Bar's Committee on Professional Responsibility and Conduct) approved a package of comment-level amendments to six rules — 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 — for a public comment period at its March 13, 2026 meeting; that comment period closed May 4, 2026. An updated version of the Practical Guidance, replacing the 2023 edition and addressing agentic AI, went before the Board of Trustees at its May 14–15, 2026 meeting.
Two features are worth understanding:
These are clarifications, not a new AI rule. The amendments are comment-level; the black-letter rule text is largely preserved. The bar's own view is that AI sharpens existing duties rather than creating new ones. Read together, the proposals form an integrated framework: understand AI (1.1), supervise its use (5.1, 5.3), protect client data (1.6), communicate significant AI use to clients (1.4), and verify AI-generated authority before filing (3.3).
Agentic workflows are squarely in scope. The independent-verification duty in Rule 1.1 and the supervisory duties in Rules 5.1 and 5.3 are directly relevant to agentic tools. Commentators have noted that a lawyer who deploys an agent to research, draft, and revise a brief with limited oversight would fall squarely within these rules.
Status matters: these amendments are proposed, not adopted. The California Supreme Court holds ultimate authority over the Rules of Professional Conduct, and whether or when the changes take effect remains to be seen. But the underlying duties of competence and supervision are in force right now.
A practical supervision framework
Treat every agentic deployment as a use of technology you remain fully accountable for. Build these checkpoints in.
Before deployment (Rule 1.1)
Understand the tool: its capabilities, data sources, limitations, and failure modes.
Define the agent's lane in writing — which tasks, on which matters, with which data.
Decide where a human checkpoint is mandatory: anything filed, anything client-facing, anything dispositive.
During the workflow (Rule 1.1)
Keep a human decision point between the agent's output and any external use. Autonomy inside the workflow is fine; autonomy out the door is not.
Require the agent to surface checkable sources, so a person can independently verify them rather than trust a conclusion.
On output (Rules 1.1, 3.3)
A licensed attorney reviews and verifies before anything is filed or sent — and personally reads and confirms every citation.
Legal judgment — strategy, analysis, the final call — stays with the lawyer. Don't hand the practice of law to the agent.
Protecting client data (Rule 1.6)
Entering client information into an AI tool can itself be a disclosure of confidential information if there's a material risk the data could be accessed, retained, or used inconsistently with your confidentiality duty. Vet where the agent sends data before you point it at a client matter.
At the firm level (Rules 5.1, 5.3)
Put an AI-use and supervision policy in writing; managerial lawyers own this.
Train everyone who touches client work on the ethical use of these tools.
Log approved tools, permitted tasks, and data limits — and revisit as the tools change.
Communicating with clients (Rule 1.4)
Where AI use is significant enough to affect the scope, cost, or decision-making of a representation, give the client enough information to make an informed decision.
Frequently asked questions
Is an AI agent a "nonlawyer assistant" under Rule 5.3?
No — that framing is a common shorthand but it misstates the mechanism. California treats AI as technology that lawyers and staff use. Rule 5.3 governs your duty to instruct and supervise the people using it; the lawyer's own competence duty (Rule 1.1) governs the output.
Are California's new AI rules in effect?
No. As of 2026 they are proposed, comment-level amendments still in the rulemaking process, subject to the California Supreme Court's authority. Your existing duties of competence, confidentiality, candor, communication, and supervision apply now regardless.
Can I let an agent file something if I review it afterward?
Review before filing, not after. The responsible attorney must personally verify the content and every citation before submission.
What's the single most important control for agentic AI?
A mandatory human verification checkpoint before anything leaves the firm — output verified against primary sources by a licensed attorney.
This article is general information, not legal advice. Confirm the current status of the rules and guidance with the State Bar before relying on any specific provision.
Sources
State Bar of California, Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law (2023; updated 2026)
COPRAC proposed amendments to California Rules of Professional Conduct 1.1, 1.4, 1.6, 3.3, 5.1, and 5.3 (approved for comment March 13, 2026; comment period closed May 4, 2026)
California Supreme Court directive to the State Bar, August 22, 2025
California Rules of Professional Conduct 1.1, 1.4, 1.6, 5.1, 5.3