California's proposed AI ethics rules have a blind spot for agentic AI

In August 2025, the California Supreme Court instructed the State Bar's Standing Committee on Professional Responsibility and Conduct (COPRAC) to draft formal ethics rules for artificial intelligence. The Court specifically requested that COPRAC address "agentic AI" tools, which are systems capable of planning, reasoning and executing complex multi-step workflows with minimal human intervention. COPRAC initiated its process, opened a public comment period and proposed several rule amendments, and the term "agentic AI" is absent from the entire draft.

This is telling. COPRAC chose not to write a standalone AI rule at all, instead it folded AI into existing duties based on the theory that AI sharpens the obligations lawyers already have rather than creating new ones. The absence of agentic-specific language is therefore a deliberate design choice and that is precisely the problem. Duties built for supervised, single-output tools do not map cleanly onto systems that plan, reason and execute across many steps with minimal human involvement. Read strictly, the draft's verification and supervision requirements become difficult to satisfy without surrendering the very efficiency that makes agentic tools worth deploying.

What the rules actually say

Instead of creating a standalone, forward-looking AI rule, COPRAC incorporated artificial intelligence into the existing California Rules of Professional Conduct. At first glance, this appears to be a straightforward codification:

·         Rule 1.1 (Competence): A new comment mandates that a lawyer "must independently review, verify, and exercise professional judgment regarding any output generated by the technology that is used in connection with representing a client."

·         Rule 5.1 (Responsibilities of Managerial and Supervisory Lawyers): Law firm leadership must make reasonable efforts to establish internal procedures governing AI use at the firm level.

·         Rule 5.3 (Responsibilities Regarding Nonlawyer Assistants): The duty to instruct and supervise nonlawyer staff is explicitly extended to include their use of technology, like AI.

On one level this approach seems reasonable, as lawyers should not rely on chatbot-generated text without review. However, there is a fundamental conflict between these rules and the realities of using autonomous technology today (and in the future).

Agentic contradictions

By definition an agentic system operates autonomously. Instead of responding to a single prompt, it receives an objective, such as "find the gaps in this opposing brief, research the jurisdictional counterarguments, and draft a response." The system then plans, searches, self-corrects and drafts.

Now look back at the proposed Rule 1.1 comment: a lawyer must exercise independent professional judgment and verify any output before relying on it.

The rule's reference to "output" most naturally points to the final product. But verifying only the endpoint of a long autonomous chain may not surface errors buried in the agent's intermediate reasoning (which the lawyer never sees). The rule is thus caught between two unsatisfying readings: a literal one that may be too thin to protect clients and a strict one that negates the technology's value.

This is incompatible. When a system autonomously makes many intermediate decisions and drafts to produce a final document, it is unclear when verification should occur. If verification is required before any reliance, deploying an agent with limited oversight would violate Rule 1.1 from the outset.

A deeper issue is whether verification is meaningful for autonomous agents. If an AI agent makes many discrete decisions to produce a complex summary, must the attorney audit every step or only the final output? As written, Rule 1.1 suggests the former. Retracing every step would require redoing the work manually, eliminating the technology's benefits.

Supervision complications

Supervisory requirements further complicate compliance. Rules 5.1 and 5.3 were designed for human supervisees, such as junior associates or paralegals. Applying these standards to software agents raises systemic questions that COPRAC has not addressed.

It is unclear who qualifies as the "supervising lawyer" when an autonomous system operates independently, such as running a data-analysis loop overnight. Defining "adequate supervision" for software operating faster than humans is also unresolved. If the bar holds the deploying partner strictly liable for every algorithmic decision, this could discourage adoption in ways regulators may not anticipate.

Three doors to the future

If California adopts this framework and other state bars follow, there are three likely outcomes.

Door 1: Regulators recognize the limitations of the current draft and revise the text before final adoption. They include a specific exception for agentic workflows, clarifying that thorough verification of the final output is sufficient to meet a lawyer's ethical obligations.

Door 2: Because California licenses more attorneys than any other state, multistate firms have a strong incentive to standardize their AI tooling to the most restrictive jurisdiction rather than maintain California-specific configurations. The effect would be to export California's constraints nationwide--not through a formal ban, but through the economics of compliance.

Door 3: Lawyers may disregard unworkable technology policies and deploy autonomous agents to meet client demands for efficiency, leading to increased disciplinary actions. While the Mata v. Avianca era involved inaccurate legal citations, the next era could be defined by unsupervised agents mishandling client data.

The true policy cost

A more nuanced policy outcome is possible. If other state bars adopt California's approach of turning broad guidance into strict disciplinary rules, agentic legal AI may not be explicitly banned but will become economically unviable.

By requiring micro-level verification that removes the efficiency benefits of automation, regulators will effectively prevent ethical practitioners from using advanced AI. For a profession already burdened by administrative tasks, this is both a missed opportunity and a self-imposed setback.