Legal technology governance

California's 2026 AI Guidance: A Firm Checklist

California's updated AI guidance addresses agentic systems and gives law firms a practical governance checklist for supervision, confidentiality, and review.

Four legal professionals reviewing an agentic AI workflow map with a highlighted human approval checkpoint in a modern law office.
Editorial illustration for this article.

Published 2026-08-17 · Updated 2026-08-17 · By The Hammer Lex Editorial Team

Quick answer

The State Bar of California approved updated practical guidance on May 14, 2026. It replaces the 2023 version and adds specific discussion of agentic artificial intelligence. The update matters beyond California because it shows how professional-responsibility duties are being turned into operational expectations for tools that can plan, act, and iterate with limited prompting.

For a law firm, the useful response is a workflow-level control system: inventory the tools already in use, classify the data and authority each tool can access, set a named human approver, preserve an audit trail, and define when an agent must stop. No single rule covers every lawyer. The guidance is a useful benchmark for designing a responsible AI program.

This article is informational and not legal advice. Professional-responsibility, confidentiality, billing, data-protection, and court-filing requirements vary by jurisdiction, client, and matter. Consult qualified counsel and the applicable bar or court authority before changing your firm's practices.

What California changed in 2026

The State Bar's updated Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law replaces the November 2023 version. The revision, the State Bar says, reflects newer developments, including agentic AI. It also clarifies how existing professional duties apply when lawyers use these technologies.

The shift underneath is the part worth understanding. A conventional drafting assistant waits for a person to ask for an output. An agentic system can be configured to break a goal into steps, call other tools, retrieve information, update a record, and keep iterating. The label counts for less than the capability. The system builds a chain of actions, and a chain is harder to review than one isolated response.

The State Bar's supporting materials describe agentic AI as automated programs that can plan and execute tasks and work toward a defined goal with minimal or no human intervention. The guidance itself still runs on familiar duties such as competence, confidentiality, communication, candor, supervision, professional judgment. The operating question gets sharper. Who approved the chain of actions, what could it reach, and what evidence shows the final work was reviewed?

California's move follows an August 2025 directive from the Supreme Court of California asking the State Bar to consider AI-related rule comments, including the use of agentic tools. The State Bar also published proposed amendments addressing competence, communication, confidentiality, candor toward the tribunal, managerial responsibilities, and supervision of nonlawyer assistants. The proposals and the practical guidance are related, but they are separate documents. Firms should track their status separately.

Why agentic AI changes the governance problem

Governance gets hard for a reason that has little to do with a bad sentence. A tool may make several individually plausible decisions whose combined result creates a professional or business risk.

Take an intake workflow that reads a web form, classifies the matter, searches a conflicts database, drafts a response, and creates a task for a lawyer. Each step looks administrative on its own. Together the workflow can affect who receives a response, what information is stored, whether a potential conflict is escalated, and what the firm implicitly promises to a prospective client.

Or a litigation assistant that retrieves authorities, summarizes a record, proposes citations, and inserts language into a draft. The document comes out looking polished. Underneath, the system may have used the wrong jurisdiction, missed a qualifying fact, or relied on a source no lawyer had checked. A review step at the end is necessary, and it may not be enough if nobody can reconstruct the inputs and intermediate actions.

The practical distinction is between output review and process control:

  • Output review asks whether the final text, classification, or task is accurate and appropriate.
  • Process control asks what the system was allowed to do, what information it touched, what it changed, and where a person had to approve the next step.

Small firms do not need an enterprise AI laboratory to implement both. They need a short, maintained record for each meaningful AI-assisted workflow.

A five-part checklist for law-firm leaders

1. Inventory capabilities, not product names

Start with the work people actually do. Ask attorneys, paralegals, intake staff, and administrators which tools can summarize, draft, search, classify, transcribe, send, update, or trigger another system. Include the AI features already embedded in practice-management, research, document, email, and meeting tools.

Record at least:

  • the workflow and business owner;
  • the tool, model, or feature involved;
  • whether the system only suggests or can take an action;
  • the information it receives and the systems it can reach;
  • the person responsible for approving the result; and
  • the date for the next vendor or workflow review.

Product-name inventories go stale quickly. A capability inventory stays useful when a vendor adds an agent mode or changes a default.

2. Define the boundary before enabling autonomy

Give every agent a written boundary. It should answer what the agent may read, what it may create, what it may change, and what it may never do without approval.

An intake agent, for example, might be allowed to normalize public web-form data and prepare a draft acknowledgment. It might be barred from deciding that the firm has accepted representation, from stating legal conclusions, from searching unrestricted client folders, and from sending a message that creates a deadline or commitment without human approval.

Use simple permission levels:

  1. Suggest: the system produces a draft or recommendation only.
  2. Prepare: the system may organize information or create a queued task, but a person approves the action.
  3. Execute with guardrails: the system may complete a narrow, reversible action within a defined scope.
  4. Prohibited: the system may not perform the action, even if a user asks informally.

Make autonomy deliberate and reviewable. A slow workflow is not the goal.

3. Put a human checkpoint where judgment is required

“A lawyer reviews the output” is too vague to work as a control. Name the decision that actually requires professional judgment, and assign it to a named role.

For legal research, the checkpoint may require verifying each authority against a primary source and confirming that the cited proposition is supported. For client communications, it may require checking scope, tone, facts, confidentiality, and whether the message could be read as advice or an engagement decision. For intake, it may require a lawyer or trained supervisor to review conflicts and urgency before the matter moves.

Put the checkpoint where the system can show it: a status, an approval field, a task, a stored review note. If the only evidence is someone's memory of having looked at a screen, the control is hard to test, train, or defend.

4. Preserve enough evidence to reconstruct the work

The firm does not have to preserve every prompt forever. It does have to decide what evidence fits the risk and the matter. A useful record might hold the tool version, date, user, source files, output, reviewer, corrections, and approval status.

Match retention to the firm's obligations and client agreements. Do not stand up a new repository of sensitive prompts without settling access controls, retention, export, and deletion first. The record is worth something only if it is protected and findable when a matter owner needs it.

This is where the AI policy and the practice-management system should meet. If the review record lives in a separate spreadsheet nobody checks, the policy describes an aspiration. It does not describe the firm's real control.

5. Test failure modes before expanding the workflow

Test agentic workflows against ordinary matters and against edge cases you pick on purpose. Feed them missing information and conflicting instructions. Try unusual names, duplicate records, stale authorities. Simulate a permission failure, a vendor outage, a user attempting to bypass the approval step.

For each test, write down the stop condition you expect. The agent should stop when it reaches an action outside its authority, when it cannot verify a source, when it detects an ambiguity that could affect the matter, or when it cannot protect the required information. A specific stop condition beats a general warning to “use caution.”

What this means for a small firm

Nothing in the State Bar guidance requires a small firm to deploy agents or buy a governance platform. It does suggest an order of operations:

First, find the unofficial use

Ask what staff already use, including consumer tools and features hidden inside approved software. A firm cannot assess confidentiality or supervision risk for a tool it does not know exists. Frame the conversation as discovery, not punishment. Otherwise people hide the most useful information.

Second, select one bounded workflow

Pick a workflow with a clear input, output, owner, and review step. Internal matter-status summarization or a draft task list from a meeting transcript makes a sound first pilot. Do not start with an autonomous client communication, a court filing, a conflicts decision, or a fee calculation.

Third, connect the approval to the matter record

Store the result, the reviewer, and the decision where the firm already manages the matter. That narrows the gap between the written policy and daily operations. It also makes it easier to answer a client or an internal reviewer who asks how a result was produced.

Fourth, measure both speed and reliability

Track turnaround time, correction rate, review completion, escalation frequency, and how long it takes to recover from an error. For a client-facing workflow, add client response time and rework. Faster output is not a success when it moves the repair work onto a lawyer.

Fifth, set a review date

Agentic tools change through vendor updates, new permissions, integrations, and model behavior. Revisit the workflow after a material product change, and on a regular calendar. The review date belongs to the workflow. It is not a one-time implementation task.

How to explain the change to clients

Clients may ask whether the firm uses AI, whether their information is used to train a model, who reviews the output, and whether the use changes fees or the quality-control process. A firm should be able to answer each of those questions accurately for the configuration it actually runs.

Answer in plain language, tied to the service, consistent with the engagement agreement and the duties that apply. Do not promise that AI is never used if embedded features run in the firm's systems. Do not describe a human review step that nobody reliably performs. A vague assurance costs more trust than a candid explanation of a bounded tool.

The ABA's Formal Opinion 512 remains a useful national reference for competence, confidentiality, communication, candor, and fees. California's 2026 guidance adds a current example of how those duties are being read as systems grow more autonomous. Neither source replaces the rules that apply in the jurisdiction and the matter in front of you.

The management takeaway

Agentic AI turns software governance into service governance. A firm no longer asks only whether a tool writes a good paragraph. It decides how much authority to delegate, what the system may know, which actions require approval, how the firm proves review occurred, and how it responds when the technology changes.

Attached to a real workflow, that problem stays manageable. Start with one process, one owner, one boundary, and one visible human checkpoint. The result is more useful than a broad promise to “use AI responsibly,” and it gives the firm a foundation for judging future tools with less guesswork.

For more operational context, read Post-AI Lawyering: What Firms Should Change Now and Legal Technology's Expanding Fault Lines.

Frequently asked questions

Does California's 2026 guidance apply to every law firm?

No. It is guidance from the State Bar of California and should not be treated as a single nationwide rule. Use it as a benchmark, and check the rules, opinions, court requirements, client obligations, and technology contracts that apply to your own work.

What makes an AI tool agentic?

An agentic tool does more than return one response to one prompt. It may plan a sequence, use connected tools, retrieve or change information, and keep working toward a goal with limited human prompting. The marketing label matters less than the system's real permissions and its ability to take action.

What is the best first agentic AI use case for a small firm?

Start with a bounded internal workflow: low-risk inputs, a reversible output, a named owner, a visible review step. Internal status summaries or draft task lists are usually easier to control than autonomous client messages, conflicts determinations, court filings, or fee decisions.

Does a human review at the end solve the risk?

Not by itself. The reviewer also needs to know what the system accessed, what it changed, and which claims or decisions require verification. Pair the final review with permission limits, stop conditions, and enough evidence to reconstruct the process.

Should a firm update its AI policy now?

It should review the policy if the firm uses tools that can act beyond a single draft or if the current policy does not describe actual usage. Focus the update on capabilities, data boundaries, human approval, records, vendor changes, and escalation. Have qualified counsel test the final policy against the firm's jurisdictional and client-specific obligations.

Source and context: This article is informational and is not legal advice. Verify current details with the linked source and qualified counsel where appropriate.

Sources

Make responsible automation visible

Connect people, matters, documents, deadlines, billing, and review steps so your firm can scale useful automation with accountability.

Book a demo

More industry articles · Free law-firm tools