What does California SB 574 mean for AI in your firm?
From January 1, 2027, section 6068.1 of the Business and Professions Code says an attorney who uses generative AI shall not enter confidential, personal identifying or other nonpublic information into a generative AI system for which access to that information is not restricted to the attorney and persons the attorney authorizes under obligations to protect its confidentiality. This page covers the technical side: ways to restrict access, who checks the output, and what to ask your IT provider. Our entry step, AI agent setup and training, is $1,999 fixed.
IG Digital Lab is a technology provider, not a law firm or compliance auditor; confirm requirements with your counsel. This page is not legal advice.
What does section 6068.1 say?
Quoted from the chaptered bill on leginfo.legislature.ca.gov, version dated 09/30/26, checked on 2026-11-07.
(a)(2) “An attorney shall not delegate the practice of law to generative artificial intelligence.”
(a)(3) An attorney who uses generative artificial intelligence to assist in the practice of law “shall do all of the following”:
(a)(3)(A) “Not enter confidential, personal identifying, and other nonpublic information into a generative artificial intelligence system for which access to confidential, personal identifying, or other nonpublic information the attorney inputs into the system is not restricted to the attorney and persons authorized by the attorney under obligations to protect the confidentiality of the information.”
(a)(3)(B) “Take reasonable steps to” verify generative AI outputs, “including, but not limited to, the accuracy of all case and statutory citations,” and to correct “any erroneous or hallucinated output in any material used by the attorney.”
(a)(3)(C) “Disclose the use of generative artificial intelligence to the court for all documents submitted to the court” and consider whether to disclose it when it is used to create content provided to the public.
The bill is SB 574, Chapter 858, Statutes of 2026, approved by the Governor on September 30, 2026. Subdivision (b)(2) defines “personal identifying information” to include driver's license numbers, dates of birth, Social Security numbers, addresses and phone numbers of parties, victims, witnesses and court personnel, medical or psychiatric information, financial information and account numbers, and content sealed by court order. A client file usually holds several of those. Sources: bill text and bill history.
When does it start?
January 1, 2027. The bill has no urgency clause, and under Government Code section 9600(a) a statute enacted at a regular session takes effect on the January 1 that follows a 90-day period from enactment. Ninety days from September 30, 2026 ends on December 29, 2026. Law firm alerts give the same date, for example Holland & Knight, October 1, 2026.
What does section 6068.1 not require?
Read from the text of the section. How it applies to your practice is for your counsel.
- Its text does not mention servers, cloud services or named products, and contains no requirement to run AI on a firm's own server.
- It defines two terms, “generative artificial intelligence” and “personal identifying information.” It does not define “restricted” or “delegate.”
- It names no approved AI tools and sets up no certification for them.
The access limit in (a)(3)(A) is the part of the section that bears most on technical setup; (a)(2), (B) and (C) are duties of the lawyer. Where the AI runs is one factor in whether access is restricted. The vendor's terms, the accounts and the records of use are others. The State Bar's 2026 Practical Guidance points the same way: a lawyer should take reasonable steps to understand how a generative AI product “collects, uses, stores, and discloses information provided by the user,” which may include consulting qualified IT professionals or cybersecurity experts. It also says a lawyer must not deploy an agentic AI system in a manner that permits “autonomous external transmission of client information” without appropriate safeguards and human review. The Board of Trustees approved the 2026 revision on May 14, 2026 (State Bar ethics news).
How can a firm restrict access to client information?
These are technical measures that support a firm's own program. Which ones are enough is for your counsel.
| Measure | What it means in practice | What we set up |
|---|---|---|
| Where the system runs | On a server or office computer the firm owns or rents in its own name, not in a vendor's shared service. The firm holds the administrator account. The agents have full access to the machine they run on, so it should hold only what the firm chooses to put there. | We install the workspace on your machine and keep the AI plans on the firm's own accounts. |
| Access control | Named people only, a second sign-in factor, access removed the day someone leaves. Where the tool allows it, separate access for each lawyer and matter. | The workspace signs in with a password plus a code from the phone. Multi-factor sign-in and same-day removal across the firm's other systems are part of our IT support plans. Roles per person and per matter are a custom build. |
| Records of use | A record of who used the AI, when, and on which documents, kept for a stated time. | The workspace keeps a timeline of what the agents did in each project, on the server. It does not record which person asked. A per-person record of document access is part of a custom build. |
| No training on firm data | Written confirmation from the provider that the firm's inputs are not used to train a model on the plan you pay for. A local model removes the question. | We choose with you between a local model and an outside one, and write down which. |
| Vendor terms review | Who at the vendor can see the inputs, how long they are kept, where they are stored, and which other companies handle them. | Your counsel reviews the provider's terms for your plan. We list what the model receives, which provider and which plan handle it, and link that plan's current terms. |
| What an agent can send out | No automatic email, upload or call to an outside service without a person approving it. | With Claude as the engine, we turn on Ask mode: a person approves each step before it runs, including email and uploads. Codex and Grok have no per-step approval in the workspace, so we write down which steps they can take. |
Who checks what the AI writes?
The lawyer, under both the statute and the State Bar guidance.
Section 6068.1(a)(3)(B) requires reasonable steps to verify the accuracy of AI outputs, including all case and statutory citations. The same chapter adds a rule to Code of Civil Procedure section 128.7(b)(2)(A): a filing “shall not contain any citations that an attorney responsible for submitting the pleading has not personally verified, including any citation provided by generative artificial intelligence.” The State Bar guidance adds that lawyers must not permit AI systems to autonomously file documents, communicate with the court, or make representations on the lawyer's behalf.
With a document search build, a draft passage can link to the firm file it came from, so the lawyer sees the basis at once. That speeds up the check. It does not verify case law, so citations still go through a research database, checked by the attorney.
Questions to ask your IT provider before January 1, 2027
Take this list to whoever runs your computers. If they cannot answer, write that down too. For the setups we build, we answer the technical ones in writing before launch.
- Which AI tools does our staff use today? Include the ones built into software we already pay for, and say who approved each.
- Who can see what we type or upload? Our people only, or also the vendor's staff and contractors? Where is that written?
- Does the vendor keep our inputs? For how long, in which country, and can we delete them?
- Is anything we send used to train or improve a model? Is that off on our plan by default, or only on a higher plan?
- Does the tool run on a machine we control? If it runs in a vendor's cloud, which model and which company handle our text?
- How does each person sign in? With a second factor? Can you remove someone's access the day they leave?
- Can an AI tool read every matter folder, or only the matters a person works on?
- Is there a record of who used the tool, when and on which documents? How long is it kept?
- Can an AI agent send email, upload files or call outside services on its own? Which steps need a person's approval?
- Are the backups encrypted, and has a restore been tested this year?
- What are staff told never to paste into an AI tool? Where is that written, and who trained them?
- Which of these answers go to our counsel? Ask for them in writing, with dates.
Where does the $1,999 step fit?
AI agent setup and training is the entry point: one workspace on a machine the firm controls.
We prepare the machine
A server or office computer the firm owns, with updates, an encrypted vault for keys and an address in the firm's name.
We install and connect
The workspace, the models you choose and the services you approve, with a written note of what each model sees.
We test and train
We run the workflows on your own documents and teach your team to give tasks and to check what comes back.
$1,999 is a fixed price for the scope agreed before work starts. Hardware or hosting, AI plans and API usage, third-party services, custom development and ongoing support are separate. A setup that works fully offline is quoted separately. More on the law firm page, AI agent setup and training and IT support in Sacramento.
Sources
Primary texts first. All checked on 2026-11-07.
- SB 574, chaptered text, leginfo.legislature.ca.gov. Among other changes, adds Business and Professions Code section 6068.1 and amends Code of Civil Procedure section 128.7.
- SB 574, bill history: chaptered by the Secretary of State on 09/30/26 as Chapter 858, Statutes of 2026.
- Government Code section 9600: when a statute takes effect.
- Practical Guidance for the Use of Generative Artificial Intelligence in the Practice of Law, State Bar of California, Standing Committee on Professional Responsibility and Conduct, 2026 version, which replaces the November 2023 version.
- State Bar of California ethics news: Board of Trustees approval on May 14, 2026.
- Analyses by law firms, not primary sources: Holland & Knight, October 1, 2026 and Baker McKenzie, October 13, 2026.
Common questions
Short answers, with the limits of what a technology provider can say.
Does SB 574 ban cloud AI tools?
The text does not name cloud tools or ban any product. Section 6068.1(a)(3)(A) conditions the input of confidential, personal identifying and other nonpublic information on access being restricted to the attorney and persons the attorney authorizes under confidentiality obligations. Whether a cloud tool meets that depends on its terms and your account. Your counsel decides.
Does a two-lawyer firm fall under it?
Section 6068.1 is addressed to “an attorney” and the text sets no firm-size threshold. Ask your counsel how it applies to your practice.
Does my firm need its own server for AI under SB 574?
The text contains no such requirement. Running the workspace on a server the firm controls, with a local model for confidential material, is one technical way to limit who can reach what the attorney enters, and it is the setup we offer. When a task uses an outside model, the provider's terms for your plan decide who at the provider can reach the inputs. Whether any setup meets (a)(3)(A) is for your counsel.
Is this page legal advice?
No. IG Digital Lab is a technology provider, not a law firm or compliance auditor. This page quotes the primary texts and describes technical measures. Confirm what applies to your firm with your counsel.
Take the checklist to your IT provider, or ask us.
Tell us what your firm uses today. You get a free written review by email within two business days, and a fixed quote for the $1,999 setup if it fits.
Tell us what's slowing you down.
You get a free written review by email within two business days.