California SB 947 Is on the Governor’s Desk. If He Says Nothing by September 30, It Becomes Law.

1. What SB 947 does and exactly where it stands today

California SB 947, the No Robo Bosses Act of 2026, would bar an employer from relying solely on an automated decision system to discipline or terminate a worker. It would require a human to corroborate the system output, a written post use notice to the affected employee, and access to the employee’s own data. It carries a $500 civil penalty per violation and, if enacted, becomes operative July 1, 2027.

The bill is not a proposal any more. The Assembly passed it on August 30, 2026 by a vote of 53 to 14, the Senate concurred in the Assembly amendments on August 31, 2026, and the enrolled version was produced on September 4, 2026. You can read the whole thing yourself in the enrolled text on the Legislature’s own site, and the vote counts are in the roll call record. As of the date of this article the bill history shows no action by the Governor.

That matters because the same author, Senator Jerry McNerney, ran an earlier version of this bill and lost. SB 7 cleared the Legislature in September 2025, was presented to the Governor on September 23, 2025, and was vetoed on October 13, 2025. The Legislature did not override. The veto was formally sustained on March 2, 2026.

So the honest framing for a small employer is this. A near identical bill was killed eleven months ago. The author narrowed it, ran it again, and it passed both chambers by wider margins than the first time. Whether it survives is a coin flip that resolves inside September. What is not a coin flip is the underlying direction, and that is the part worth planning around.

2. Why September 30 matters more than a signature

September 30, 2026 is the deadline because of a timing rule in the California Constitution, and the rule means that gubernatorial silence enacts the bill rather than killing it. Article IV, Section 10(b)(2) provides that any bill passed by the Legislature before September 1 of the second calendar year of the session, and in the possession of the Governor on or after September 1, that is not returned on or before September 30 of that year becomes a statute.

SB 947 passed on August 31, 2026. That is one day inside the September 1 cutoff. The enrolled bill went to the Governor in September. Every element of the rule is satisfied, which means there are three possible outcomes and only one of them stops the law.

Two of the three paths end with the same compliance date. That asymmetry is the reason a small employer should not wait for a headline. The planning question is not whether the Governor signs. It is whether your discipline and termination process can survive a rule that took effect on July 1, 2027, and you have roughly nine months either way.

3. Who SB 947 covers, and why there is no small business exemption

SB 947 would cover every employer in California with no employee count threshold, no revenue threshold and no phase in for small businesses, subject to two narrow exemptions covered at the end of this section. The enrolled text defines an employer as a person who directly or indirectly, or through an agent or any other person, employs or exercises control over the wages, hours or working conditions of any person. That definition sweeps in a five person shop on the same terms as a five thousand person one. The employment law firm CDF makes the same reading in its September 3, 2026 analysis of the bill as sent to the Governor, noting that it applies to public and private employers regardless of size.

The definition reaches further than a payroll headcount in two ways that matter to staffing firms and to companies that use them. It expressly includes a labor contractor of a covered employer, and it expressly includes farm labor contractors and foreign labor contractors. If you place workers at client sites and your platform scores those workers, you are the employer for purposes of this part.

The realistic exposure is larger than most owners assume, because the underlying behavior is already common. Resume Builder surveyed 1,342 United States managers with direct reports and found that among managers using AI in people management, 78 percent used it for raises, 77 percent for promotions, 66 percent for layoffs and 64 percent for terminations. More than one in five said they let the system make the call with no human input at all. Those findings were reported by HR Dive in June 2025. SB 947 is aimed squarely at that last group.

There are two real exemptions and both are narrow. Section 1526.5 removes parties covered by a collective bargaining agreement, but only where the agreement explicitly waives the part in clear and unambiguous terms, expressly provides for wages and working conditions, and provides protection from algorithmic management. Section 1526.6 removes uses required by or reasonably necessary to comply with federal law or a binding federal contract relating to aircraft for the national airspace or products and services for national security, military, space or defense purposes, and only for those covered operations. Neither is available to an ordinary small employer.

4. What counts as an automated decision system

An automated decision system under SB 947 is any computational process derived from machine learning, statistical modeling, data analytics or artificial intelligence that issues a simplified output, including a score, classification or recommendation, that is used to assist or replace human discretionary decisionmaking and materially impacts natural persons. Read that once more and notice what it does not require. It does not require a large language model. It does not require the vendor to call the product AI. Statistical modeling and data analytics are enough.

The bill then carves out a short exclusion list: spam email filters, firewalls, antivirus software, identity and access management tools, calculators, databases, datasets and other compilations of data. That is the whole list. It is an infrastructure exclusion, not a productivity software exclusion.

The tools most small employers forget they have

The gap between what an owner thinks of as AI and what the statutory definition reaches is where the compliance risk lives. Scheduling and workforce management platforms that generate attendance reliability scores are covered when those scores feed a write up. Productivity monitoring that flags idle time and produces a ranked list is covered. Customer review aggregation that produces a service score used in performance management is covered, and SB 947 says so directly: the corroborating information a human reviewer may use expressly includes witness interviews, which may include relevant online customer reviews.

Note that this cuts both ways. Customer ratings were the specific example Governor Newsom raised when he vetoed SB 7, and the current bill answers that objection by naming customer reviews as legitimate corroboration rather than banning their use.

5. The two operative duties: no sole reliance, and human corroboration

SB 947 creates two distinct duties at Section 1522, and small employers routinely collapse them into one. The first is a flat prohibition: an employer shall not rely solely on an automated decision system when making a disciplinary or termination decision. The second is a process requirement that attaches at a lower threshold: if an employer primarily relies upon system output to make that decision, the employer shall direct a human to corroborate the decision.

Solely and primarily are not the same word. You can violate the second duty while satisfying the first. A manager who glances at an attendance score, agrees with it, and issues a final warning has not relied solely on the system, but has done no corroboration and creates real risk that the employer will be found to have relied primarily on it. Exactly where that line sits will be worked out through Labor Commissioner enforcement and litigation, because the bill does not define primarily.

The corroboration standard is specific about what counts. The human must corroborate using either the data that was collected or used to produce the output, or other relevant supporting information. The bill lists supervisory or managerial evaluations, personnel files, employee work product, peer reviews and witness interviews. This is a documentary standard, not an attitude. A reviewer who agrees with the score without looking at anything else has not corroborated it.

Then comes the provision with the sharpest operational teeth. Under Section 1522(c), if the employer cannot corroborate the output, or if the human reviewer concludes the output is inaccurate, incomplete or misleading, the employer shall not use that output to make the decision at all. The reviewer is not a rubber stamp with a veto they are discouraged from using. A reviewer who finds the flag unreliable removes it from the decision.

Section 1522(a) adds three flat prohibitions that sit on top of all of this. An employer shall not use an automated decision system to prevent compliance with or violate any federal, state or local labor, occupational health and safety, employment or civil rights law, to infer an employee’s protected status under Government Code Section 12940, or to predict and take adverse action against a worker for exercising their legal rights.

Read the trigger on those three carefully, because it is different. The two duties above attach only when a disciplinary or termination decision is being made. The Section 1522(a) prohibitions are written as freestanding limits on using an automated decision system at all, with no decision type qualifier. Treating SB 947 as a rule that touches nothing but your termination workflow understates what it would actually require.

6. The post use notice and the employee data request right

The notice under SB 947 is a post use notice, delivered at the time the employer informs the employee of the decision, and this single design change is the clearest answer to the veto that killed SB 7. Governor Newsom’s veto message objected to what he called “unfocused notification requirements” reaching businesses using the most routine digital tools, as Fisher Phillips reported in October 2025. You can read the message yourself in the veto document posted by the Governor’s office. SB 947 drops advance notice entirely. Nothing is owed until the employer actually disciplines or terminates someone in primary reliance on a system.

Section 1524 sets four requirements for how the notice is delivered. It must be in plain language as a separate, stand alone communication. It must be in the language the employer uses for routine communications with employees. It must go out through a simple and easy to use method, including email, a hyperlink or another written format. And it must arrive at the time the decision is communicated, not afterward.

The notice content is a fixed list of four items:

• That the employer primarily relied upon an automated decision system to make the disciplinary or termination decision.
• That a human reviewed the decision and corroborated the system output.
• Contact information for a human the employee may contact for more information about the decision and about the right to access a description of their own data.
• That the employer is prohibited from retaliating against the employee for exercising rights under this part.

Separately, Section 1522(d) gives the employee a right to request, and obliges the employer to provide, a meaningful and objective description of the employee’s own data used by the system. Section 1522(e) requires that the description anonymize the personal information of customers, other employees and other individuals. In practice this means a small employer needs to know, before the request arrives, whether its vendor can produce a per employee data description at all. That is a contract question and a vendor capability question, and it is far cheaper to ask at renewal than after a termination.

One piece of relief is built in. Section 1526.3 provides that an employer who complies with the notice requirements of this part is not required to comply with substantially similar automated decision system notice provisions under other state law. That relief expressly does not extend to quotas as defined in Labor Code Section 2100, the warehouse distribution center quota statute, or to other automated standards applied to working conditions. Section 1526.4 also preserves the California Privacy Protection Agency’s automated decisionmaking rules for businesses covered by the CCPA.

7. Enforcement, the burden shift, and the $500 penalty

The civil penalty is $500 per violation, and that statutory figure is only one component of the exposure, because notice, data access and retaliation counts can accumulate separately across workers and decisions. The provision that should change how you document decisions is the burden shift at Section 1526.1(c). Its operative language is worth quoting exactly: in a civil action or administrative proceeding brought for a violation of the sole reliance or corroboration duties, “once it has been demonstrated that an ADS was used to make a disciplinary or deactivation decision, the employer must demonstrate that the employer did not primarily rely upon an ADS or that the employer complied with Sections 1522 and 1524 when making the disciplinary or termination decision.”

That sentence moves the evidentiary weight onto the employer at the point where most small employers have nothing written down. The worker’s showing is that a system was used. Your showing is everything else. An employer with a dated corroboration record, a copy of the notice and the reviewer’s notes meets that burden in an afternoon. An employer with a manager’s recollection does not.

Three enforcement routes exist and they are not mutually exclusive. The Labor Commissioner may investigate, order temporary relief to maintain the status quo pending a full investigation, and issue citations using the existing procedures at Labor Code Sections 98.3, 98.7, 98.74 and 1197.1. A public prosecutor may bring a civil enforcement action. And in a civil action, a petitioner may seek temporary or preliminary injunctive relief, punitive damages and reasonable attorney fees and costs.

Attorney fees are the practical multiplier. A $500 penalty does not fund litigation. A fee shifting provision does, and it is the reason a single contested termination can become an expensive matter for a company with twelve employees. The bill authorizes punitive damages but does not displace the ordinary California standard for awarding them, so expect availability in a given case to be argued rather than assumed. Section 1526 also bars discharging, threatening to discharge, demoting, suspending or otherwise retaliating against an employee for using rights under the part, filing a complaint with the Labor Commissioner or cooperating in an investigation.

One more provision worth flagging for multi jurisdiction employers: Section 1526.2 states that the part does not preempt any city, county or city and county ordinance that provides equal or greater protection. Local rules stack on top rather than being displaced.

8. What already applies in California right now

Two sets of California rules already govern automated decision systems in employment, and both are in force or already scheduled regardless of what happens to SB 947. Treating September 30 as the start of your obligations gets the timeline backwards by roughly a year.

The first is the Civil Rights Council’s employment regulations on automated decision systems, which took effect October 1, 2025 and apply under the Fair Employment and Housing Act to employers with five or more employees. The final regulatory text is published by the Civil Rights Department. These rules make clear that using an automated decision system to make employment decisions can violate FEHA where it produces discrimination, directly or by disparate impact, on the basis of a protected characteristic. They also amend the FEHA recordkeeping rule at 2 CCR Section 11013 from two years to four and add automated decision system data to the records that must be preserved, running from the date the record was made or the date of the personnel action, whichever is later. The same four year duty reaches anyone who sells or provides an automated decision system to an employer. And they treat certain assessments that elicit information about a disability as potentially unlawful medical inquiries.

The second is the California Privacy Protection Agency’s automated decisionmaking technology framework. The CCPA updates, risk assessment, cybersecurity audit and ADMT regulations were approved by the Office of Administrative Law on September 22, 2025 and became effective January 1, 2026. Businesses that use automated decisionmaking technology to make a significant decision, a category that expressly includes decisions about employment, must comply with the ADMT article no later than January 1, 2027 under Section 7200(b). The timing rules are worth reading precisely rather than collapsing into one date. A processing activity that was already underway when the regulations took effect and that continues must have a documented risk assessment by December 31, 2027, and the first submission to the agency, covering risk assessments conducted in 2026 and 2027, is due by April 1, 2028. Consumers, which in this context includes employees of covered businesses, get a right to notice, a right to opt out in defined circumstances and a right to meaningful information about how the technology functioned.

Read together, the three layers answer different questions. The Civil Rights Council rules ask whether your tool discriminates. The CPPA rules ask whether you told people and gave them choices. SB 947 asks whether a human actually looked. A small employer can fail the third while passing the first two.

9. How SB 947 compares to Illinois, Colorado and New York City

SB 947 is narrower in subject matter than the other major state frameworks but deeper in process, because it regulates only discipline and termination and then dictates exactly how those decisions must be made. The comparison matters for any employer with workers in more than one state, and it matters for staffing agencies placing candidates across state lines.

Illinois moved first, and its law is broader in subject matter than California’s, reaching the full employment lifecycle rather than just discipline and discharge. Public Act 103-0804 amended the Illinois Human Rights Act at 775 ILCS 5/2-102 and took effect January 1, 2026. It makes it a civil rights violation for an employer to use artificial intelligence that has the effect of subjecting employees to discrimination on the basis of a protected class with respect to recruitment, hiring, promotion, renewal of employment, selection for training or apprenticeship, discharge, discipline, tenure or the terms, privileges or conditions of employment. It separately bars using zip codes as a proxy for a protected class, and it requires notice to an employee that the employer is using artificial intelligence for those purposes. The Illinois definition of employer reaches any person employing one or more employees in Illinois during 20 or more calendar weeks, which is about as small as employment thresholds get.

Colorado went the other way, toward a broad consequential decision framework. As we covered in our analysis of what actually applies to employers in Colorado on January 1, 2027, the original Colorado AI Act never took effect and was replaced by SB 26-189, signed May 14, 2026 and effective January 1, 2027, which covers automated decisionmaking technology across education, employment, housing, financial or lending services, insurance, health care and essential government services.

New York City is the outlier in the other direction: narrow subject matter, heavy documentation. Local Law 144 requires a bias audit within one year before use of an automated employment decision tool, public disclosure of the audit results, and notice to candidates or employees. The Department of Consumer and Worker Protection began enforcement July 5, 2023. It does not reach every hiring tool everywhere. It applies where the tool meets the ordinance’s definition of an automated employment decision tool, meaning it substantially assists or replaces discretionary decision making, and where the role is a job in New York City.

The pattern across all four is worth naming. No jurisdiction is banning these tools. Every one of them is requiring the employer to be able to explain a decision after the fact. That is a records problem before it is a technology problem, which is the same conclusion we reached in our review of the first state AI audit law.

10. A 90 day plan for a small employer

A small employer can reach a defensible position on SB 947 in 90 days without hiring anyone, because the work is inventory, process and documentation rather than technology. The plan below assumes an organization under 200 employees with no dedicated compliance staff, and it works whether or not the Governor signs, because the Civil Rights Council rules already apply and the CPPA obligations arrive January 1, 2027.

Days 1 to 30: find the systems

List every tool that produces a score, ranking, flag, classification or recommendation about a worker. Go past the HR system. Include scheduling, time and attendance, productivity monitoring, quality scoring, customer review aggregation, sales performance dashboards and anything in your CRM or help desk that ranks staff. For each one, record what it outputs, who sees the output, and whether that output has ever appeared in a performance conversation. Ask each vendor two written questions: can you produce a description of one named employee’s data used by the system, and what documentation do you provide about how the score is generated.

Days 31 to 60: fix the decision process

Write a one page corroboration standard covering what a reviewer must look at before acting on a system flag, drawing on the categories the bill names: supervisory evaluations, personnel files, work product, peer reviews and witness interviews. Add a required field to your discipline and termination documentation that records whether a system output contributed, which one, what the human reviewed, and what the reviewer concluded. Then decide the harder question in advance: what your managers do when the reviewer finds the output inaccurate, incomplete or misleading, because SB 947 requires that the output be dropped from the decision entirely.

Days 61 to 90: build the paper and train the humans

Draft the post use notice template now, with all four required elements and a named human contact, and have it translated into whatever languages you use for routine employee communications. Set your retention practice to hold automated decision system data for at least four years from the date the record was made or the date of the personnel action, whichever is later, which is what the amended Civil Rights Council recordkeeping rule now requires. Confirm the specific record types with counsel, because litigation holds and wage and hour rules can run longer. Train every manager who can issue a write up on one point above all others: under the burden shifting provision, the company has to prove what the human did, so an undocumented review is functionally the same as no review.

None of this requires you to stop using the tools. It requires you to be able to show your work.

This article is provided for general informational purposes and reflects the state of the law as of September 2026. It is not legal advice. Regulations in this area are changing quickly. Confirm current requirements and consult qualified counsel before making decisions for your organization.