California signs the No Robo Bosses Act. AI-led firings now need a human check

California Governor Gavin Newsom signed SB 947, the No Robo Bosses Act, on September 30, 2026. The bill, by Sen. Jerry McNerney, bars employers from relying solely on an automated decision system (ADS) to discipline or fire someone, and requires a human to corroborate the call when the system is the primary basis for it. It takes effect July 1, 2027.

What the law requires before a firing

When an employer "primarily relies" on an ADS for a disciplinary or termination decision, SHRM reports that a human must independently examine and corroborate it with other information: managerial evaluations, personnel records, work product, peer reviews, or witness interviews. Kilpatrick's alert adds the data used to generate the output and relevant online customer reviews to that list.

The reviewer has to do more than rubber-stamp the result. Kilpatrick writes that the human must exercise independent judgment to verify the output before the decision is final. If the output cannot be corroborated, or turns out to be inaccurate, incomplete, or misleading, SHRM says the employer cannot base the discipline or firing on it.

The notice the worker gets

After the decision, the employer owes the worker a written notice. Per Kilpatrick, it must be a separate, stand-alone communication in the language the employer normally uses with that employee, simply worded and easy to access by email, hyperlink, or written document. It has to say the employer primarily relied on an ADS, and confirm that a human reviewed and verified the output.

The notice also gives the worker a person to contact, states that retaliation is prohibited, and explains what Jackson Lewis quotes as "the employee's right to request a meaningful and objective description of the employee data used." Before answering that request, the employer must anonymize personal information about other people, according to Kilpatrick.

What counts as an automated decision system

The definition is wide. Jackson Lewis quotes the law as covering "any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues simplified output." Kilpatrick describes that output as a score, ranking, classification, or recommendation used to assist or replace human discretion in an employment decision. Spam filters, firewalls, antivirus software, calculators, and datasets are excluded.

The law also bars using an ADS to infer an employee's protected status under the Fair Employment and Housing Act, per Jackson Lewis. The Labor Commissioner and public prosecutors enforce it, and Kilpatrick lists a civil penalty of $500 per violation, plus possible injunctive relief, punitive damages, and attorneys' fees.

Why this version got signed

Newsom vetoed McNerney's first attempt, SB 7, in October 2025. HR Dive reported that his veto message said the bill "imposes unfocused notification requirements on any business using even the most innocuous tools." When McNerney reintroduced it in February 2026, Quartz reports he dropped the advance-notice requirement and the language covering gig workers.

What is left is narrower: discipline and termination, with notice after the fact. "No worker should ever be fired or disciplined by a robo boss," McNerney said, per Quartz.

Why a build studio cares

Read SB 947 as a spec and it turns into a list of fields. An employer cannot send that notice, answer a data request, or contest a citation unless the software that scored the worker kept a record of how. Kilpatrick advises keeping contemporaneous records of what the ADS did, what the reviewer considered, and how and when the notice went out. A tool that stores only the final score cannot answer any of those questions. The bill does not list fields. Our reading is that, per decision, the record needs at least this:

decision_id
system:         name, model version, config hash
inputs:         which employee data fields, as of when
output:         raw score / ranking / classification / recommendation
reliance:       primary or supporting
reviewer:       named human, role, reviewed_at
corroboration:  evidence reviewed (evaluations, work product, interviews), verdict
notice:         sent_at, channel, language, contact person
data_request:   received_at, answered_at, other people's data anonymized

This is what we check. In a Deep Audit, the data-flow map traces which employee fields reach a scoring system, and we test whether the model version behind a past score can still be recovered. The AI workflows we build already ship with logs and a human approval gate for risky actions, and that gate is the hook SB 947 needs: record the inputs, the model version, the output, and the person who signed off. Adding provenance to a system that never stored it is the slow, expensive path, and July 1, 2027 is nine months out.

Next step: read the bill text on leginfo next to Kilpatrick's alert, then ask your HR software vendor which of the fields above it stores today. If you run or build a workforce tool that scores people and want that answer checked before July 2027, write to us at hello@gattyworks.com.