California's No Robo Bosses Act makes HR reviewers bring their own evidence before an AI-backed firing

Take an HR team whose discipline step today is a flagged case from a monitoring tool and an Approve button. From July 1, 2027, a click on that button is the thing SB 947 says does not count ^[2][6][4]. The reviewer has to arrive with material the system did not produce. Gizmodo's examples are manager evaluations and peer reviews ^[3].

Teams that run automated decision systems can tell themselves a human is in the loop because a person signs every case. The law, as both outlets describe it, separates signing from corroborating, and only corroborating meets the bar ^[3][4]. The state's labor federation backed it on those terms. "By signing these four first-in-the-nation bills, Governor Newsom has shown the country how we can begin to put guardrails on AI at work to ensure human oversight over this technology, limit surveillance, and put workers first," said Lorena Gonzalez, president of the California Federation of Labor Unions, AFL-CIO ^[7].

The duty turns on how heavily the employer leans on the system. The law bars relying on an automated system solely, and the corroboration requirement applies when the employer mainly relies on one ^[1][3]. The coverage does not define "mainly," and it does not describe a requirement to keep records of the review. I'd expect employers to file the corroborating evidence with each case anyway, since a review nobody wrote down is hard to show later, though on the reported text that is a practice choice.

Teams have time. Newsom signed the bill on September 30, 2026, about nine months before it takes effect ^[6][10].

He pitched the package as a state answer to Washington. "We are deeply mindful we have to do a lot more in the absence of federal leadership," Newsom said ^[9].

To sort the rebuild, put each firing or discipline workflow on two axes: whether the system's output mainly determines the outcome, and whether the reviewer brings evidence the system did not generate. A system that decides plus a reviewer who adds nothing is the case the law was written for, and it needs a new review step before July 2027 ^[3][4]. Swap in a reviewer who pulls manager evaluations or peer reviews and the same workflow matches the corroboration the law describes ^[3]. Workflows where people decide from their own evidence, with the system only informing them, carry less exposure. Check the advisory tool with a rubber-stamp reviewer hardest, because a reviewer who adds nothing can turn an advisory score into the main basis of the decision.

My recommendation is to require independent evidence at the review step everywhere AI touches discipline. The cost is reviewer time on every case, plus some of the speed that made the system attractive in the first place.