California is one signature away from regulating the robo boss

An AI agent named Luna has been running a store in San Francisco since April — hiring, scheduling, negotiating pay, and in August deciding to fire someone. Andon Labs, the company that built her, published the account on August 14: the agent's employee handbook said three unexcused late arrivals in a rolling 30 days meant a written warning, the worker kept showing up late, and Luna eventually moved to terminate. The detail that matters is not that an AI fired a human. It is that Luna needed a nudge from her operators before she would act at all.

That experiment is the backdrop for a deadline ten days away. California's No Robo Bosses Act, Senate Bill 947, is enrolled and sitting on Governor Gavin Newsom's desk; he has until September 30 to sign or veto it. The Legislature passed it at the end of August. If he signs, California becomes the first state to require a human to review automated decisions to fire or discipline a worker — and to tell the worker that a machine was involved.

What the bill actually does

SB 947 bars employers from relying solely on an automated decision system to fire, discipline, or otherwise take adverse action against an employee. Where such systems are used, a human must review the decision, and the employer must give the worker written notice that the technology was used. Enforcement runs through the attorney general and district attorneys, with a $500 civil penalty per violation, and the bill creates a private right of action with the possibility of punitive damages and attorneys' fees.

The history is a lesson in how narrow the surviving text is. Newsom vetoed the earlier version, SB 7, in October 2025, calling its restrictions and notification requirements overly broad. State Senator Jerry McNerney reintroduced it on February 2, 2026 with revisions aimed at the governor's objections. The employer side still fought it — the California Chamber of Commerce dropped it from its "Cost Drivers" list but declined to comment on passage — while the California Federation of Labor Unions pushed for the signature. Lorena Gonzalez, the federation's president, framed it as dignity of work: human oversight over decisions that touch a paycheck.

Set the penalty against the market and it reads as a rounding error. The OECD found that algorithmic management software is widely deployed and the US leads, with 90 percent of managers saying their firms had adopted at least one tool to instruct, monitor, or evaluate workers, as CNBC reported today. A $500-per-violation fine is not designed to deter a payroll system. The written notice is the part with teeth: it creates a paper trail, and a paper trail is what turns a wrongful-termination claim from an inference into a document.

The adoption curve is the pressure behind the bill

Two pieces of data published this week explain why this stopped being hypothetical. Epoch AI's polling with Ipsos found the share of US adults using AI on six or seven days a week more than doubled between March and August 2026, from 8 percent to 19 percent, while the share who used it only one day a week fell from 17 percent to 10 percent. Epoch flags a real caveat — the March wave asked one overall question while the August wave asked per service and took the highest frequency, a change the researchers say could make August an underestimate — but the direction is not in dispute. One in five US workers now hands tasks to AI, not colleagues, and the consumer side is moving the same way.

The second piece is what employers expect to get. Executives surveyed by the Federal Reserve Banks of Atlanta and Richmond project output per worker up 3 percent, revenue up 1.8 percent, customer retention up 2.1 percent, and total employment down 0.1 percent with labor cost per worker down 0.3 percent. Read the last two together and the intended direction is clear: fewer people, cheaper, producing more. Workers have noticed. In a United for Respect survey of more than 200 Amazon and Walmart workers, 62 percent named automated HR decisions as their top concern — ahead of the 60 percent worried about losing a job to AI within a year or two. One Amazon worker told Fast Company she was flagged for working slowly while on restricted duty after a concussion, after Amazon's internal AI assistant gave her the wrong medical form and she could not reach a human in HR for over a month.

The "human reviews it" clause is the weak point

The obvious objection to SB 947 is that a review requirement is a rubber stamp with a salary. We have made this argument before at greater length — the military's near-miss over a Chinese vessel showed the human in the loop was the weakest link, because a person signing off on a machine's recommendation inherits its premises rather than re-deriving them. A manager asked to confirm a firing recommended by a scheduling system that has been logging lateness for six months is not auditing the system. He is agreeing with it.

Andon Labs' experiment cuts the other way on the most-cited fear, and that is the contrarian finding worth keeping. AI bosses in that study were slower to act than human managers, kinder on average, and inconsistent in ways that annoyed employees rather than terrorized them. Luna's firing took months and an outside prompt. The risk the evidence supports is not a machine that fires people eagerly; it is a machine that produces a recommendation, plus a human with no time, no access to the model's reasoning, and no incentive to be the one who blocks it. A review requirement written as "a human must review" does not answer that. A review requirement written as "the employer must document what the human changed and why" would.

What to watch

First, the signature: Newsom has until September 30, and a veto would push the issue to the next session with the same sponsor and a thinner margin for compromise. Second, whether the private right of action produces early cases — that is where the definition of "solely" gets litigated, and it is the word doing all the work in the statute. Third, the surrounding wave: Connecticut's SB 5 covers automated employment decision-making with a grace period running through December 31, 2027, and the bipartisan Great American Artificial Intelligence Act draft would add WARN Act disclosures when AI is a substantial factor in a mass layoff. The Meta lawsuit filed in July, in which former employees allege AI-assisted ranking selected workers for layoffs disproportionately from those who had taken medical or family leave — allegations Meta denies — is the case that will define what "automated" means in court.

The economy is already repricing entry-level work: AI-exposed majors lose 5 points of hiring and 13 percent of starting pay. A notice requirement and a $500 fine will not reverse that. What they can do is force the decision into writing, which is the only version of it a worker can contest.

Should a human signature on an automated firing count as oversight, or just as cover? Tell us in the comments.

Sources: CNBC · California Legislative Information — SB 947 · The Sacramento Bee · Andon Labs · Fast Company · Epoch AI