California SB 947 Is Not a Robot Ban. It’s a $500 Human-Oversight Trade
California SB 947 just passed. The market commentary around it is already overpriced on emotion and underpriced on structure. The No Robo Bosses Act is not a robot ban. It is a settlement layer for employment decisions. It does not care about model architecture. It cares where final authority lives. An automated decision system can make a termination recommendation, but it cannot be the execution layer. A qualified human must independently verify. Each failure starts with a civil penalty floor of 500 dollars per violation. Then it gets worse. Add punitive damages, attorney fees, Labor Commissioner enforcement, and a private right of action, and this legislation looks less like technology policy and more like a liability vector attached to every HR workflow.
The bill is the reworked version of an earlier California AI employment bill that died, and it has been narrowed to the sharp end of the relationship. It prohibits firing or disciplining a worker solely through an automated decision system. It bans predictive analytics used to infer protected characteristics. It demands written notice before those systems are used in consequential decisions. And it explicitly addresses AI Agents: if an agent is authorized to make workplace decisions, the employer cannot use that delegation to escape responsibility. That is the real mechanism. A black box cannot fire anyone. There has to be a record, and there has to be a human who owns the outcome.
But the legislators did not define key terms. What does independent mean? What qualifies as verification? Is a single checkbox from a supervisor who never looked at the underlying data sufficient? SB 947 is deliberately technology-neutral. That neutrality has a cost. The standard will be written by court cases, not by compliance officers. Between now and July 1, 2027, every employer with an AI-driven people function is building inside an open legal question.
Let me put this into trading terms. This is a regulation that creates an arbitrage for planners and a trap for procrastinators. Imagine an organization that uses an automated system to run a five-thousand-person reduction in force. The expected efficiency gain from that decision is material. But if the system is later found to be operating without genuine human verification, the damage analysis starts with five thousand discrete decision events. At a five-hundred-dollar statutory floor, that is two and a half million dollars before any damages, fees, or punitive award. California plaintiffs already know how to turn process failure into a verdict. The upside of automation gets eaten by the downside of procedural invalidity.
The overlooked variable is not penalty size. It is proof quality. The first round of enforcement will not be decided by whether an employer had an AI policy. It will be decided by whether the employer can reconstruct who looked at which data, when, and why they overrode or accepted the model output. That is an audit trail problem. Most companies do not have that infrastructure today. They have a chat log, a score, and a manager who vaguely remembers the case. That is not evidence. That is a liability.
I have seen this exact mistake in other code. In 2022, I audited a staking contract that had a clear integer overflow. The team treated the audit as a ritual, not as a go-no-go control. They launched anyway and lost millions. HR AI compliance will follow the same pattern. If the documentation layer is bolted on after deployment, the error is already embedded. Logs cannot resurrect a missing conversation. Decision provenance is not an optional compliance luxury. It is the product. The smartest employers will treat the next twenty-one months as a redesign cycle, not a waiting period.
Now here is the view nobody wants to hear. This bill may be bad for workers, not because it is weak, but because it persuades everyone that human oversight is meaningful. Independent human verification is an attractive phrase. In practice, a supervisor who receives a model-generated recommendation, has three other tasks, and is rewarded for processing cases will approve the output. That is not independent verification. That is approval theater. And it is worse than no human at all because a plaintiff must fight both a stubborn algorithm and the person who signed it.
The bill’s entire alignment claim rests on the assumption that the human in the loop is not a source of systemic failure. That assumption is false. I have traded with automated agents long enough to know that the most dangerous latency is not model inference. It is human ego inserting itself after the calculation. Ego is the ultimate systemic risk. The same ego that makes a manager trust their gut over a defensible process is about to become a discovery exhibit. All the hand-wringing about untrained supervisors eventually turns into legal data. Chaos is data waiting to be quantified, and the first round of lawsuits will quantify exactly how little independent verification actually took place.
That is the bear case for AI autonomy and the bull case for probative evidence. The companies building deterministic decision logs, human review workflows, and explicit override rationales are not just checking a box. They are building the settlement infrastructure for an entire labor market. The companies that keep selling black-box agents with a human approval button will be shorted by reality.
Whether Newsom signs the bill or sends it back, the direction is set. A veto only pushes the same structure into another version. California has already told the market where employment AI is heading: the firing circuit must contain a human who can be named and held accountable. The smart AI stack will no longer be marketed as autonomous. It will be marketed as defensible. It will log the model version, the feature inputs, the human review, the override rationale, and the written notice. That is the new competitive frontier. Liquidity vanishes. Conviction remains. The question any buyer should ask from now until July 2027 is simple: if your model recommends termination, where is the kill switch? If the answer is a product feature, move on. The only correct answer is a named human.