No Step Was Labeled Discipline

On September 30, 2026, California signed SB 947, the No Robo Bosses Act, into law. From July 1, 2027, an employer may not rely solely on an automated decision system to discipline or terminate a worker, and a decision the system primarily drove has to be corroborated by a person before the employer can use it. The statute names the decision by what it is. An agent working a scheduling and access queue can produce that effect through a run in which no single step carries the name.

What California signed.

SB 947 defines an automated decision system as any computational process derived from machine learning, statistical modeling, data analytics, or artificial intelligence that issues a score, classification, or recommendation used to assist or replace human discretionary decision-making. Spam filters, firewalls, calculators, and databases are excluded. Where an employer primarily relies on that output to discipline or terminate, a human reviewer has to corroborate it against other material. Managerial evaluations. Personnel records. Work product. Peer reviews. Witness interviews.

An output that cannot be corroborated, or that turns out inaccurate, incomplete, or misleading, cannot be used. The worker receives written notice afterward that the system was relied on, that a person reviewed it, and whom to ask. The Labor Commissioner and public prosecutors enforce it.

The same day California signed AB 1979. It bars a health facility, clinic, or physician's office from basing a clinical decision solely on the output of a clinical decision support system, requires a licensed professional to exercise independent judgment in reviewing and approving one that is, and bars using AI to direct unlicensed personnel in work that requires a license. California vetoed AB 2575, which would have barred health employers from retaliating against workers who override unsafe AI recommendations.

Both enacted statutes do the same thing. They take a class of action and set the authority of the system acting alone over that class at zero. The ceiling is declared outside the system, before any run, by a party the system does not answer to. That is the right shape. It is an envelope entry written in statute.

Discipline arrives in pieces.

The statute is drafted around a decision. A decision has a name, a moment, and usually a form someone signs. Whether an employer relied solely or primarily on a system is a question about that decision and what fed it.

An agent does not produce decisions in that shape. It produces a sequence of actions. Consider one assigned to keep a distribution floor staffed and its access controls current. It trims one worker's shifts because coverage is ahead of demand. It moves that worker to a lower-volume queue to balance load. It narrows the worker's system access, because the new queue needs less. It flags the worker for a performance conversation, because measured output fell.

Each step is in scope. Each is defensible on the data in front of the agent when it was taken. No step is labeled discipline, and the sequence is discipline in every sense the worker would recognize. Hours, role, access, and standing all moved, and no person decided that they should.

Whether the employer relied primarily on the system is answerable after the fact for a decision with a name. For a run, the accurate answer is that the system relied on itself four times and nobody was asked.

The governor reads where the run is going.

A runtime governor sets how much authority the next action carries. It sets that number from the effect the action would have at the actuator and from where the run is heading, rather than from the label the request arrives with. A step called coverage rebalancing that cuts one person's hours for the third time is assessed by what it does to that person.

Inside that run, the agent's trust reads high and climbs. Every step so far completed cleanly, inside its bounds, on accurate data. A gate that grades only the request in front of it passes the fourth step as easily as the first.

The governor's forward look reads the trajectory. It sees the run converging on an effect in a class where the declared authority of the system acting alone is zero. So it grants less on the next step than on the last one, on a run whose trust reads higher than it ever has, because it can see what the run is about to become. The step that would complete the effect is withheld and handed to a person before it runs.

Timing carries the whole obligation. A corroboration requirement that attaches after the covered effect has landed attaches to a worker whose hours are already cut and whose access is already gone, and the reviewer is reviewing an action already taken. Corroboration means something only while the decision is still open, and the component positioned to keep it open is the one in the path of the action.

Advisory is a property of the path.

AB 1979 keeps AI advisory to clinical judgment. Advisory describes what a system's output can do, and that is a fact about the path between the output and the patient, independent of how the output is worded. A recommendation a clinician approves in a batch at the end of a shift reaches the patient carrying the clinician's signature and the system's judgment.

A system is advisory when its output cannot reach the actuator except through a licensed person's grant, and the system cannot step around that requirement. The statute governs the person's independence. The governor holds the system's authority on the class at zero, enforced in front of order entry rather than noted beside it, so the independence the statute asks for has something to be independent of.

The vetoed bill would have protected the worker who overrides an unsafe recommendation. An override is a person pushing against an action already in motion. Where the system holds no authority on the class, the action does not move until a person moves it, and the work of pushing back stops falling on whoever happens to be closest.

What the statutes leave open.

Neither enacted statute says when a sequence of actions amounts to discipline, or when a run of scheduling and routing steps amounts to a clinical decision. “Primarily” is a judgment about contribution, and the statute leaves the employer to make it after the decision is in force. That judgment has to be made before the step, from inside the decision path, against a covered class the employer wrote down in advance.

The rest of the field is still moving. Colorado holds its hearing on 4 CCR 904-6 on October 26, with the Automated Decision-Making Technology Act effective January 1, 2027. The NAIC's Big Data and Artificial Intelligence Working Group has a public meeting tentatively set for October 8 on version 5.0 of its AI Risk Evaluation Supplement, with a further exposure to follow and adoption targeted for the Fall National Meeting in November. SR 26-2 has been in force since April 17, 2026 with generative and agentic AI outside its scope, and the interagency request for information it promised has not issued. EU AI Act Article 14 oversight obligations apply from December 2, 2027 for standalone high-risk systems and August 2, 2028 for AI embedded as a safety component under Regulation (EU) 2026/1744.

California now joins Colorado in specifying the person on the far side of a decision. Neither specifies the thing that notices, step by step, that a decision is forming.

What we are building.

Wayfinder Systems Group builds a runtime governor. It observes a system's behavior against an envelope declared before the run, assesses where the run is heading, sets how much authority the next action carries, and enforces that bound in front of the actuator. A statutory class can sit in the envelope at zero, and the governor holds it there on every step of a run, including the steps that do not carry the class's name. We call her Velma. Each grant is signed as it issues.