A Fraction of a Decision

On August 31, 2026 the California Legislature approved SB 947, the No Robo Bosses Act of 2026, by 53 to 14 in the Assembly and 28 to 10 in the Senate. It bars an employer from relying solely on an automated decision system to discipline or terminate a worker. The bill is with the Governor, and its requirements become operative on July 1, 2027 if it is signed.

Solely is a word about how much authority the system carried on one decision. No legislature had written the bound that way before.

What the statute bounds.

The bill reaches any computational process derived from machine learning or statistical modeling that displaces human judgment in a consequential decision about a worker. It requires human review and independent corroboration before a recommendation for discipline or termination is acted on. It requires written notice to the worker afterward. The notice names the system, the data behind it, and a reviewer who can be reached. It carries a per-violation penalty and a private right of action, so the party who will test it in practice is the worker.

Most AI legislation bounds one of three things. What data may enter. What purpose the output may serve. Who has to be told. SB 947 bounds a fourth thing. How much of the decision the system was allowed to be.

A fraction of a decision.

Read literally, solely describes a share. The system supplied less than all of the decision, so some remainder came from somewhere else. The trouble is that a decision does not divide. A worker is terminated or is not. There is no ninety percent of a firing.

What divides is the authority behind it. Authority is a quantity a system can hold more or less of on a given action, and it is the quantity the word is reaching for.

Two arrangements satisfy the statute identically on paper. In the first, a reviewer reads the file, has access to what the system did not see, and reaches a different outcome often enough to matter. In the second, a reviewer receives a queue of recommendations and approves them at the rate they arrive. Both produce a human in the record. In the second, the system's effective authority over the outcome is total, and the statutory language is satisfied.

The bill anticipates this and asks for independent corroboration rather than review alone. Corroboration is the stronger word. It means a second determination reached from something other than the first system's output. What makes a determination independent is a property of the path the decision travelled, and a statute cannot inspect a path. It can only require that one exist.

The ceiling and the grant.

SB 947 sets its bound by category and sets it once. Discipline and termination carry the ceiling. Everything else carries none. That line is drawn before any particular case arrives, by a legislature that will never see one.

A category is a coarse instrument for reasons that have nothing to do with drafting quality. Inside discipline and termination sit cases that are not alike. One is a documented policy breach with a clean record behind it. Another is a worker whose performance signal has been drifting for six months for reasons that live entirely outside what the system measures. The statute treats both the same.

A governor works at the other end. It sets how much authority the next action carries, per action, at runtime, against an envelope declared before the run and outside it. It starts from withheld and grants only what it can assure. A statutory ceiling becomes the maximum that envelope will admit for that class of action, and the grant on any individual case sits at or below it.

The statute fixes a ceiling for a class. A governor sets the grant for the case.

Two kinds of forward look.

SB 947 also prohibits predictive behavior analysis. An employer may not run an automated system over a worker's data to forecast that worker's future conduct, intentions, or disposition.

That prohibition and a governor's forward look are easy to confuse. They are different objects and they point in opposite directions.

The banned forecast takes a person's data and predicts what the person will do. The prediction is then used to act on the person, ahead of anything the person has actually done. It extends the reach of the system into a future the worker has not arrived at.

A governor's forward look takes the governed system's own realized behavior and predicts where that behavior is heading relative to what the system was assured to do. Its subject is the system, not a person. Its consequence runs the other way. A governor can grant less authority on a decision whose present trust reads higher, because its forward look caught a divergence coming before the outcome landed. The forecast is used to make the system do less.

A rule against predicting people does not reach a control that predicts its own charge. Keeping the two apart is a design obligation, and it belongs to the builder, because from the outside both look like a model producing an estimate about the future.

What holds the bound.

If the bill is signed, employers have until July 1, 2027. The artifact most of them will produce is a policy stating that no automated system decides alone, and a review step added to the workflow that routes recommendations to a person.

A policy asserts the bound. A control holds it. The difference appears at the only place it can appear, which is the path between the recommendation and the effect. If the termination can be executed without the grant, the bound describes intended behavior rather than actual behavior, and the first case that goes wrong is the test of which one it was.

Four questions about that path answer themselves by inspection, and all four are fixed at build time. Whether the enforcement point sits on the only path to the effect. What the action carries when no grant arrives. Which direction an uncertain assessment rounds. Whether a grant already set can be widened from inside the run.

Five days earlier, on August 26, 2026, the Australian Securities and Investments Commission published its Corporate Plan for 2026-27 and wrote that AI use must not weaken accountability. Two jurisdictions, one week, the same shape. The obligation is named in public. The mechanism is left to whoever builds it.

What we are building.

Wayfinder Systems Group builds a runtime governance substrate. It sits above control and below intelligence. It does not retrain the model that produced the recommendation and it does not redesign the workflow around it. It observes what the system is doing. It assesses how far that has moved from what the system was assured to do. It modulates how much authority the next action carries, and it enforces that grant in the path, before the action reaches its effect. A signed record of what was granted and what was withheld comes out of it. We call her Velma.