There is a difference between a manager making a decision with software and a manager approving the decision the software has already made. On an organisation chart, both can look like human supervision. To the person facing discipline or dismissal, the difference can be everything.
California has moved to draw that line. Governor Gavin Newsom signed SB 947, the revised No Robo Bosses Act, on 30 September. The measure restricts sole reliance on automated decision systems in disciplinary and termination decisions. Its relevant provisions become operative on 1 July 2027—not on the day of the signing announcement. [1][2]
The immediate change is therefore a legal and organisational task ahead, not evidence that workplace decisions have already improved. Employers have to understand what the law covers. Workers need to know what the promised review is supposed to mean. And both have an interest in ensuring that the human in the process has more to do than click an approval button.
Assistance is not authority
The central restriction is not a blanket ban on using AI at work. The law distinguishes sole reliance on an automated system from circumstances in which a system assists a decision. Where the employer primarily relies on its output for the relevant action, the new framework requires human corroboration with additional information and includes notice and data-description requirements. [2]
That distinction is important because the word AI can obscure the actual workflow. A system might summarise a file, produce a score, flag a pattern or recommend an action. The consequential question is how the employer uses that output and what happens before the decision reaches the employee.
A meaningful review has to be capable of changing the result. If the reviewer cannot understand the basis of a recommendation, cannot obtain relevant context or has no practical authority to disagree, the presence of a person may add little to the decision.
This is an analytical test of the workflow, not an allegation that any particular employer is already failing it. It also explains why compliance cannot be judged solely from whether a human name appears at the end of the process.
The difficult work happens before the signature
A recommendation may appear persuasive because it compresses a complicated record into a simple output. The compression is useful only if the information omitted does not change the decision that should be made.
An absence in a dataset, for example, can mean different things depending on how the data were collected. A performance measure can be relevant without describing every part of the job. These are reasons to inspect the context, not assumptions that every automated assessment is wrong.
The human reviewer needs enough information to ask the appropriate questions. What was measured? Over what period? Is the comparison meaningful? Is there contradictory material? Can the employee explain an apparent discrepancy? The answer should depend on the evidence of the particular case rather than on whether the system’s output looks precise.
The new framework’s requirement for corroboration points towards that distinction. SHRM’s account describes additional material such as managerial evaluations, personnel records, work product or other relevant evidence. A second source is useful only if it adds genuine information; repeating the machine’s own conclusion in another field would not perform the same function. [3]
There is a related organisational question. Who is responsible for discovering that the source material itself is incomplete or wrong? Moving a process to software does not make that responsibility disappear. It can make the point at which it needs to be exercised less visible.
Employers have objections too
SHRM says it strongly opposed SB 947 and intends to engage on implementation. Its position is part of the policy debate, not an independent finding that the law will fail. Likewise, supporters’ descriptions of the law as a major protection do not establish that its benefits have already materialised. [3]
The substantive issue is how to preserve useful tools while making consequential decisions answerable. Automated assistance may help organise information or apply a process consistently. Human decisions can also contain errors, bias or poor judgment. The comparison is not between a fallible machine and an infallible person.
The case for human involvement is that someone can examine context, take responsibility and reconsider the recommendation. Its strength depends on the design of that involvement. A review with insufficient time, information or authority risks becoming a formality. A review that genuinely tests the decision may require a different allocation of work.
Those practical demands should be discussed openly. A rule can have a legitimate protective objective and still require clear implementation guidance. Conversely, the existence of an administrative burden does not establish that the underlying safeguard is unnecessary.
A notice is the start of accountability
For an affected worker, knowing that an automated system played a major role can change the questions they ask. It can direct attention to the information used and to the person responsible for explaining the decision.
But disclosure alone does not resolve a disagreement. The explanation must be understandable, the relevant information usable and the review process capable of addressing a mistake. An employee should not have to infer the reason for a consequential decision from the existence of a score.
It is equally important not to promise a legal remedy that the particular law does not provide. Eligibility, enforcement and interaction with other employment protections are jurisdiction-specific matters. This is an account of the new policy and its practical questions, not advice about an individual dismissal.
The transition period makes accurate communication particularly important. “California has signed a law” and “these obligations already apply” are different statements. Employers and workers should consult the operative text and appropriate guidance, rather than rely on the shortest version of a headline.
Who can say no to the system?
The test next July will not be how many organisations announce a human-in-the-loop policy. It will be whether their processes preserve a real capacity to question, corroborate and explain.
That can be assessed through concrete practice: the information available to reviewers, the reasons recorded for decisions, the ability to identify errors and what happens when a recommendation is rejected. Those measures should not be confused with a claim that every disagreement must end in the employee’s preferred outcome.
The central principle is narrower and more demanding. An employer choosing to use a system remains responsible for the employment decision. The person designated to review it should be able to do more than make the decision look human.
AI may supply an answer quickly. The point of the new rule is to ensure that speed does not settle who has the final say.
Sources & notes
Explore the sources cited in this article.