On September 30, 2026, California enacted SB 947, the No Robo Bosses Act of 2026. Beginning July 1, 2027, its workplace provisions prohibit sole reliance on an automated decision system for discipline or termination. When an employer primarily relies on automated output, a human must corroborate the decision.
The important step comes when that review fails. If the output cannot be corroborated, or the reviewer concludes it is inaccurate, incomplete, or misleading, the employer cannot use it for the disciplinary or termination decision.
A manager’s signature therefore does not answer the central question. What evidence supported the decision, and what happened when the recommendation was unsupported?
The Review Must Reach Beyond the Recommendation
The enacted law permits corroboration using data collected or used to produce the automated output, or other relevant supporting information. Its examples include supervisory or managerial evaluations, personnel files, employee work product, peer reviews, and witness interviews, which may include relevant online customer reviews.
The evidence need not come from a source independent of the system’s inputs. That is a meaningful limit on how the requirement should be described. The law does not mandate a second investigation conducted without access to the recommendation. It requires human corroboration and prohibits use when the prescribed review does not support the output.
Checking the arithmetic is only one possible task. A score can correctly summarize recorded activity while giving a misleading account of the employee’s work. The reviewer needs to examine whether the information supports the contemplated action, including context that the score omitted.
An Accurate Count Can Support a Misleading Conclusion
Consider a hypothetical employer whose software recommends discipline because an employee completed fewer transactions than colleagues. The count is accurate. But the employee handled complex cases, performed work outside the tracked application, and lost time to an equipment failure.
Confirming the transaction count would not resolve whether the comparison supports discipline. The reviewer would need to examine the work and the explanation for the difference. A second report repeating the same count would add little to that inquiry.
A different hypothetical makes the stopping rule clearer. Software recommends termination for repeated unauthorized absences. The reviewer checks the dates against approved-leave records and discovers that the recommendation treated authorized leave as misconduct.
The attendance allegation has failed. The employer needs a process that stops the action based on that output, rather than allowing a reviewer to acknowledge the error and approve the recommendation anyway. These examples illustrate the analysis; they describe no reported incident under the new law.
The Employer Needs a Record of What the Person Examined
SB 947 does not prescribe a review form, minimum review time, or universal evidentiary threshold. The practical task is to make the review specific enough that another person can understand what was checked and why the resulting decision followed.
A useful contemporaneous record begins with the output and the contemplated action. What did the system claim? What did the employer propose to do? How much did the recommendation influence that proposal? Preserving the original output avoids reconstructing those questions after a dispute arises.
The reviewer can then identify the underlying records actually examined, the work samples reviewed, and any people interviewed. Identifying the particular evidence is more useful than stating that the personnel file was available.
The record should address contradictions and missing context. In the transaction example, it would explain whether complex assignments and equipment downtime account for the difference. In the attendance example, it would identify the approved-leave records that defeated the allegation.
Finally, it should explain which propositions were supported and which were rejected. The existence of a personnel file does not corroborate every conclusion in it. A managerial evaluation or customer review can itself contain an error. The statutory examples identify possible sources; their relevance and substance still matter.
These are recommended implementation practices. The statute does not enumerate this documentation checklist. A checkbox stating that a human reviewed the recommendation provides much less information about whether corroboration occurred.
Disregarding the Output Means Stopping Its Influence
An employer cannot directly demonstrate a reviewer’s private thought process. It can preserve observable evidence that an unsupported recommendation stopped controlling the action.
In the attendance example, that would include recording the rejected allegation, stopping the pending recommendation, and ensuring that downstream personnel records and approval steps do not continue to treat the absences as unauthorized. Preserving the original output for review is compatible with excluding it from the grounds for discipline.
The practical distinction is between retaining evidence of the mistake and continuing to rely on the mistake. Deleting a recommendation would not, by itself, show that its influence ended. Relabeling it as a human decision would not establish that either.
If the employer considers a separate, supported ground for action, the record should identify those facts and explain why they justify the decision. That is not an automatic permission to reach the same result under a different label. The unsupported output cannot supply a missing factual premise.
A workflow can make that boundary concrete. A failed review can block the proposed action, require correction of the stated grounds, and send any revised proposal for review on its own evidence. Such controls are practical choices, rather than mechanisms expressly prescribed by SB 947.
The Employee Receives More Than an Assurance
The bill’s author describes the law as requiring both human verification and notice when employers use automated systems to assist discipline or termination decisions. The enacted notice provisions make the representation specific: the employer must tell the affected employee that a human reviewed the decision and corroborated the output.
Employees also receive a conditional right to request a meaningful, objective description of their own data used by the system. These provisions do not create unrestricted access to other people’s information or the vendor’s entire model. They do make the employer’s account of the review something the affected employee can question.
The review record, notice, and final decision should tell the same factual story. If the record shows that a material allegation was rejected, a notice asserting corroboration of that allegation would require attention before it is sent. A standard notice cannot substitute for the underlying review.
The Enforcement Provision Makes Demonstration Consequential
Labor Code section 1526.1(c) adds a significant evidentiary mechanism in proceedings alleging a violation of section 1522(b) or (c): after the stated showing of ADS use, the employer must demonstrate that it did not primarily rely on the system or that it complied with the review and notice requirements. The provision uses the term deactivation in its trigger; the operative decision provisions address discipline and termination.
The distinction between some use and primary reliance therefore matters. A signature or policy calling every result a human decision does not establish the actual role of the output. Records of the sequence, evidence, and reasons can help explain that role.
Coverage and exceptions also matter. The law includes specified collective-bargaining and federal-purpose exceptions. Its corroboration requirement should not be presented as a universal rule for every automated employment decision. Compliance also does not resolve obligations under discrimination law or other applicable protections. The law does include a safe harbor, but it reaches notice alone: an employer satisfying this part’s notice requirements need not meet substantially similar automated-decision notice provisions under other state law. That relief excludes quotas and other automated standards applied to working conditions, and it preserves applicable Privacy Protection Agency automated-decisionmaking regulations for businesses covered by the Consumer Privacy Act.
