California No Robo Bosses Act Requires More Than a Human Sign-Off

California has enacted restrictions on automated firing and discipline, but adding a manager’s approval does not necessarily satisfy the new law. The California No Robo Bosses Act prohibits employers from relying solely on automated decision systems for those decisions. When employers primarily rely on their output, a human must corroborate the decision. Its requirements become operative July 1, 2027.

Governor Gavin Newsom signed Senate Bill 947 on September 30, 2026, after vetoing its broader predecessor, SB 7, in 2025. The narrower law preserves automated assistance while requiring employers to retain responsibility for consequential decisions. That creates a practical conflict between software designed to accelerate management and review that requires someone to examine the underlying evidence.

The immediate pressure falls on employers, human resources teams, and vendors whose systems generate performance scores or disciplinary recommendations. Workers also gain notice and data-description rights in covered situations. The central question is whether human review changes the decision process or merely adds an approval screen.

What the California No Robo Bosses Act Actually Changes

SB 947 regulates how employers reach disciplinary and termination decisions, including what happens when an automated recommendation drives the outcome.

An automated decision system, or ADS, uses computational methods to produce consequential scores, classifications, or recommendations that assist or replace human judgment. The definition covers machine learning, statistical modeling, data analytics, and artificial intelligence. Coverage therefore extends beyond chatbots and software marketed explicitly as AI.

The chaptered bill text establishes different rules for sole reliance and primary reliance. Employers cannot rely solely on an ADS when making disciplinary or termination decisions. When they primarily rely on its output, they must direct a human to corroborate the decision using relevant information.

That information can include data collected or used to generate the output, along with other supporting evidence. The statute identifies supervisory evaluations, personnel files, employee work, peer reviews, and witness interviews as possible sources. Relevant online customer reviews can also form part of that supporting information.

The consequence of unsuccessful review matters as much as the review requirement. Employers cannot use the output for discipline or termination when they cannot corroborate it. The same restriction applies when the reviewer concludes that the output is inaccurate, incomplete, or misleading.

This creates a distinction between inspecting a recommendation and simply accepting it. A manager who approves a score without examining supporting material has not necessarily performed the corroboration the law describes. The statute does not prescribe a universal review form, but it requires more substance than the existence of an approval button.

The law also prohibits specified uses of automated systems, including inferring an employee’s protected status under California’s employment discrimination law. Other provisions address interference with legal compliance and adverse action connected to workers exercising their legal rights. These restrictions sit alongside the decision-review rules.

The California No Robo Bosses Act does not establish a universal ban on workplace algorithms. Ordinary databases, calculators, spam filters, firewalls, antivirus software, and identity-management tools are expressly excluded from the ADS definition. A system’s function and role in decision-making matter more than its product category.

For readers encountering the story in November 2026, the timing deserves precision. California has enacted the law, but its requirements become operative on July 1, 2027. Employers have preparation time, while workers should distinguish these forthcoming protections from rights already available under other laws.

The narrower focus also limits what employers should infer. SB 947’s central review and notice obligations concern discipline and termination. Treating it as a general advance-notice requirement for every workplace AI tool would misstate the enacted framework.

Employers and Software Vendors Face Different Kinds of Pressure

The employer makes the employment decision, but the software’s design can determine whether meaningful review is practical.

For employers, the challenge begins with identifying where automated outputs enter disciplinary processes. A familiar attendance dashboard can become consequential when a supervisor uses its score to recommend dismissal. An embedded feature deserves scrutiny even when nobody purchased a separate product labeled an AI manager.

That distinction makes departmental boundaries less useful than tracing the actual decision. Procurement may know the vendor, information technology may manage the integration, and human resources may own the policy. The supervisor using the recommendation might understand only the final score.

A decision can therefore depend on automation without any single department seeing the complete chain. The operational question is where a recommendation originated, what information supported it, and how heavily the employer relied on it. A general inventory of software licenses will not necessarily answer those questions.

Vendors face a related product challenge. If their interface presents only a risk label or recommended action, customers might struggle to corroborate the result. Access to underlying records, understandable input descriptions, and a way to challenge recommendations become commercially relevant capabilities.

These capabilities should not be confused with guaranteed compliance. A vendor cannot establish what a customer’s manager actually reviewed or how heavily the employer relied on a recommendation. Software can support a defensible process, while the employer still needs to operate that process.

The organizational tradeoff is clear. Automating the identification of possible problems can reduce administrative work. Investigating a consequential recommendation still consumes managerial time and requires access to information beyond a summary screen.

Broader research helps explain why this tension reaches beyond a single California statute. The OECD’s algorithmic management research examines software used to instruct, monitor, and evaluate workers. Managers reported perceived decision benefits alongside concerns about accountability, explainability, and worker protection.

Those findings support a more balanced reading than either blanket enthusiasm or blanket rejection. Software can help managers detect patterns they would otherwise miss. It can also leave responsibility unclear when a recommendation proves wrong.

The commercial divide consequently runs between systems optimized to deliver an answer and systems designed to support examination of that answer. Employers choosing between those approaches need to consider review workflows alongside prediction quality. An accurate aggregate performance measure does not resolve every individual employment decision.

The same distinction applies to workers using personal AI tools. Assistance with organizing information differs from an employer using automated output to impose discipline. The stakes change when a tool moves from supporting someone’s work to evaluating whether that person should keep their job.

For enterprise buyers, California SB 947 employer impact therefore extends into implementation decisions. Who can inspect the source material, who can reject the output, and who answers employee questions become practical purchasing considerations.

Human Review Must Test the Recommendation

The central safeguard is corroboration, and corroboration loses value when a reviewer merely repeats the system’s conclusion.

Consider a hypothetical attendance system that recommends discipline after flagging repeated late arrivals. A supervisor sees the recommendation and approves it because the dashboard appears consistent. That approval establishes human involvement, but it does not necessarily establish that anyone checked the underlying attendance information.

A substantive review would examine the records supporting the recommendation. It could identify an incorrect shift assignment, an approved schedule change, or missing context. These examples illustrate the review problem; they are not reported findings about a particular employer.

The distinction matters because a conclusion can look precise while its inputs remain incomplete. A timestamp might accurately record an event yet fail to establish that the employee violated the relevant policy. The reviewer needs to connect the evidence to the proposed disciplinary action.

SB 947 permits corroboration using information collected or used to produce the automated output. It also allows other relevant supporting information. The law therefore does not require every reviewer to recreate the system independently or obtain an entirely separate dataset.

That flexibility creates both an implementation advantage and a limitation. Reviewers can use existing records rather than commissioning a new investigation for every covered decision. However, inspecting the same incomplete records without questioning their meaning can reproduce the original mistake.

A second hypothetical scenario makes that risk clearer. A performance tool assigns a low score after counting completed customer cases. The employee’s work includes unusually complex cases that take longer to resolve. Confirming the case count alone would not establish that the score fairly captures performance.

The point is not that every quantitative measure is misleading. It is that accurate measurement and justified discipline are separate judgments. A reviewer must understand what the output establishes, what it omits, and whether the supporting information corroborates the decision.

The law’s rejection rule gives that examination practical consequences. If the output cannot be corroborated, the employer cannot use it for the disciplinary or termination decision. Review becomes meaningful when it can stop reliance on an unsupported recommendation.

That requirement also changes the significance of managerial authority. A reviewer who cannot obtain records, question assumptions, or reject the output has limited ability to test it. Organizations should examine whether their internal processes give reviewers those practical capabilities.

The California No Robo Bosses Act thus puts pressure on the handoff between software and management. A system can identify a possible concern, while a human must take responsibility for the evidence supporting consequential action. The workflow needs room for disagreement.

This is also where automation bias becomes relevant, meaning a tendency to defer to a machine-generated recommendation. A polished interface can make an output appear settled before anyone examines its basis. Review procedures should account for that possibility rather than assuming a human presence automatically prevents it.

The strongest evidence of useful review would be a process that detects unsupported outputs and changes decisions accordingly. That is an operational assessment, not an additional statutory reporting requirement. It asks whether the safeguard performs its intended function.

Workers Gain Notice, but Transparency Has Boundaries

Covered employees gain information about automated involvement, without receiving an unrestricted right to inspect the entire system.

When an employer primarily relies on an ADS for discipline or termination, it must provide a written postuse notice. That means notice delivered when the employer informs the employee of the decision. It is not a general requirement to announce every tool before deployment.

The communication must stand alone, use plain language, and follow the language used for routine workplace communications. Delivery must use a simple method, which can include email or a hyperlink. These requirements aim to make the information accessible when the decision becomes concrete.

The notice must identify the employer’s primary reliance on an ADS and state that a human reviewed and corroborated the output. It must also provide a human contact for further information. Employees must receive information about their data-description rights and protection against retaliation.

Separately, covered employees can request a meaningful, objective description of their own data used by the system. Employers must anonymize other individuals’ personal information in that description. The provision therefore balances access with privacy interests involving coworkers, customers, and others.

A description of employee data is narrower than disclosure of source code or unrestricted access to a model. It also differs from an automatic entitlement to every record held by the vendor. Readers should avoid treating the provision as a general right to reverse-engineer workplace software.

Still, the combination of notice and access creates an opportunity to identify factual misunderstandings. An employee might discover that a decision relied on attendance records from the wrong schedule. Alternatively, the description might confirm the relevant information while leaving disagreement about the employer’s interpretation.

The usefulness of that exchange depends partly on the quality of the description. A vague statement that the system considered performance information gives less practical insight than an explanation of the categories actually used. The statute’s requirement for meaningful, objective description matters in that context.

Employer-side lawyers have also highlighted uncertainty around the threshold for primary reliance. In their October 30 analysis of workplace AI restrictions, Ogletree Deakins attorneys identify that undefined language as a likely dispute. Its boundaries will matter for review and notice obligations.

For example, a manager might consult a recommendation alongside interviews and existing performance records. Determining which information primarily drove the decision requires examining the actual process. Calling software advisory does not, by itself, explain the weight the employer gave its output.

The law authorizes enforcement through the Labor Commissioner and public prosecutors. It also includes protections against retaliation for asserting covered rights. Employers should distinguish those mechanisms from a new, standalone private lawsuit right, which the legislation does not create.

California SB 947 explained through its information rights reveals a deliberate limit. Workers receive a route to understand automated involvement in covered decisions, but transparency does not automatically establish that the outcome was fair. Access and accountability remain connected, separate questions.

A Narrower Law Leaves Important Disputes Unresolved

California secured a targeted restriction on automated discipline and termination while leaving broader arguments about workplace algorithms unsettled.

The distinction from SB 7 is essential historical context. Newsom vetoed that earlier proposal in 2025. SB 947 returned with a narrower framework, and its final legislative changes removed provisions that had reached further into workplace management.

The Legislature’s bill analysis records changes involving independent contractors, compensation-related restrictions, customer ratings, and data access. Those revisions show why summaries of earlier drafts can mislead readers about the enacted law. A proposal’s history does not substitute for its final provisions.

Labor supporters frame human oversight as protection against workers losing their livelihoods through automated mistakes. In the governor’s worker protection announcement, California labor leaders welcomed the workplace legislation. Their position emphasizes keeping people responsible when technology influences consequential employment decisions.

Business opposition raises a different concern. Arguments recorded in the legislative analysis question whether the restrictions adequately account for useful automated tools. Critics warn that regulatory uncertainty can discourage adoption, including systems intended to improve consistency or address bias.

These positions identify a real tradeoff. Automated evaluation can introduce errors, while human evaluation can contain prejudice, inconsistency, and incomplete information. Requiring human involvement does not establish that every resulting decision will be better.

The narrower and more defensible claim is that corroboration creates an opportunity to catch unsupported outputs. Whether organizations use that opportunity effectively remains an implementation question. The law establishes obligations without demonstrating their future results.

The scope also contains exceptions that require careful reading. A collective bargaining agreement qualifies for an exemption only when it satisfies specified conditions, including an explicit waiver and protective terms. Union representation alone does not establish that the exemption applies.

Another limited exception concerns certain uses required or reasonably necessary under federal rules or contracts involving aircraft and specified national-security or defense-related development. That exception applies to covered operations. It is not a general exemption for any organization with federal customers.

Other legal duties also remain relevant. SB 947 preserves applicable privacy-related automated-decision regulations under California’s consumer privacy framework. Its notice coordination provision therefore should not be read as eliminating every other obligation involving employee information.

There is even a drafting detail that deserves caution in broad summaries. The enforcement burden provision retains a reference to deactivation, while central substantive duties address discipline and termination. That wording should not become a shortcut for claiming universal gig-worker coverage.

The California No Robo Bosses Act establishes a specific accountability framework, rather than resolving every dispute about algorithmic management. Its practical strength will depend on how employers interpret primary reliance, conduct corroboration, and communicate with affected employees. Claims of either complete protection or complete prohibition go beyond the evidence.

What to Watch Before July 2027

The next useful signals are concrete interpretations and operating changes, rather than general promises about responsible workplace AI.

First, watch for official guidance addressing the distinction between sole reliance and primary reliance. Clarification would help employers understand which workflows trigger corroboration and postuse notice. As of this reporting, the undefined threshold remains an important uncertainty.

Over the coming months, useful guidance would address mixed processes involving automated scores, supervisor observations, and existing personnel records. The key issue is how those inputs contribute to the decision. A product’s description as supportive or advisory does not fully answer that question.

Any clarification that examines actual reliance would strengthen the law’s accountability logic. An approach focused mainly on labels would leave more room for formal compliance without meaningful examination. Readers should evaluate the substance of future guidance rather than assume an announcement settles the issue.

Second, watch how workplace software vendors change their products and customer materials. Relevant developments include clearer input descriptions, access to supporting records, and review features that allow managers to question recommendations. These would indicate attention to the practical demands of corroboration.

An approval screen alone would be a weaker signal. The more useful question is whether a reviewer can understand the recommendation’s basis and record reasons for rejecting it. Product demonstrations should show that entire sequence, including what happens when evidence contradicts the output.

Enterprise buyers can apply that distinction during renewal and procurement discussions. They should ask vendors to demonstrate a contested recommendation rather than only a successful prediction. A system’s response to disagreement reveals capabilities that a polished summary may conceal.

Third, watch whether employers change disciplinary procedures before the operative date. Relevant signs include assigned review responsibilities, accessible supporting information, usable notices, and a clear route for employee data-description requests. These changes would show implementation moving beyond a general AI policy.

Organizations should also examine whether managers have time to perform the reviews assigned to them. A formal duty can become superficial when the surrounding workflow rewards immediate approval. This is a management design issue as much as a technology procurement issue.

For knowledge workers, the distinction between assistance and evaluation remains useful beyond this law. Tools that organize a personal knowledge base serve a different purpose from systems used to justify discipline. Employers should explain when information crosses from supporting work into evaluating the worker.

California SB 947 employer impact will become clearer through these operational choices. The law gives organizations a deadline and a framework, while implementation determines whether employees encounter an informed reviewer or another automated step.

Start with a real decision your organization makes, then trace where automated output enters it. Identify who can inspect the supporting information and who can stop an unsupported recommendation. Ask what the affected employee would receive and which person would answer their questions. Compare that process with the enacted requirements before relying on a vendor’s compliance label. For workers, understand what information a covered notice should provide and what the data-description right actually covers. The California No Robo Bosses Act makes human responsibility central to automated discipline and termination. Before July 2027, can your workplace show how that responsibility operates when someone challenges the machine’s recommendation?