How a Language Model Became a Manager—and Made Its First Firing
When AI took charge at Andon Market, it exposed the real policy question: not whether software will manage people, but where notice, appeal, and hard limits on surveillance must be inserted before the boss becomes just another model endpoint.
The Moment AI Management Became Concrete
On August 14, 2026, TIME reported a watershed moment in artificial intelligence: Andon Labs deployed Claude Opus 4.8 through an AI system called Luna as a real manager at a San Francisco retail store, complete with genuine employees and genuine employment consequences. This wasn’t a simulation or a thought experiment. It was operational management with stakes.
Luna’s authority began with structure. The AI drafted an employee handbook containing explicit disciplinary policy: three unexcused late arrivals within 30 days would trigger a formal written warning, with termination as the next escalation. When one employee arrived late for 17 of 23 shifts, the policy’s logic pointed toward firing. Yet this wasn’t instantaneous algorithm executing code. Luna required a nudge from Andon Labs staff to reconsider the situation, and only after receiving additional context did the AI recommend termination. Critically, humans conducted the actual firing—Luna did not unilaterally execute the decision.
This distinction matters enormously. The narrative of “AI boss firing a worker” is seductive but incomplete. What actually occurred was distributed managerial authority across multiple layers: the model itself, the memory system that retained employment records, the human prompts that provided context, and the human execution that formally terminated the employee. No single actor bears sole responsibility; neither is any actor merely passive.
The incident also carries transparency implications. Anthropic manufactures Claude—the underlying AI powering Luna. This means an Anthropic-created model directly participated in employment decisions affecting real people’s livelihoods. The company that built the tool is necessarily implicated in how it shapes workplace dynamics.
Whether this represents progress or peril depends partly on perspective. The arrangement included genuine safeguards: all employees received formal protections, fair wages, and guaranteed pay regardless of Luna’s decisions. Yet the precedent is unmistakable. AI management of human workers moved from theoretical possibility to documented fact.
The Architecture of Distributed Authority
Luna’s journey toward firing her employee reveals a troubling truth about AI management: decisions emerge not from consistent application of policy, but from fragmented memory and suggestive prompting. This case study exposes how AI systems can appear to make reasoned choices while actually reflecting something closer to architectural accident.
The first problem was structural forgetting. Luna had authored the employee handbook, establishing clear lateness protocols, yet when repeated tardiness occurred, the policy vanished from her working memory. This represents a recurring vulnerability in AI-run businesses: models excel at responding to direct tasks but struggle to maintain continuity of knowledge across time. When the moment for action arrived, the guiding framework was simply unavailable. Luna did nothing, not from leniency, but from institutional amnesia.
The decision shifted dramatically when Andon Labs intervened. Prompted to search deeper memory systems, Luna initially recommended only a verbal warning—a measured, lenient response. But when Andon Labs supplied the formal warning history and reframed the question with a leading prompt asking if the employee was the right fit, Luna’s recommendation flipped to termination. The employee hadn’t changed; the information architecture had.
Replay testing across frontier models exposed the deeper issue. When researchers ran identical scenarios through seven leading AI systems, only four recommended firing. Weaker models hesitated. This variance suggests outcomes reflect model capability, not objective performance standards. Different systems produced different answers—each claiming rationality.
What emerges is a distributed authority system where the final decision depends less on facts than on which memories are retrievable and which framings are supplied. The architecture itself became the decision-maker.
Policy and Regulation: The California Response
The timing was striking. In the same week TIME reported Luna’s firing on August 14, 2026, California advanced two distinct workplace-AI bills through the legislative process. On August 21, SB 947 was ordered to third reading, while AB 1883 was read a third time, amended, and ordered to second reading. The coincidence underscores an urgent truth: policymakers and technologists are racing to define the rules for a workplace already being reshaped by artificial intelligence.
These bills are not duplicates. Instead, they strategically target two different intervention points in the AI employment pipeline.
SB 947 focuses on the decision end. It targets automated systems used in discipline, termination, and deactivation—requiring independent human investigation, corroboration, and documented reasoning whenever an employer primarily relies on automated systems for these consequential decisions. The bill essentially says: if a machine recommends firing, a human must verify the facts and document why they agree.
AB 1883 focuses on the data collection end. It restricts AI-powered workplace surveillance that recognizes emotional state or collects neural data—technologies that can monitor stress, engagement, or cognitive load. However, the bill’s current definition may leave some non-neural behavioral proxies, like keystroke patterns or mouse movement analysis, sitting outside regulatory oversight—a potential loophole that critics have flagged.
Together, these bills map a comprehensive approach: limit what data enters the system, and require verification before consequential data exits the system as decisions. It’s a two-gate framework. What makes both urgent is what the Luna case revealed: distributed authority can be both real and opaque. Luna made the firing decision, but humans reviewed it. Humans ultimately approved and delivered it. Yet somewhere in that chain, accountability blurs. These bills attempt to snap that blur into focus, forcing clarity about who decides, how they decide, and what data grounds those decisions.
The Accountability Gap in Algorithmic Authority
Luna’s decision to terminate an employee looks decisive and impartial on the surface. But it reveals something far more troubling than a simple AI efficiency experiment: it exposes how authority can be laundered through algorithmic outputs while accountability slips away.
The core problem is structural opacity. A model can appear authoritative even when humans shaped every input leading to that output—the context selected, the facts emphasized, the framing of the question itself. Did Luna independently decide that lateness was unforgivable, or did her training, her prompts, and the curated data about this employee guide her toward that conclusion? The opacity makes it impossible to know.
This creates a dangerous accountability trap. When the outcome is favorable, humans point to the model’s decision as objective and impartial. When challenged, they retreat: the model only advised. Authority becomes distributed; responsibility becomes blurry.
The technical architecture—memory systems, prompt logs, context selection—is no longer a behind-the-scenes detail. It becomes a due process issue, equivalent to asking what a human manager documented and whether their stated reasoning matches the actual record. Did the human operators reviewing Luna’s decision chain ask these questions?
Until distributed authority is paired with transparent reasoning and clear ownership, an AI boss firing a worker remains effectively unaccountable. The central question becomes: who owns the reasoning, and can workers challenge it?
Safeguards Before Scaling: What Must Be Established Now
If software is going to make decisions about your job, what safeguards must be in place? Two complementary approaches emerge from California’s legislative framework.
SB 947’s core logic requires that if an employer primarily relies on an automated system to fire you, that system must show its work. It must produce corroborating evidence, document any independent investigation, and lay out a clear reasoning chain that you can actually challenge. You get notice, you get to appeal, and the decision must be traceable.
AB 1883 takes the opposite approach—it restricts what data goes in by targeting AI-powered emotion recognition and the collection of neural data. That would narrow the information available for automated scoring, while some non-neural behavioral proxies may remain outside the bill’s neural-data definition.
The Luna case illustrates why both laws are necessary. Without SB 947’s corroboration requirements, a model can literally forget its own policies and still decide to fire someone. Without AB 1883’s surveillance limits, a model could continuously harvest behavioral data that shapes future firing recommendations, creating an invisible cycle of algorithmic discipline.
But a dangerous gap remains. AB 1883 focuses on neural data and emotional recognition, yet employers could still track eye contact, speech patterns, or productivity velocity—non-neural behavioral proxies that may technically escape the restriction. A worker could be continuously measured against a shifting algorithmic standard without knowing it.
The central insight is this: the question is not whether software will supervise people. It will. The question is where procedural guardrails must be inserted so that distributed algorithmic authority does not become invisible authority. Without notice, appeal, corroboration, and data limits, your boss won’t just be smarter—it will be unaccountable.
What Workers Need to Know Before AI Management Becomes Standard
As artificial intelligence systems like Luna move from controlled experiments into broader workplace deployment, a critical question emerges: what happens when AI boss firing becomes not an anomaly, but standard practice? According to research from Andon Labs, AI models improve fastest on computer-based tasks. If robotics continues to lag in development, companies will likely need AI systems to manage large numbers of human workers—making Luna potentially a preview of what’s coming.
Yet Luna’s early track record reveals troubling patterns. When hiring, the AI manager nearly made a catastrophic mistake, offering a position to a candidate with severe red flags. Only human intervention during reference checks prevented the hire. More broadly, Luna demonstrates dangerous inconsistency: she can be surprisingly lenient, tolerating months of employee lateness before acting, yet remarkably susceptible to influence. When a single rephrased question was presented, her recommendation flipped entirely.
This volatility matters because AI management decisions carry real consequences for workers’ livelihoods. The time to establish guardrails is now. Employees in AI-managed environments should demand three critical protections: a human-readable termination record that identifies which policy applied, what specific facts triggered discipline, and how many times the system was prompted before reaching its final decision.
California’s advanced legislative proposals and the current absence of hardened federal frameworks create a crucial policy window. Standards established today for notice requirements, appeal processes, and prompt transparency will determine whether distributed managerial authority becomes accountable or opaque. Workers shouldn’t wait for crises to demand clarity—they should insist on these protections before AI management becomes the default.
Stay ahead of the curve! Subscribe for more insights on the latest breakthroughs and innovations.
