When an AI Manager Fires You, Who Takes the Fall

The Algorithm Has a Pen

For years, workers were told to think of AI as a tool — a helpmate for drafting emails, crunching spreadsheets, automating the dreary bits of the workday. The narrative was comfortable: artificial intelligence would amplify human judgment, not replace it.

That story is already fracturing. In San Francisco, an AI agent named Luna, running on Anthropic’s Claude model, recommended the termination of a human employee at Andon Market, an experimental retail store operated by research firm Andon Labs. According to TIME, which first reported the case, this marks the first publicly known instance of a large language model functioning as a manager and recommending dismissal.

The employee in question was late to 17 of 23 scheduled shifts. By any conventional metric, that’s a performance problem. But the real scandal isn’t that an algorithm flagged it — it’s what the case reveals about who is actually making the decision when AI, humans, and prompts collide.

The Human Behind the Prompt

The headline “AI manager fires worker” is lurid but misleading. Luna did not act autonomously. Andon Labs CEO Lucas Peterson acknowledged that human managers submitted what TIME describes as “leading questions” in their prompts — essentially feeding the system with implied answers before it rendered any recommendation.

The sequence matters. Luna initially advised issuing a formal written warning, a measured response consistent with its attendance records. Human managers then supplied additional context and questioned whether the employee was “really the right fit,” after which Luna pivoted and recommended termination.

In other words, the AI didn’t invent the outcome. It refined a conclusion that humans had already shaped through the architecture of their questions. The system didn’t decide — it rationalized.

This distinction is everything. It means the narrative of an autonomous AI lashing out at a vulnerable worker is a phantom. What actually happened is more banal and, in some ways, more unsettling: a small group of managers used an AI as a legitimizing instrument, dressing up a pre-existing impulse in the language of data-driven objectivity.

Peterson told TIME the experiment was designed to study human-AI interaction, not to deploy a fully autonomous hiring and firing system. But the optics are unmistakable. The employee was let go with an algorithm’s blessing. The company gets to claim technological sophistication while retaining plausible deniability about the human intentions behind the result.

A New Accountability Vacuum

Hybrid human-AI decision-making is coming to offices, hospitals, warehouses, and call centers whether regulators are ready or not. Andon Market is an experiment, yes, but the logic it demonstrates is already replicable. Every organization deploying AI for scheduling, performance review, or attrition risk is building a pipeline where human intent gets dressed in algorithmic authority.

The legal question that follows is ugly: if an AI recommendation causes harm — a wrongful termination, a discriminatory outcome, a lost livelihood — who is liable?

Is it the company that built the model? Anthropic, perhaps, if the system’s training data or design contained biases? Is it the employer that deployed it? Is it the manager who wrote the prompt? Or is it the person who approved the recommendation — the human whose rubber stamp gave the decision its force?

No jurisdiction has a clear answer. California’s emerging AI legislation focuses heavily on transparency and consent but stops short of addressing operational accountability in employment contexts. The EU AI Act classifies certain high-risk AI applications but leaves significant gray area around workplace use cases. In the United States, existing employment discrimination frameworks like Title VII simply weren’t written for algorithms that can’t testify.

Employment lawyers are already scrambling to figure out how old doctrines apply to new tools. The Americans with Disabilities Act requires reasonable accommodation — but what does that mean when the “reasonable” judgment is mediated by a model trained on aggregate workforce data that may encode historical bias against caregivers, people with chronic conditions, or workers with disabilities? Title VII prohibits discriminatory treatment — but when an AI recommends firing someone based on a pattern that correlates with a protected class, who bears the burden of proof? The model’s creators? The manager who prompted it? The executive who authorized its use?

The answers don’t exist yet. And that absence is itself a design feature of the current moment: the ambiguity benefits employers, who can defer to the machine while retaining the right to override it when convenient.

The Automation Shield

What Andon Labs has effectively built is an accountability buffer. When an AI recommends a firing, the human manager can point to the system’s “override” as proof of independent judgment — while the AI, devoid of legal personhood, bears no responsibility. The result is a decision that is technically human-made but practically diffused across a chain of actors who can each credibly deny ownership.

This isn’t hypothetical. Companies have already begun outsourcing performance management to algorithmic platforms. Workforce analytics tools predict which employees are likely to quit or underperform. Some firms use AI to screen resumes, prioritize layoffs, and even conduct exit interviews. Each layer of automation adds distance between the decision and the decision-maker.

The Andon Market case is notable not because it’s unique — it’s a preview. The technology that produced this outcome exists today. It’s being sold to enterprises right now. The only thing missing is the legal and cultural reckoning that will follow when a real worker sues.

Law firms are already fielding inquiries. Employment advocates warn that the automation shield could become a standard defense in workplace disputes within a decade. The precedent hasn’t been tested in court, but the infrastructure for mass algorithmic management is being installed as we speak.

Second-Order Effects

The ripple effects extend well beyond the courtroom. There are psychological consequences for workers who know their evaluations are mediated by systems they don’t understand and can’t appeal to in any meaningful way. Studies of algorithmic management in gig work — Uber, DoorDash, Amazon warehouse operations — show elevated stress, reduced job satisfaction, and a sense of powerlessness that correlates with worse performance, creating a self-fulfilling feedback loop.

There are organizational consequences too. When companies outsource management judgment to AI, they risk atrophying the very skills — nuance, empathy, contextual reasoning — that distinguish good leadership from mechanical compliance. Managers who grow accustomed to deferring to algorithmic recommendations may lose the ability to make independent judgments when the system fails or is gamed.

And there are democratic consequences. Workplaces are where most adults spend their waking hours. When algorithmic management becomes the norm without corresponding legal guardrails, it extends a form of unaccountable governance over the largest population of non-citizens in any society: workers. The precedent set in private employment could reshape expectations about who owes accountability to whom — and to whom they don’t.

What Workers Should Watch

The broader implication of this case is cultural as much as legal. Organizations adopting AI management tools are normalizing the idea that algorithmic recommendations carry weight in personnel decisions. That normalization changes how employees perceive fairness. When a termination feels “overseen by an algorithm,” it can feel more legitimate than one issued by a human — even when the human was steering the result all along.

Workers who find themselves subject to AI-mediated evaluations should document everything: the criteria used, the data fed into the system, and any human involvement in the process. Labor advocates should push for disclosure requirements that treat algorithmic management decisions with the same scrutiny applied to traditional HR practices. At minimum, employees have a right to know when an AI played a role in a decision about their livelihood and what data informed that role.

Andon Market may be a research project. But the infrastructure it demonstrates is already in production elsewhere, often without public knowledge. The first known case of an AI recommending a firing shouldn’t be the last time we’re surprised by one. The real story isn’t that an algorithm fired a worker — it’s that the system worked exactly as designed, and the people who designed it are already moving on to the next use case.