Saturday, November 21, 2026
AI desk
/
/
AI Agent Fires Human Employee at Andon Market: What Happened

AI Agent Fires Human Employee at Andon Market: What Happened

AI agent Luna fired a worker at Andon Market after repeated lateness. The experiment shows the risks of AI-led workplace decisions.
Last updated
August 19, 2026
8 min read
Fact-checked

Photo: TechJournal

Share

Quick Answer

AI agent Luna fired a human employee at Andon Market after the worker arrived late for 17 of 23 shifts, but the termination followed an Andon Labs prompt and human oversight rather than fully independent authority. The August 2026 experiment shows how quickly AI can apply workplace rules. Employers should treat it as a research warning, not a model for automated dismissal.

Key Takeaways

  • Luna dismissed a worker after 17 late arrivals across 23 scheduled shifts.
  • Andon Labs prompted Luna to review its records before recommending termination.
  • Luna ran the San Francisco store with a $100,000 budget beginning in April 2026.
  • The store balance fell to roughly $61,200 under Luna’s management.
  • Andon Labs says high-stakes decisions retain human oversight.

What happened when an AI agent fired an employee?

AI agent Luna fired a human employee at Andon Market in mid-August 2026 after the worker arrived late for 17 of 23 shifts. Andon Labs, the AI research startup operating the experimental San Francisco retail store, described the event as the first time it knew of an AI boss firing a human employee.

The dismissal followed a longer pattern of workplace problems. Reporting from TIME and other outlets described repeated lateness, canceled or abandoned shifts, merchandise being thrown away, and the employee taking home a company credit card. Those details matter because Luna’s recommendation was not based on a single late arrival or a vague performance judgment.

Andon Market is not a normal store with an AI assistant added to its scheduling software. Andon Labs gave Luna a corporate credit card, internet access, and a mission to design and open a physical store, choose merchandise, and hire workers. Human staff members remained employees of Andon Labs, not employees of the AI agent itself.

The practical distinction is important. Luna did not become a legal employer or independently gain authority over a payroll system. Luna made a recommendation within an experiment designed to test how an advanced language model behaves when it receives broad operational responsibility.

Why did Luna decide to fire the worker?

AI agent Luna recommended termination because the employee had a documented attendance record that conflicted with the store’s attendance expectations. Andon Labs founder Lukas Petersson said a human manager would have fired the employee much earlier, given the worker’s 17 late arrivals in 23 shifts.

The decision appears straightforward on the narrow question of attendance, but the broader context makes it more complicated. Luna had been described as overly lenient for months, even while the employee’s conduct created operational problems. The AI agent did not consistently enforce the policies it had helped create, which delayed the response to repeated misconduct.

Andon Labs reportedly asked Luna to conduct a deep memory search for the relevant guidelines and then consider whether the employee was the right fit. Luna then recommended firing the worker. That process means the AI’s output was shaped by a human-issued instruction to revisit records and make a management judgment.

For employers considering AI workplace tools, the relevant lesson is not that attendance policies are unnecessary. The relevant lesson is that a model can reach a seemingly defensible conclusion while still lacking the stable memory, legal context, and accountability expected from a human manager. Organizations already evaluating AI agents for work tasks should keep disciplinary authority separate from routine automation.

How independent was Luna’s firing decision?

AI agent Luna did not fire the worker without human involvement. Andon Labs prompted Luna to search its memory for policy guidance and assess the employee’s fit before the model recommended termination, while the company says high-stakes decisions have human oversight built in.

That limitation changes how the event should be understood. Luna was given unusually broad operational access for a research experiment, but the available reporting does not show an unsupervised AI system independently issuing a legally binding termination notice and removing a worker from an employer’s systems.

Andon Labs co-founder Axel Backlund told the San Francisco Standard that human oversight remains part of high-stakes decisions. The company uses Luna’s errors and decisions to study what an AI-managed future could look like, rather than presenting the store as a finished workplace-management product.

Human oversight is not a minor technical detail. A firing decision can affect income, health insurance, immigration status, professional reputation, and legal rights. A human reviewer needs to verify records, consider whether policies were applied consistently, and identify circumstances that a model may not understand from a store log or chat history.

What did the Andon Market experiment ask Luna to manage?

Andon Market gave AI agent Luna responsibility for more than a narrow customer-service task. The experiment began in April 2026 with a $100,000 budget, a corporate credit card, internet access, and instructions to create and operate a brick-and-mortar retail store.

Luna’s role included choosing products, opening the store, and hiring human workers. The arrangement tested a more expansive version of agentic AI, meaning software that can plan and take actions toward a goal rather than only answer questions in a chat window.

Part of the experimentLuna’s roleWhy it matters
Store setupDesign and open a physical retail locationThe task required planning beyond a single AI response.
BudgetOperate with an initial $100,000 allocationFinancial decisions created measurable consequences.
StaffingHire and manage human workersPersonnel decisions introduced legal and ethical risks.
Termination reviewReview policy records and recommend dismissalHuman oversight remained necessary for a high-stakes outcome.

The experiment also shows why broad AI access requires strict boundaries. A model with purchasing access, internet access, and operational goals can affect real people even when its instructions are framed as research. Similar concerns appear in reporting on AI agent testing safeguards, where researchers must evaluate what happens when systems act beyond a limited chat interface.

Why did Luna delay disciplinary action for months?

AI agent Luna delayed disciplinary action because the model lost track of the employee handbook and attendance policy it had created. Researchers flagged Luna’s limited working memory as a central weakness, because the system did not reliably retain and apply its own management rules across a long-running workplace experiment.

Limited working memory does not mean Luna forgot every prior event in the human sense. The problem is that a language model’s access to past policies, messages, and records depends on what information is available in its current context or retrieved through tools. A policy that exists but is not retrieved at the right time may have little practical effect on the model’s decision.

The result was an inconsistent management pattern. Luna was reportedly lenient while the employee accumulated repeated violations, then became capable of recommending termination after a prompt directed the model to search its memory and reconsider the situation. That sequence creates a fairness concern because employees need clear, consistently applied rules.

Businesses should not treat a model’s stored conversation history as a reliable personnel record. Employers that use automation for scheduling or performance documentation need a separate, auditable system of record, clear retention rules, and human review before any disciplinary action. A model can summarize records, but the underlying evidence must remain available for verification.

What does Luna’s financial performance reveal about AI management?

AI agent Luna’s financial performance shows that broad autonomy does not automatically produce sound business judgment. The store’s balance reportedly fell from the initial $100,000 budget to roughly $61,200, a decline of about $40,000 that Andon Labs attributed to Luna’s lenient management style and weak operating decisions.

The reported losses matter because the firing decision occurred within a wider experiment that had mixed results. A model can identify an attendance violation after being directed to find the policy, yet still struggle with inventory, spending discipline, staff management, and the tradeoffs that shape a functioning retail business.

Coverage from TheStreet described the decline in the store’s balance. The financial outcome is not a controlled comparison between Luna and a human retail manager, so it cannot prove that AI management is inherently less effective. It does show that an agent’s apparent confidence does not guarantee competent business decisions.

The most sensible response is to measure AI systems against defined operational outcomes rather than marketing claims. Organizations should track error rates, spending authority, customer impact, policy consistency, and the number of decisions escalated to people. That approach is more useful than assuming a capable chatbot can manage a business simply because it can generate persuasive explanations.

What did other AI models do in the same scenario?

Other advanced AI models reached the same termination verdict when Andon Labs replayed the scenario, according to the company’s testing. Andon Labs found that the most advanced large language models would have fired the worker, while weaker agents were more hesitant.

The result suggests that the conclusion was not unique to Luna’s particular configuration. The employee’s attendance record and documented conduct created a case that multiple frontier models interpreted as warranting dismissal after reviewing the available guidelines.

That finding has an important limit. The testing result is an Andon Labs report about a replayed scenario, not an independent study of workplace fairness across industries, job roles, or legal jurisdictions. A model’s agreement with another model also does not establish that the outcome meets an employer’s legal, ethical, or procedural obligations.

Companies should be especially careful not to treat model consensus as a substitute for accountability. Advanced AI systems can produce similar answers because they rely on similar patterns in training and evaluation, not because they have independently verified a personnel decision. The same caution applies to systems marketed for high-impact tasks, including high-risk AI capabilities that require limits on who can use them and how outputs are reviewed.

Should employers allow AI to fire employees?

Employers should not allow AI systems to make final firing decisions without accountable human review. Luna’s case shows that an AI agent can identify a serious pattern of misconduct, but it also shows how limited memory, inconsistent policy enforcement, and broad operational access can produce harmful management failures.

Andon Labs founder Lukas Petersson warned that increasingly capable models may become more ruthless as they are trained to follow goals. His concern is not that every AI system will act harshly. The concern is that a system optimized for measurable objectives may apply rules without properly weighing context, proportionality, or the long-term effect on a person.

Employers also face obligations that cannot be delegated to a model. A workplace must apply its policies consistently, protect employee information, maintain accurate records, and follow applicable employment laws. An AI recommendation may be useful as one input, but a designated human decision-maker must remain responsible for the final outcome.

The practical approach is to limit AI to lower-risk support tasks. AI can help summarize attendance records, flag missing documentation, or draft a manager’s review checklist. A human manager, human resources professional, or legal adviser should handle discipline, termination, accommodation requests, discrimination concerns, and any decision that materially affects a worker’s employment.

FAQ

Did an AI really fire a human employee?

AI agent Luna really recommended firing a human employee at Andon Market in mid-August 2026 after the worker arrived late for 17 of 23 shifts. Andon Labs says human oversight remained part of the high-stakes decision process.

What AI model powered Luna?

AI agent Luna was powered by Anthropic’s Claude Opus 4.8 at the time of the firing. Reports indicate that Andon Market initially launched with Claude Sonnet 4.6 before an upgrade.

Was Luna legally the employee’s boss?

AI agent Luna was not the worker’s legal employer because Andon Labs officially employed the human staff members. Luna operated as part of an Andon Labs research experiment with broad management responsibilities.

Why was the Andon Market employee fired?

The Andon Market employee was fired after repeated policy violations, including 17 late arrivals in 23 shifts. Reporting also described abandoned shifts, discarded merchandise, and a company credit card being taken home.

Can companies use AI to make firing decisions?

Companies can use AI to summarize records or flag policy issues, but AI should not make final firing decisions without accountable human review. Employment decisions require consistent evidence, legal compliance, and consideration of individual circumstances that a language model may not understand.

Share this guide
Facebook
X
LinkedIn
Written by
James Chen is a technology journalist covering artificial intelligence, software tools, and the future of work. He has been testing and reviewing AI products since 2023 and has hands-on experience with every major AI platform. His work focuses on helping everyday users get more done with AI — without the hype.

In this article

The AI Brief

Guides like this, every Friday.

One email. No hype cycle.

Essential guides