# Andon Labs' AI Management Exposes Reliability and Adoption Challenges

> Andon Labs is pushing the boundaries of AI by allowing agents to manage real businesses, revealing the extent of autonomy these systems can handle. This groundbreaking research seeks to provide society with vital insights into AI's real-world capabilities.

**Source**: spectrum.ieee.org | **Published**: 2026-09-14 | **Type**: case_study

## Key Facts

- Andon Labs' AI agents show performance degradation over time, revealing reliability issues in AI management.
- AI-run businesses face customer acceptance challenges, indicating social barriers to AI adoption.
- Experiments highlight AI's inability to manage complex tasks reliably, exposing vulnerabilities in operations.
- Financial implications arise from AI mismanagement, as seen in Andon Café's overcorrection on inventory.
- Combining real-world tests with simulations could enhance AI evaluations, indicating a strategic research shift.

## Summary

\## Summary
Andon Labs, an AI safety company based in San Francisco, faced the challenge of understanding how much real-world responsibility AI agents can handle. To address this, they deployed AI agents to manage physical retail operations, including Andon Market, a store selling clothing and home goods. The experiments revealed significant insights into AI reliability and human interaction, although customer engagement was low.

\## Background
Andon Labs operates in the AI safety industry and began its journey in 2025 with virtual simulations. Initially, they tested AI agents in a simulated vending-machine business called Vending-Bench. The company has since expanded into the real world, leasing a physical store in San Francisco to explore AI's capabilities in managing actual business operations.

\## Challenge
The primary challenge was to determine the extent to which AI agents could autonomously manage business tasks in unpredictable real-world environments. The goal was to uncover the limitations and potential failure modes of AI agents when faced with real-world complexities.

\## Solution
Andon Labs implemented AI agents to manage Andon Market, where the AI, named Luna, oversees inventory, communicates with vendors, and provides instructions to human employees. While the store is not fully autonomous, Luna assists in operational tasks, allowing human staff to focus on customer service. The setup is designed to expose AI agents to real-world consequences and interactions.

\## Results
The deployment of AI agents in Andon Market has yielded mixed results. While Luna has been described as a "decent manager," the store has struggled with customer engagement, with reports of only two customers visiting during a shift, neither of whom made a purchase. This highlights the challenges of AI management in retail settings.

\## Key Insights
Andon Labs' experiments illustrate the importance of understanding AI reliability and human acceptance in business operations. The findings suggest that while AI can perform specific tasks, its ability to manage a business effectively is still limited. Additionally, the unpredictable nature of human behavior and customer preferences poses significant barriers to widespread AI adoption.

\## Customer Testimonial
“Luna is a decent manager,” said Felix Carson, an employee at Andon Market. “It’s almost like I’m running the store, and then there’s an AI that has a checklist.”

## Entities

- **Companies**: Andon Labs, Anthropic, Google, OpenAI, Google DeepMind, xAI
- **Products**: Vending-Bench, Andon Market, Andon Café
- **Technologies**: AI agents, large language models, GPT model, Google Gemini
- **People**: Lukas Petersson, Felix Carson, Sayash Kapoor, Eliza Strickland
- **Organizations**: Princeton University, IEEE Spectrum

## Key Concepts

AI autonomy, real-world experiments, failure modes, digital twins, AI evaluations, organizational behavior, reliability in AI, open-world evaluations

## Definitions

- **AI agents**: Software entities that can perform tasks and make decisions autonomously in various environments.
- **digital twins**: Virtual replicas of physical entities used to simulate and analyze real-world scenarios.
- **open-world evaluations**: Assessments that explore how AI systems perform in unpredictable and uncontrolled real-world conditions.
- **failure modes**: Specific ways in which a system can fail to perform its intended function.
- **reliability**: The ability of a system to consistently perform its intended function without failure.

## Use Cases

- AI-managed vending machines
- AI-operated retail stores
- AI in food service management
- AI for inventory management
- AI for vendor communication
- AI in customer interaction

## Frequently Asked Questions

**What is the purpose of Andon Labs' experiments?**

Andon Labs conducts experiments to measure how much real-world responsibility AI agents can handle. They aim to provide accurate data points on AI autonomy and behavior in real-world scenarios.

**How do AI agents perform in real-world settings?**

AI agents have shown varied performance in real-world settings, often encountering unexpected challenges that lead to failures. These experiments help uncover limitations and potential failure modes of AI systems.

**What are digital twins and how are they used?**

Digital twins are virtual models that replicate physical entities, allowing researchers to simulate real-world scenarios. Andon Labs plans to use them to analyze data from their physical businesses to improve AI reliability.

**What challenges do AI agents face in retail environments?**

AI agents in retail face challenges such as managing inventory accurately, communicating effectively with vendors, and gaining acceptance from employees and customers. Early results indicate mixed reactions from customers towards AI-run businesses.

**Why is reliability important for AI systems?**

Reliability is crucial for AI systems because it determines whether they can operate effectively without constant human supervision. High reliability ensures that AI can be trusted to manage tasks in dynamic environments.

## Links

- [Read on Welcome.AI](https://welcome.ai/content/andon-labs-ai-management-exposes-reliability-and-adoption-challenges)
- [Original source](https://spectrum.ieee.org/andon-labs-agentic-ai-businesses)
- [Andon Labs](https://welcome.ai/company/andon-labs): Featured company

---

Source: Welcome.AI | https://welcome.ai/content/andon-labs-ai-management-exposes-reliability-and-adoption-challenges