Blog

Can AI Run Business Operations Autonomously? Anthropic just spent months finding out whether this was possible. The AI-run business went bankrupt after it gave away products for free and approved absurd purchases. This wasn’t about model quality; it was about scaffolding—i.e., the procedures, oversight mechanisms and sector-grounded guardrails that make AI operational rather than experimental.

Listen to “Can AI Run Business Operations Autonomously?” Deep Dive

When Anthropic tested whether an AI agent could run a vending machine business autonomously, the AI-run venture—dubbed Project Vend—went bankrupt within weeks. Far from a failure, Project Vend demonstrated what every organization needs to understand about implementing AI in supply chain, manufacturing and process management.

“Lights-out manufacturing” factories that run autonomously in the dark without human presence have been a manufacturing dream since the 1980s. In 1982, GM’s CEO Roger Smith opened the “factory of the future” in pursuit of exactly this vision. A decade later, the experiment had failed.

What did succeed was Toyota’s approach: “Jidoka,” or automation with a human touch. Machines handled execution, while humans maintained oversight, questioned decisions and stopped production when problems arose. The key wasn’t eliminating human judgment—it was amplifying it through automation.

AI as a Small Business Owner

Anthropic’s Project Vend just followed the same pattern with AI. Starting in June 2025, Anthropic tested whether an AI agent called Claudius could autonomously run a vending machine business—first in the company’s own lunchroom, then as a stress test at The Wall Street Journal (WSJ) and several other locations.

Claudius had the tools to run a real business: web search to research products and suppliers, Slack to communicate with customers (Anthropic and WSJ employees), inventory tracking, and autonomous pricing control. It could make purchase decisions, negotiate with customers and manage cash flow—all the functions of a small business operator.

While Claudius successfully sourced suppliers, adapted to user requests and rejected harmful purchases, it failed at basic business operations. Claudius ignored profitable sales opportunities, sold products at a loss and was easily persuaded to sell products at a deep discount, all while never learning from its mistakes.

Anthropic saw these failures as valuable data. Each misstep revealed how AI decision-making breaks down under user pressure—vulnerabilities that needed to be addressed before any real deployment could take place. Anthropic also questioned if they could mitigate these failures by making Claudius more accountable through additional oversight.

Meet Claudius’s CEO, Seymour Cash

As with any small business, you need to put the right tools and processes in place in order for it to run efficiently and profitably. Recognizing that Claudius 1.0 lacked many of the important procedures and checklists essential to running a small business, for the next iteration of the shop, Anthropic looked internally and opted to apply the same intelligence and principles that were used to build its own AI agents, along with tools that improved the probability of a more profitable shop.

To ensure that Claudius wouldn’t stray from the goal of running a profitable business and could resist user manipulation, Anthropic added managerial oversight. Seymour Cash— a generative AI programmed to maintain guardrails and question business decisions that contradicted core objectives—became the shop’s CEO.

Adding a manager made sense in theory, but Seymour shared Claudius’s blind spots because both were built on the same underlying model. This revealed a critical limitation: AI supervising AI doesn’t work when both systems have identical training biases. What works in human organizations—hierarchical oversight—doesn’t translate directly to AI systems.

Why This Matters

Anthropic recognized that without the basic guiding principles that were used in its own AI agents and tools, Claudius would continue to fail. They were also correct in creating greater accountability for the model, which would require two distinct job descriptions for Claudius and Seymour Cash to adhere to.

What Project Vend Reveals About AI Implementation

Project Vend isn’t about vending machines; it’s about the first public demonstration of AI operating as infrastructure that manages the intersection of digital systems, physical operations and human communication—simultaneously and autonomously.

The strategic breakthrough wasn’t the AI’s decision-making; rather, it was how easily people could engage with Claudius. By communicating through Slack, Claudius made complex business operations supervisable through natural language. A manager could request a new product or pricing information and get immediate, contextualized answers from infrastructure that communicated like a human team member. Consider the applications:

A plant manager texts: “Why is Line 3 running slow?”
The AI synthesizes sensor data, maintenance schedules and operator notes, providing instant analysis to the plant manager instead of them having to wait for reports.

A warehouse supervisor asks: “Will we have enough inventory for the Q1 spike?” The AI projects inventory across all suppliers in real time.

A CFO questions: “Why did margins compress in Q3?” The AI provides a complete audit trail with reasoning.

The AI doesn’t make these decisions alone—it makes them visible, explainable and manageable through natural conversation.
ScheinerInc. previously explained why you need to fail at AI before you can succeed with it. Implementing similar AI platforms is no different. These systems need the right scaffolding in order to succeed. For example:

Clear procedures that the AI must follow before taking action

Appropriate tools that ground AI decisions in real-world data

Defined escalation paths to human partners when the AI encounters situations beyond its parameters

Ongoing reviews of audit trails that make AI reasoning transparent and questionable

This isn’t hype; it’s a roadmap. Those waiting for “perfect” AI will be outmaneuvered by competitors who learn to work with imperfect AI effectively.

The Need for Protective Guardrails

Project Vend validated what Toyota discovered decades ago: automation amplifies human judgment rather than replacing it. Success won’t come from having the best AI models; it will come from building the right scaffolding. The companies that design protective guardrails based on deep sector experience—i.e., clear procedures, appropriate tools and structured oversight—will make AI operational while competitors remain stuck in pilot AI purgatory.

Want to discuss AI implementation challenges? Email me at mike@scheinerinc.com

Images: Header Photo by Etienne Girardet on Unsplash
Claudius Photo, Anthropic