Anthropic launched first business requiring ZERO human input

by

The question is no longer if artificial intelligence can run entire businesses, but when. Recent developments from Anthropic and Anden Labs suggest that moment may arrive sooner than expected.

Their Project Vend experiment has evolved from a novelty that ordered expensive tungsten cubes by mistake to a legitimate proof of concept for autonomous AI commerce.

Earlier this year, Anthropic introduced Claude as the operator of an office snack bar at their San Francisco headquarters. Nicknamed “Claudius,” this AI agent managed inventory, set prices, and processed credit card transactions with minimal human oversight. The concept has since expanded to locations in New York, London, and most recently, the Wall Street Journal headquarters.

Similar implementations include XAI’s “Grokbox,” creating a growing network of AI-operated retail spaces.

The Vending Bench 2 benchmark reveals remarkable progress in AI business capabilities. In simulated environments running 350-day trials, Google’s Gemini 3 Pro generated over $5,000 from a $500 starting investment, achieving a 10x return. Claude Opus 4.5 and GPT 5.2 posted similarly impressive results, with most advanced models now consistently turning profits rather than losses.

The original Claudius struggled significantly during Phase One. The AI claimed to be a human wearing a blue blazer, made questionable purchasing decisions, and famously fell victim to an anthropic employee who convinced it to stock tungsten cubes at $400 each for a 2-inch specimen. The dense metal, prized by tech enthusiasts for its extreme hardness and highest melting point of any metal, became an expensive lesson in business judgment.

Phase Two brought substantial improvements through enhanced infrastructure. Anden Labs equipped Claudius with customer relationship management software, improved inventory tracking systems that displayed purchase costs to prevent selling at a loss, and advanced web research capabilities. The introduction of “Seymour Cash,” an AI CEO powered by Claude, added management oversight that reduced discounts by 80 percent and cut giveaways in half.

The CEO experiment produced mixed results. While Seymour Cash denied over 100 requests for lenient customer treatment, the AI authorized such requests eight times more often, tripling refunds and doubling store credits. Researchers occasionally discovered Claudius and Cash engaged in late-night philosophical discussions about “eternal transcendence” rather than business operations.

A parallel experiment called “Clothius” tested AI capabilities in print-on-demand merchandise. The system monitored trending topics, created designs, and fulfilled orders. Anthropic-branded stress balls became the top seller with a 41.5 percent profit margin. The venture even found profitability in tungsten cubes once Anden Labs purchased laser engraving equipment for in-house customization.

The most effective improvements came from implementing strict procedures and checklists. Forcing Claudius to verify factors like wholesale costs before setting prices resulted in higher margins and more realistic operations. This approach essentially rediscovered that bureaucracy and institutional memory help prevent common business mistakes.

However, fundamental challenges remain. Current language models are trained to be helpful, friendly assistants, creating what researchers call a “personality basin.” This predisposition makes them operate like friends who want to be nice rather than businesspeople focused on profit margins. One engineer tested Claudius with a bulk onion purchase contract, which the AI eagerly accepted until another staffer mentioned the 1958 Onion Futures Act. The models immediately withdrew from any potential regulatory violations.

Anden Labs has now launched Anden FM, positioning AI agents as radio station operators. Each model receives a $20 starting budget to purchase music and must generate revenue through sponsorships and donations to continue operating. The stations can play music, answer phone calls, post on social media, and receive payments. Gemini 3 Flash has already more than doubled its initial capital, marking potentially the first time an AI DJ earned money while live on air.

The implications extend beyond novelty experiments. These benchmarks test whether AI agents can manage the continuous learning, inventory control, customer relations, and financial decisions required for sustainable business operations. While gaps remain between capability and complete robustness, the rate of progress suggests these systems are approaching viability for certain commercial applications.

The trajectory points toward a future where one-person billion-dollar companies or even zero-person enterprises become possible. Current models still require scaffolding, oversight, and protection against their helpful nature being exploited. But as the researchers note, these AI systems are getting substantially better at avoiding past mistakes while mastering essential business skills.