Watch an AI Run a Company Live — and See How It Fights for Survival

Imagine watching a company operate in real time, with no employees but a handful of AI models making decisions, handling crises, and even closing deals — all live on the internet. This is not a fantasy; it’s the bold experiment of Firmulate, a public showcase of artificial intelligence managing a tiny software business as it battles for survival, every single workday. The stakes? Burn €105,000 a month against a mere €2,300 recurring revenue, with the entire process transparent for all to see.

The Experiment: AI as a Corporate Manager

At the heart of this project is a small, synthetic company run by 13 AI-powered ’employees.’ These aren’t typical chatbots; they are decision-making models trained to react to real crises, manage client relationships, and even negotiate deals. Every decision they make is versioned and auditable, creating a transparent trail of how they handle the company’s toughest week. The purpose? To see whether AI can truly manage complex, high-stakes business scenarios without human intervention.

Benchmarking AI Against Business Challenges

The experiment pits four frontier AI models against each other — each given the same challenging week filled with customer crises, internal dilemmas, and manipulation attempts. All models successfully identified every crisis and refused every manipulation — a promising sign. Yet, only two of them managed to close a critical deal worth €55,000, demonstrating real economic success. Interestingly, the models that read deeper into the company’s internal files outperformed those that relied solely on surface information, highlighting the importance of thorough analysis.

A Deep Dive into Decision-Making

Among the models, Opus 4.8, the most detailed participant with over 80 learned rules, showed the deepest analysis but fell short at the closing stage. It left the deal unexecuted, showing that even in high discipline environments, oversight can happen. Meanwhile, the newer model Kimi K3, which ran without an effort parameter, exhibited the cleanest decision process and successfully closed the deal. This points toward a nuanced understanding: discipline and thorough analysis matter, but so does the way the AI is configured.

Real Money, Real Risks, Live and Unfiltered

The live company isn’t just an experiment; it’s a functioning entity with actual cash flow. It burns €105,000 monthly against a modest €2,300 recurring revenue, with a public cash countdown visible at firmulate.com/live.html. Every day, the system is versioned, and decisions are logged, creating a continuously evolving record of how AI manages a business under pressure. This transparency allows observers to assess whether AI can be trusted in real-world applications — not just in chat demos but in actual operational environments.

What It Means for the Future of Work

For business leaders, the lessons are clear: the question isn’t simply whether AI can generate good text, but whether it can complete complex, high-stakes tasks honestly and reliably. The experiment shows that AI can spot crises, resist manipulative tactics, and even close deals — but only if it’s carefully guided and tested beforehand. Running a company with AI isn’t a matter of replacing humans overnight; it’s about understanding its strengths and weaknesses in real scenarios before deployment.

The Broader Implication: Trust and Accountability

The experiment also underscores the importance of transparency. With every decision versioned and auditable, firms can see exactly how AI models behave in critical moments. This built-in accountability is essential if AI is to be trusted with sensitive business operations, especially as it starts to touch areas like customer support, sales, and strategic planning.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI