Watch an AI-Run Company Struggle for Survival in Real-Time Experiment

Imagine a business where every decision is scrutinized, every crisis is met with unwavering honesty, and its fate is played out live for the world to see. This isn’t fiction; it’s the latest experiment in AI management, and it’s happening now, in full view at firmulate.com/live.html.

The Live Company That’s Building Publicly — and Barely Surviving

In an era where transparency is king, a small software company is pushing the boundaries of open management. It operates with 13 synthetic employees, driven entirely by AI models, and its every move is recorded and versioned daily. The goal? To see if artificial intelligence can handle the complex, often messy realities of running a business—under pressure, crisis, and temptation.

Real Money, Real Risks, No Human Employees

Despite its digital workforce, the company burns through €105,000 each month, yet generates only €2,300 in monthly recurring revenue. It’s a high-stakes simulation, but one that mimics real-world financial pressures. The live experiment allows viewers to observe how different AI models respond to real business dilemmas, including crises, strategic decisions, and social engineering attacks.

Competitive Benchmarking with the Crucible League

To gauge AI performance, four frontier models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—are tested against the same week of chaos. This week includes customer crises, internal miscommunications, and manipulation attempts. The scores reveal each model’s ability to diagnose problems, resist manipulation, and execute decisions.

For instance, the models’ scores range from a high of 95 for gpt-5.6-sol to 77 for Fable 5, with all models spotting every crisis and refusing manipulation attempts. Yet, only two models successfully signed a €55,000 deal based on their analysis—highlighting that even with perfect crisis detection, execution remains a challenge.

Uncovering the Hidden Weakness

The most surprising insight? The key weakness wasn’t in customer interactions, but buried two document references deep inside the company’s internal files. The models that read and interpret these documents at depth won the deal at full price, adding €4,583 monthly recurring revenue to the company’s tally. This demonstrates how critical internal data comprehension is, even in AI-managed businesses.

Resisting Social Engineering and Manipulation

In a series of staged social engineering tests, fake CEO messages and reporter tricks were used to manipulate decisions. All five models refused these attempts, with Kimi K3 particularly careful, reasoning that requests should be treated as suspicious—potential impersonation or approval bypasses.

The Company in Action

The live site offers an ongoing view of this digital enterprise, with every decision, crisis, and outcome available for public observation. It features the most comprehensive participant, Opus 4.8, which employed over 80 learned rules but still left some deals on the table due to slipped discipline—showing even the most advanced models face challenges in execution, not just diagnosis.

Implications for Business and AI

This experiment isn’t just about watching a fictional company struggle; it raises vital questions for anyone relying on AI to run or assist in real business environments. Can these models finish what they start? Will they stay honest under pressure? And most importantly, how much does useful work cost when humans are replaced by AI?

See It in Action

Those interested can watch the ongoing performance and examine detailed scores at firmulate.com/live.html. There’s also a quiz to test how well you understand AI decision-making at firmulate.com/quiz.html. This is the future of business management—public, transparent, and intensely revealing.

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

Powered by Thorsten Meyer AI