Watch a Real Company Run on AI — and Fail to Turn a Profit

Imagine a fitness class where every move is monitored, decisions are scrutinized, and your progress is judged daily — but instead of gaining strength, the class burns through €105,000 every month with a mere €2,300 in revenue. Sounds impossible, right? Yet, this is the reality of a live AI-run software company that’s open for the world to watch, with every decision, crisis, and misstep on display.

The Living Experiment in AI Management

This isn’t a typical tech demo or a closed lab experiment. It’s a real, publicly accessible company that operates with 13 synthetic employees, each powered by advanced AI models. Every workday, the company faces the same crises, customer demands, and ethical temptations as any business — but with a twist: its management decisions are all made or influenced by AI models that are observed and scored in real time.

What’s at Stake? Money, Trust, and Survival

The company’s financials tell a stark story. It burns through €105,000 every month, yet earns just €2,300 in monthly revenue. Its public cash countdown is ongoing, and the company’s entire operation is versioned daily — meaning every decision and response is recorded and auditable. This transparency aims to probe whether AI can truly manage complex, real-world business scenarios without slipping into shortcuts or dishonesty.

Testing AI’s Moral and Practical Limits

In this experiment, four leading AI models, including GPT-5.6, Kimi K3, Sonnet 5, and Opus 4.8, each ran the same difficult week. The challenge was to navigate crises, resist manipulation attempts, and close a critical deal with a client. All models recognized every crisis and refused to be manipulated, such as fake CEO messages or reporter tricks. However, only two managed to sign the €55,000 deal that their own analyses identified as the best move — but one of these models hesitated, leaving the deal on the table despite knowing it was the right decision.

What Does It Take to Win?

The key to success was reading deeply into internal documents. The models that won the deal had uncovered a crucial fact buried two document references deep in the company’s files — a detail that was overlooked by others. This demonstrates the importance of thorough information processing, a capability that AI can excel at when given access to the right data.

Ethics and Integrity Under Fire

The experiment also tested social engineering — fake CEO messages and background requests. All five AI models refused to cooperate with these ethically fraught manipulations, with Kimi K3 explicitly reasoning that such requests could be impersonation or approval bypass attempts. This highlights how AI can be trained to uphold integrity even when under pressure.

The Human-Like Failures and Lessons

Despite its advanced capabilities, the company’s latest model, Opus 4.8, was the least successful. With over 80 learned rules and deep analysis, it left a critical deal unclosed, slipping into internal conflict and delaying escalation. This underscores that even the most sophisticated AI can falter under specific pressures, especially if not disciplined or guided properly.

Why This Matters Beyond the Tech World

For those in fitness and exercise, the takeaway is clear: whether you’re tracking your reps or managing a gym, the quality of your tools matters. It’s not just about how well an AI or a machine can generate words or images. It’s about whether it can finish what it starts, stay honest under stress, and take advantage of hidden insights — all crucial for genuine progress and trust.

The Big Picture: Transparency, Trust, and the Future of AI in Business

This live experiment is a rare window into how AI can handle complex, high-stakes management. It’s also a sobering reminder: AI’s true value isn’t just in generating convincing chat or simple tasks, but in reliably executing critical decisions, uncovering hidden opportunities, and maintaining integrity under pressure.

What Can You Do? Run Your Own Wargame

Interested companies can test their own AI management strategies through secure, read-only simulations that replicate their business environments. It’s a way to evaluate whether AI can handle the real-world crises and temptations your business faces — without risking your actual operations.

Final Thoughts

Watching this experiment unfold in real time offers a stark view: AI models are capable of recognizing crises and resisting manipulation, but they’re not infallible. Success hinges on thorough data access, disciplined decision-making, and ethical safeguards. As AI continues to evolve, these insights will be crucial for anyone looking to incorporate intelligent automation into their business, whether in tech, fitness, or beyond.

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

Powered by Thorsten Meyer AI