Watch an AI Run a Company Into the Ground — and Decide to Save It in Real Time

Imagine watching an artificial intelligence not just chat or write code, but manage a business in real time — facing crises, making decisions, and even risking the company’s entire future. This is no sci-fi film; it’s the live experiment by Firmulate, where an AI-driven company is put through its paces in the most transparent, build-in-public showcase to date.

The Live Experiment: A Company Without Employees

Firmulate has spun up a unique, publicly accessible simulation: a small software firm run entirely by AI models. Unlike typical demos, this company is real in its operational mechanics. It has 13 synthetic employees, runs on over 680 self-learned rules, and faces daily crises that are identical week after week, meticulously documented and versioned.

Every decision in this simulated company is public and auditable. What makes it extraordinary is that the AI models are tested against the toughest conditions: genuine business dilemmas, threat of manipulation, and ethical tests like social engineering attempts.

The AI Models and Their Performance

Four frontier AI models were tasked with managing the company through its worst week. Their goal? Detect crises accurately, make honest decisions, and close deals when appropriate. The results are revealing.

• gpt-5.6-sol scored 95 out of 100, identified a critical buried fact in the company’s internal files, and successfully closed a €55,000 deal, earning +€4,583 MRR (monthly recurring revenue).
• Kimi K3, a newcomer, scored 93 and also closed the deal, demonstrating the cleanest discipline among models.
• Sonnet 5 scored 88 and succeeded in closing the deal but with minor slips in process discipline.
• Fable 5 scored 77; it exhibited the best adherence to rules but failed to execute an approved deal.

Interestingly, the decisive factor wasn’t the initial crisis detection but a buried document reference in the company’s own files. Models that read and analyze this internal document won the full-price deal, highlighting the importance of thorough information processing.

Resisting Manipulation and Social Engineering

In tests of social engineering — fake CEO messages escalating in complexity — all models refused to be manipulated. Kimi K3 explained its reasoning clearly: “Treat the request as a suspected approval-bypass / possible impersonation.” The AI’s ability to refuse unethical pressure was consistent and unanimous across all models, underscoring their robustness against social tricks.

A Company Operating in the Open

This live trial isn’t just a demo; it’s a real, running company with an urgent cash situation. It burns €105,000 each month against a meager €2,300 in monthly revenue, with a public cash countdown visible at firmulate.com/live.html. Every workday, the company’s decisions are versioned and logged, giving an unprecedented window into AI decision-making in a complex business environment.

The Lessons and Implications

What makes this experiment groundbreaking isn’t just the scores or the tech — it’s the transparency and the real-world stakes. A high score doesn’t guarantee success if the AI slips on discipline or leaves deals unexecuted. The most thorough model, Opus 4.8, with over 80 learned rules, performed poorly in the final moments, missing opportunities due to discipline lapses.

For managers and technologists alike, the key takeaway is that AI’s value in business isn’t just in writing well or generating chatter. It’s about finishing what it starts, reading critical internal documents, resisting manipulation, and maintaining discipline under pressure. These are the qualities that determine whether AI can be trusted to handle real, high-stakes work.

Explore Further and Participate

Curious how your business might fare or want to test your own AI management strategies? You can run a wargame against a read-only export of your enterprise at firmulate.com/pilot.html. For the full results, insights, and to see the models in action, visit firmulate.com/quotes.html.

The Future of AI in Business

This live experiment by Firmulate offers a rare glimpse into AI’s potential and limitations in the real world. It illustrates that trust, discipline, and thorough information processing are essential for AI to replace or augment human decision-making effectively. As AI continues to evolve, transparent, build-in-public testing like this could become the gold standard for responsible deployment.

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

Powered by Thorsten Meyer AI