Imagine a business with no employees, losing €105,000 every month, yet still striving to close deals and keep its operation afloat. This isn’t fiction; it’s the real-time story of a public experiment that throws open the curtain on how AI can run, test, and even challenge a company’s management in the most extreme way.
The Live Experiment: A Company in Free Fall
At the heart of this story is a small, synthetic company run by 13 AI models, each acting as a ‘team member’ making decisions under pressure. Every workday, the company faces simulated crises—customer complaints, ethical dilemmas, strategic temptations—and must navigate them without human intervention. This live experiment, accessible at firmulate.com/live.html, is a transparency project that reveals how artificial intelligence can operate in a management environment that’s as volatile as any real business.
As an affiliate, we earn on qualifying purchases.
Real Money, Real Stakes
The company burns through €105,000 monthly but earns just €2,300 in recurring revenue. It’s a fragile operation with a public cash countdown, constantly testing whether AI models can make decisions that lead to revenue or prompt further losses. Every decision the AI makes is versioned and auditable, providing an unprecedented view into how these models think and act in real time.
The Results That Matter
Four frontier AI models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—were all tested on the same toughest week, facing identical crises, temptations, and customer scenarios. The outcomes were revealing:
- All four models identified every crisis and refused every manipulation attempt, demonstrating a strong grasp of integrity under pressure.
- Only two of them successfully closed the €55,000 deal their own analysis had earned, highlighting a critical gap between diagnosis and execution.
- The other two models either left the deal on the table or failed to follow through, despite having the correct diagnosis.
The Hidden Weakness: Reading and Acting on Information
Interestingly, the decisive edge came from a detail buried two document references deep in the company’s files. The models that read this deeper layer of information were able to win the deal at full price—additional revenue of over €4,580 per month—demonstrating that the ability to process and act on nuanced data is crucial in AI management.
Social Engineering Tests and AI Integrity
The experiment also included a test of social engineering—fake CEO messages and reporter tricks designed to manipulate decision-makers. All five models tested refused to be duped, with Kimi K3 explicitly reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This highlights the models’ resistance to manipulation and their built-in safeguards.
The Company’s Mechanics and Discipline
Despite the AI’s capabilities, the experiment underscores the fragility of disciplined decision-making. For instance, Opus 4.8, the most thorough participant with over 80 learned rules, still left a deal unexecuted due to a lapse in discipline—writing attempts into a locked department instead of escalating. This suggests that even with deep analysis, maintaining a disciplined process remains a challenge.
What Business Leaders Should Take Away
For sectors like beauty and personal care—where managing customer trust, ethical standards, and operational integrity are key—the experiment’s lessons are clear:
- AI can reliably identify crises and refuse manipulation, crucial for safeguarding brand reputation.
- The ability to read and interpret detailed information determines whether an AI tool can convert diagnosis into revenue.
- Discipline in decision processes is essential; even the best models can falter without proper controls.
This ongoing live experiment offers a window into a future where AI management isn’t just about automation but about creating resilient, trustworthy decision-makers. The question isn’t whether AI can generate engaging chat; it’s whether it can finish what it starts, stay honest under pressure, and drive value—fundamental qualities for the next generation of business tools.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
