Imagine managing a small business where every decision, crisis, and ethical test unfolds live—without any human employees. Now, picture witnessing AI models navigating this chaos, with some closing deals and others faltering under pressure. This is not science fiction; it’s the ongoing experiment at Firmulate, a public showcase of AI management in action.
The Live Company Experiment: AI in the Driver’s Seat
At the heart of this bold venture is a small, simulated software company operated by 13 synthetic employees and driven by real money mechanics. The company burns through €105,000 every month while earning just €2,300 in monthly recurring revenue—making it a high-stakes testbed for AI decision-making under financial and ethical pressure. Every workday, the company’s operations are versioned and publicly visible, providing a transparent window into how AI models handle complex management challenges.
As an affiliate, we earn on qualifying purchases.
How the Experiment Works
Four frontier AI models—each representing a different approach—are put through the same grueling week of crises, customer interactions, and ethical tests. These models are examined for their ability to identify problems, make sound decisions, and resist manipulation or deception attempts. Interestingly, all four models successfully detected every crisis and refused every manipulation, including social engineering: fake CEO messages and reporter tricks.
However, their ability to close deals varied. Two models, gpt-5.6-sol and Kimi K3, were able to sign the €55,000 deal their own analysis had identified as the optimal choice, earning full additional revenue worth over €4,500 monthly recurring. The other two—Sonnet 5 and Fable 5—missed the opportunity, leaving potential profit on the table due to process slips or unexecuted approvals.
What Do These Results Reveal?
The key finding? The decisive weakness in the company’s management wasn’t in crisis detection or ethical resistance but buried two document references deep in its files—details that models capable of reading and understanding these documents went on to close the deal at full price. This illustrates how subtle informational advantages can make or break outcomes, even in high-pressure environments.
The Ethical and Practical Stakes
Throughout the experiment, all models refused manipulation attempts, such as escalating fake CEO messages or background approval requests. Kimi K3, in particular, explained its reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This demonstrates a level of ethical reasoning and cautious judgment that is often missing in typical AI demos.
The Real-World Implications
For companies deploying AI today, the question is not just about how well an AI writes or responds but whether it can truly complete what it starts—reading relevant documents, resisting manipulation, and acting ethically under pressure. This experiment highlights the importance of building AI systems that do not just produce plausible responses but adhere to disciplined, goal-oriented decision-making processes.
The Ongoing Saga of a Company in Crisis
As of now, the live experiment runs in real-time at firmulate.com/live, with the company burning €105,000 monthly against minimal revenue, under the constant threat of running out of cash. Every decision, crisis, and ethical test is versioned and publicly available, offering unparalleled transparency into AI management capabilities and vulnerabilities.
Conclusion: Why This Matters
This extreme build-in-public experiment demonstrates that AI management can be more than just chatbots or simple automation. It can serve as a real-time testing ground for ethical judgment, decision-making under pressure, and strategic thinking—all critical skills for AI to function effectively in complex business environments. As AI models continue to evolve, their ability to navigate these challenges will determine whether they become trustworthy partners or unpredictable liabilities.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
