Imagine a startup with no employees, yet battling daily to survive — burning over €105,000 every month against just €2,300 in recurring revenue. This is not fiction, but a real-time experiment that offers a rare, unfiltered look at how artificial intelligence can manage a company on the brink. Welcome to the world of Firmulate, where every decision, crisis, and temptation is publicly recorded and scrutinized.
The Live Experiment: A Company in Crisis, Powered by AI
At the heart of this experiment is a fully operational, yet heavily strained, micro-company run by 13 synthetic employees. These digital workers handle real money mechanics, with a stark cash countdown that makes clear the company’s precarious position. Every workday, the company’s decisions are versioned and made public, creating a transparent laboratory for studying AI management under pressure.
How It Works
The experiment pits four frontier AI models against the same challenging scenario: running a small software business during its worst week. Each model faces identical crises, customer demands, and ethical temptations, with their decisions fully auditable. The goal? To see which AI can best navigate the turmoil and finish what it starts.
The Key Findings
- All four AI models detected every crisis and refused every manipulation attempt, demonstrating a baseline of honesty and crisis awareness.
- Only two models managed to close a deal worth €55,000, the same analysis that identified the opportunity and pitched it convincingly. The other two did not sign because they left crucial information unread or unacted upon.
- The difference wasn’t in their diagnoses but in their discipline. For example, the most thorough model, Opus 4.8, analyzed deeply but faltered at closing due to discipline lapses, such as writing attempts into a locked department instead of escalating.
- Surprisingly, a buried fact in the company’s own files—hidden two document references deep—was the decisive factor in winning a high-value deal. The models that read deeper won the full-price contract, adding over €4,583 monthly recurring revenue.
Resisting Social Engineering
The experiment also tested whether the AI would fall for social engineering tricks, such as fake CEO messages and staged reporter requests. All models refused these manipulative tactics, with Kimi K3 explicitly treating suspicious requests as potential impersonation, highlighting their resistance to deception.
As an affiliate, we earn on qualifying purchases.
The Reality of Building Public AI Management
This isn’t just an academic exercise. The firmulate.com/live site offers a live, transparent window into this ongoing struggle, making it a pioneering example of build-in-public in its most extreme form. The company burns €105,000 each month with only €2,300 in recurring revenue, illustrating just how delicate the balance is when relying on AI for real-world management.
What the Results Mean for Business
In a landscape where AI is increasingly integrated into customer service, sales, and decision-making, these findings raise critical questions. Will AI agents be able to finish what they start, stay honest under pressure, and read critical information buried deep in documents? According to the experiment, the answer depends heavily on both the model and the discipline built into its rules.
Broader Implications
- The experiment shows that AI can identify all crises and reject manipulation — but winning deals requires not just diagnosis but disciplined execution.
- The buried fact in the company’s files was a game-changer, emphasizing the importance of thorough information reading and analysis.
- Even the most rule-abiding model, Kimi K3, which ran without effort parameters, performed strongly but still missed opportunities, demonstrating that discipline and depth matter.
- Ultimately, the experiment underscores that AI’s value in management isn’t just about writing well but about finishing tasks, reading deeply, and resisting shortcuts or deception.
The Future of AI-Managed Companies
What does this mean for the future of work? For businesses considering AI integration, the experiment suggests a need for rigorous testing before deployment. Management quality isn’t measured in chat quality but in tangible outcomes — closing deals, reading critical info, resisting manipulation, and managing resources.
Visitors can see this experiment unfold live at firmulate.com/live. Here, the company’s daily struggles are laid bare, offering a rare glimpse into AI’s potential and limitations in real-world management.
This live experiment reveals that AI models can detect crises and resist deception but still struggle with closing deals and disciplined execution. It’s a vivid window into AI’s emerging role in managing real companies on the edge of survival.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
