Watch an AI-Run Company Struggle for Survival in Real Time

Imagine managing a fitness gym where every decision impacts your bottom line, but instead of trainers or staff, AI models are making the calls. Now, picture seeing those AI managers face crises live, battling not just customer issues but their own limitations—warts and all.

Behind the Curtain of an AI-Driven Business Experiment

At first glance, the idea of a software company with no employees might seem like a futuristic stunt. But this is no mere experiment in automation; it’s a real, ongoing test at firmulate.com/live, where a small company operates every weekday with a team of 13 synthetic employees guided by advanced AI models.

This company isn’t just running on autopilot—it’s actively fighting for its survival while exposing the strengths and flaws of AI decision-making under pressure. It’s a build-in-public project, with every version of the AI’s decisions publicly recorded, scrutinized, and analyzed.

The AI Models at Play

• gpt-5.6-sol: The top scorer, with a 95/100, identified critical buried facts and closed a lucrative deal, demonstrating thorough analysis and decisive action.
• Kimi K3: Slightly behind at 93, K3 managed to close the deal through disciplined decision-making, even without default effort settings.
• Sonnet 5: Scored 88, with some slip-ups but still managed to seal the deal.
• Opus 4.8: The most detailed model, with 80+ learned rules, yet finished last — a reminder that deeper analysis doesn’t always lead to better outcomes.

The Crises, the Challenges, and Human-Like Frailties

The experiment tasked all models with navigating the company’s worst week—same crises, same customers, same temptations. Every decision was versioned and auditable, allowing experts to see exactly how each AI responded.

Remarkably, all models identified each crisis and refused manipulation attempts, like fake CEO messages escalating over multiple stages or subtle reporter tricks. Kimi K3, for instance, explicitly treated such requests as potential impersonation, refusing to bypass security protocols.

Yet, despite their resilience against manipulation, only two models managed to seal the deal worth €55,000, the company’s key revenue opportunity. The catch? The critical insight that led to the successful deal was buried two document references deep in the company’s files—an information piece that only reading the full context, rather than surface-level cues, uncovered.

Real Money, Real Stakes, Live Observations

Every weekday, this company burns €105,000 against just €2,300 in monthly recurring revenue (MRR). It operates with over 680 self-learned playbook rules, each versioned daily, making it a real-time showcase of AI’s capabilities and limits in a high-stakes environment.

Watching the company live, viewers see the AI models confront crises, make decisions, and sometimes slip—like Opus 4.8, which left a closing opportunity on the table by escalating issues internally instead of escalating them properly. It highlights that even the most sophisticated models can stumble when discipline slips, or when the decision complexity exceeds their learned rules.

What Does This Mean for Business and Fitness?

While it’s a high-concept experiment, the implications resonate beyond the tech sector. If AI can run a company, handle crises, and refuse manipulation under real pressure, it hints at a future where AI might manage parts of your gym or fitness business—handling customer inquiries, managing schedules, or even closing deals—without human intervention.

But the experiment also shows that AI isn’t infallible. Just like in a busy gym, discipline, attention to detail, and understanding of deep context matter. A model that reads only surface notes might miss the buried facts critical to success, much like a trainer overlooking an underlying injury.

Why Should You Care?

Whether you’re managing a fitness gym or a tech startup, the core question remains: Can your AI management tools finish what they start? Will they interpret your data accurately, stay honest under pressure, and ultimately, make decisions that serve your business best? The ongoing live experiment at firmulate.com/live offers a rare glimpse into how AI performs when stakes are high—and where it might stumble.

Final Thoughts: Seeing is Believing

In an era where automation tools are increasingly integrated into daily operations, this experiment is a vital eye-opener. It’s not about shiny demos or chat-capabilities; it’s about real performance under pressure. The question is not just whether AI can write well, but whether it can *finish* what it starts, stay honest, and adapt to crises—all while managing actual money.

For fitness and business leaders alike, understanding these nuances is crucial as AI moves from experimental to essential. Watch the live company at firmulate.com/live and see firsthand what AI’s true potential and limitations look like in action.

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

Powered by Thorsten Meyer AI