Watching an AI-Run Startup Fight for Survival in Real Time

Imagine observing a startup that has no human employees, yet faces the same brutal challenges of cash flow, crises, and tough decisions—live, every workday. For astrophotography enthusiasts, this may sound like a distant universe, but in the world of AI-driven business simulation, it’s happening now. Welcome to a groundbreaking experiment where artificial intelligence models run a real company, navigating crises, making decisions, and even risking financial ruin—all in the open.

The Live Company That’s Watching Itself

At the heart of this experiment is Firmulate, a platform that models a small software company with real money mechanics, 13 synthetic employees, and a public cash countdown. Every business day, the company’s operations are versioned and logged, giving observers unparalleled insight into how AI handles the complexities of running a business. Currently, the company spends €105,000 each month but earns just €2,300 in recurring revenue, illustrating the stark reality of a startup burning through cash while trying to close deals.

Testing the Limits of AI Decision-Making

Four advanced AI models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Opus 4.8—were each tasked with guiding this company through its worst week. They faced the same customer crises, internal dilemmas, and manipulative attempts. The results were revealing: all the models successfully identified every crisis and refused every manipulation, including social engineering tactics like fake CEO messages and reporter tricks. This indicates a strong capacity for ethical safeguards and crisis recognition in AI decision-making.

What Really Made the Difference?

While all models performed well in crisis detection and resisting manipulation, only two closed the deal with a major customer at full price (€55,000). The other two, despite diagnosing the opportunity accurately, left the deal on the table. The critical secret lay in a buried fact—information stored in the company’s own files that the models who read and understood this data won the deal at an additional €4,583 MRR. This underscores a vital point: thorough information access can be the difference between closing or losing a lucrative deal.

Built-in Public, Built for Reality

This experiment is not a demo or a controlled test; it’s happening in real time, with every decision versioned and auditable. The company’s daily operations are openly accessible at firmulate.com/live.html, allowing anyone to watch the AI models in action—making decisions, facing crises, and fighting to survive. The entire setup offers a raw look at how AI can manage complex, high-stakes environments without human intervention.

How AI Resists Social Engineering

In one notable test, social engineering tactics—such as escalating fake CEO messages and subtle reporter tricks—were deployed. Remarkably, all five models refused to be manipulated, with Kimi K3 explicitly treating such requests as potential impersonation or approval-bypass attempts. This resilience demonstrates that AI models trained with proper safeguards can maintain integrity even under pressure, an essential trait for deployment in real business environments.

The Grim Reality of the Business Model

Despite these impressive decision-making feats, the company is currently losing money at a significant rate, with a burn rate of €105,000 per month against a monthly revenue of €2,300. Every workday, the company’s cash position diminishes, creating a public countdown to potential insolvency. This stark reality highlights the experimental nature of the setup: it’s a real business fighting for survival, not just a theoretical sandbox.

Insights from the Performance League

The models are scored on their ability to diagnose issues, close deals, and maintain discipline. The top performer, gpt-5.6-sol, scored 95 points, successfully closing the deal after revealing the buried fact. Kimi K3 scored 93—a newcomer showing the cleanest discipline. Sonnet 5 and Opus 4.8 scored 88 and 77 respectively, with Opus being the most thorough but still leaving opportunities unexploited, such as closing the deal or escalating issues appropriately.

Implications for AI in Business

This experiment vividly demonstrates that AI models can be more than just chatbots—they can be responsible for managing complex operational decisions, resisting manipulation, and even closing high-value deals. But it also exposes the fragility of current AI systems under real-world pressures, especially in financially precarious situations. As AI begins to touch areas like CRM, support, and forecasting, the question is no longer about whether it can generate convincing text but whether it can reliably deliver real, honest work under stress.

Participate and Observe

Business leaders and AI developers can run similar tests using their own data through the platform’s pilot mode, ensuring the AI’s decisions align with their standards before deploying it in live environments. For those interested, the ongoing company story and decision data are openly available, fostering transparency and continuous learning.

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

Powered by Thorsten Meyer AI