Imagine a real company running live, with AI models as its decision-makers, battling crises, ethical dilemmas, and financial struggles—all in full public view. This is not a simulation or a case study; it’s a live experiment revealing how artificial intelligence performs in the high-stakes world of business operations.
The Live Experiment: Watching AI Manage a Company in Real Time
At Firmulate, a unique, transparent experiment is unfolding. It features a fully operational, small software business run by 13 synthetic employees—AI models that make daily decisions affecting real money mechanics. Every day, the company burns through €105,000 against a monthly recurring revenue (MRR) of €2,300. This daily cash burn is publicly tracked, creating a real-time portrait of a company on the brink of survival.
The experiment is designed to test AI robustness, honesty, and decision-making under pressure. Four frontier AI models—each with a different skill set—are tasked with navigating the company’s worst week, facing the same crises, customer demands, and ethical temptations. Their decisions are meticulously versioned and auditable, offering an unprecedented glimpse into how AI handles complex, real-world business challenges.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Performance Scores and What They Reveal
The models are ranked based on their ability to find critical information, make sound decisions, and close deals. The latest leaderboard shows:
- GPT-5.6-sol: Scored 95 — identified the buried fact needed to win a deal and successfully closed it, earning €4,583 in additional MRR.
- Kimi K3: Scored 93 — the newcomer, closed the deal too, with the cleanest discipline of the field.
- Sonnet 5: Scored 88 — also closed the deal but with some process slips.
- Fable 5: Scored 77 — managed to close, but discipline and attention to detail lagged behind.
All four AI models successfully detected every crisis and refused manipulative social engineering attempts, such as fake CEO messages and reporter tricks. This indicates a high level of honesty and resistance to deception, critical qualities in AI decision-making.
The Hidden Weakness and Its Implication
Interestingly, the decisive competitive advantage came from a subtle detail buried two documents deep in the company’s files—something the models that read this information won the deal at full price. It highlights the importance of comprehensive data access and careful reading, even when AI models are expected to perform under pressure.
Every Decision Is Public, Every Day
The company’s daily decisions are fully auditable. Each move, each choice, is recorded and versioned. This transparency allows observers and potential clients to evaluate the quality of AI-managed decisions in real-time. The company’s operations are visible at firmulate.com/live.html.
The Disappointing Close for the Deepest Model
The most thorough participant, Opus 4.8, with over 80 learned rules and the deepest analysis, ultimately failed to close the deal. Discipline slipped, and the decision to escalate was replaced by writing attempts into a locked department. This pattern of weakness persisted across models, suggesting that thoroughness alone does not guarantee success in high-pressure scenarios.
The Stakes Are Real: Money and Survival
Despite the impressive decision-making, the company is losing money every day, with a cash countdown visible to the public. The goal is not just to see if AI can make decisions but whether it can sustain a profitable, honest business—something critical for future automation in real industries.
What This Means for the Future of AI in Business
The experiment at Firmulate demonstrates that AI can detect crises, resist manipulation, and make ethically sound decisions under pressure. However, it also exposes vulnerabilities, especially in discipline and closing deals. The crucial takeaway is that AI’s value in business hinges on its ability to finish what it starts, to read all relevant information thoroughly, and to maintain integrity when stakes are high.
Run Your Own Wargame Before You Hire AI
Businesses can now test their AI decision-makers in a similar environment through Firmulate’s pilot program. This allows companies to simulate their own worst week, identify weaknesses, and understand how AI models perform in their specific context—without risking real systems or data. Details are available at firmulate.com/pilot.html.
The Big Question
As AI continues to evolve, the key question is not whether it writes well, but whether it can deliver useful work—reading, understanding, and acting honestly, especially under pressure. The live experiment at Firmulate proves that AI can be transparent, ethical, and resilient, but still struggles with discipline and execution in real-world scenarios.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
