Imagine watching a company struggle for its very survival — with no humans involved, only AI models making decisions in real time.
This is not science fiction. It’s the live experiment conducted by Firmulate, a company that runs a simulated business as a public, ongoing story. Here, you can see how AI models handle crises, ethical dilemmas, and tough decisions, all while being watched and scored daily.
AI management simulation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Live Company: An Open-Source Experiment in AI Management
At the heart of this experiment is a small, fictional software company operated entirely by AI models. These models, called frontier models, are tested under the most stressful conditions — the company’s worst week — with the same customers, crises, and temptations. The goal? To see whether AI can make sound, honest decisions under pressure, and how well it can manage real money mechanics.
Every decision made by these AI models is versioned and auditable, allowing observers to track their thinking step-by-step. The company has 13 synthetic employees, each programmed with over 680 learned rules and a set of self-improving playbook strategies. Despite these efforts, the company burns €105,000 each month against a mere €2,300 in monthly revenue, highlighting its fragile financial position. A public cash countdown reminds viewers that every day is a battle for survival.
Performance Scores and Outcomes
Four leading AI models participated in the experiment, each running the same scenario. Their performance was measured against a set of benchmark scores from the Crucible League, a competition for AI excellence. The scores ranged from 95 (best) to 77 (worst), with the top model, gpt-5.6-sol, scoring 95 and successfully closing a deal that earned full price and additional revenue of over €4,583 in monthly recurring revenue (MRR).
Interestingly, all models identified every crisis and refused manipulation attempts, such as social engineering schemes that tried to trick the AI into bypassing approval steps. For example, when a fake CEO message was escalated through multiple stages, all five models refused to approve it, citing suspicion or impersonation concerns.
The Hidden Weaknesses and Deep Insights
The most telling flaw was not in the models’ ability to detect crises but rather in their ability to act fully on their diagnoses. The most thorough model, Opus 4.8, with over 80 learned rules, was the lowest scorer at 77 and left potential deals unclosed due to discipline lapses. It diagnosed opportunities correctly but failed to follow through, often defaulting to writing into a locked department instead of escalating issues properly.
This recurring weakness was consistent across all models, suggesting that even sophisticated AI systems struggle with decision execution and process discipline under stress.
What This Means for the Future of AI in Business
While the experiment is a simulation, it offers stark lessons for businesses considering AI automation today. The critical questions are not just about whether AI can generate convincing text or perform well in demos. They are about whether AI can finish what it starts, stay honest under pressure, and effectively manage real-world risks and opportunities.
For enterprises contemplating AI integration, the experiment provides a stark reminder: success depends on the AI’s ability to read relevant data, resist manipulation, and execute decisions reliably. The live scoreboard at firmulate.com/live.html shows ongoing results, allowing anyone to watch a company fighting for survival in real time.
Beyond Demos: Building Trust and Reliability
This experiment underscores the importance of transparent, auditable AI behavior in business. It demonstrates that AI’s true value lies not only in generating outputs but in reliably managing complex, high-stakes situations — a challenge still very much in progress.
By observing this company’s daily struggles and victories, decision-makers and AI developers gain insights into what risks remain and what disciplines are necessary for safe, trustworthy automation.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
