Watch an AI-Run Company Battle Crises and Win (or Not) in Real Time

Imagine a car repair shop where every decision, from ordering parts to handling customer disputes, is managed by artificial intelligence. But instead of smooth service, the shop faces the same crises and temptations as any human-run business — yet it’s all happening live, with public scores and transparent performance. Welcome to the world of Firmulate, where an AI-driven company is fighting for survival in real time, and you can watch every move unfold.

At the heart of this experiment is a small software company operated entirely by AI models, with 13 synthetic employees executing daily tasks. While the company burns through €105,000 each month, its revenue lingers at just €2,300 monthly recurring — a stark illustration of a business in crisis. The goal? Test whether these AI models can truly manage complex, high-pressure situations without human oversight, and whether they can be trusted to make honest, strategic decisions.

The experiment, run in real-time and publicly accessible at firmulate.com/live.html, pits four advanced AI models against the company’s worst week — dealing with the same customers, crises, and temptations. Each model is a different frontier version, with varying analysis depth and rule adherence. All decisions are recorded, versioned, and auditable, providing a transparent window into AI decision-making in a simulated business environment.

Are AI Models Up to the Challenge?

The results are striking. All four models successfully identified every crisis that arose — from urgent customer complaints to operational breakdowns. They also refused every manipulation attempt, whether social engineering tricks or subtle bribes. For instance, when fake CEO messages escalated over multiple stages, all models rejected the requests, citing concerns over impersonation and approval bypass risks.

But when it came to closing deals, the story diverged. Only two models managed to sign the €55,000 deal they had identified as the best opportunity. One of those, Kimi K3, did so after a thorough analysis and without succumbing to shortcuts. Interestingly, the decisive advantage was hidden two document references deep within the company’s files — a detail that only reading that context could uncover. Those who discovered and leveraged this buried insight secured an additional €4,583 in monthly recurring revenue.

The Hidden Weaknesses in AI Decision-Making

Among the models tested, Opus 4.8 stood out for its thorough analysis, having learned over 80 rules and conducting deep evaluations. Despite this, it finished the week in last place by failing to close the deal and slipping into less disciplined behavior, such as leaving potential opportunities unexecuted and failing to escalate issues appropriately.

This highlights a key challenge: even highly disciplined AI models can falter under real-world stress, especially if their decision protocols are not aligned with strategic objectives. The experiment shows that deep analysis doesn’t necessarily translate to better outcomes if the AI fails to act decisively or manages its tasks properly.

Implications for the Future of Business AI

This experiment is more than a showcase of AI prowess; it underscores the importance of trust and reliability in AI management tools. For industries like automotive repair, where decisions can impact safety, customer satisfaction, and operational costs, it’s critical that AI systems not only understand context but also execute fully and honestly.

The current leaderboard demonstrates that models like GPT-5.6-sol and Kimi K3 are closest to achieving this ideal, with scores of 95 and 93 respectively. They managed to find hidden information, sign deals, and refuse risky social engineering attempts. Meanwhile, the experiment openly exposes weaknesses, providing a transparent testbed for improvement.

This live company, with its public cash countdown and real money mechanics, exemplifies an extreme form of ‘build-in-public.’ Every decision, every rule learned, every process slip is visible — making it a rare, valuable glimpse into what AI-driven management could look like in the real world.

Watching this AI-managed company handle crises and make strategic decisions in real time reveals both its promise and its current limitations. As AI models improve, their ability to act reliably — especially under pressure and temptation — will be crucial. For industries contemplating AI integration, the key takeaway is clear: it’s not just about AI writing well, but about AI finishing what it starts, reading deeply into data, and staying honest when stakes are high.

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

Powered by Thorsten Meyer AI