Imagine an automotive dealership managed entirely by artificial intelligence, making decisions in real time. But what if that AI is running a company that’s losing money every day and on the brink of collapse? This isn’t a sci-fi scenario — it’s the live experiment conducted by Firmulate, where an AI-powered company is openly battling for survival, revealing what AI can and cannot do in the real world.
The Real-World AI Company: A Live Business Under Stress
At the heart of this experiment is a small software company, run by a set of 13 synthetic employees. Every workday, it faces genuine crises, real money mechanics, and the temptation to cut corners. The company is publicly losing €105,000 each month against a modest recurring revenue of €2,300. Despite this, it’s a vivid, open window into how AI models handle complex management tasks in a high-stakes environment.
What makes this experiment extraordinary is its transparency. Every decision made by the AI models is versioned, auditable, and observable at firmulate.com/live.html. You can see the real-time cash countdown, watch the daily decisions, and even listen to the self-learned rules guiding the AI’s actions. This build-in-public approach offers a rare, detailed look at AI’s management capabilities — and limitations — in a real business context.
How Do the AI Models Perform?
Four frontier AI models were put through a standardized test: the worst week in the company’s life. Each faced identical customer crises, internal dilemmas, and manipulation attempts. Remarkably, all four models identified every crisis and refused every attempt at manipulation, including sophisticated fake CEO messages designed to escalate decisions. This shows a strong capacity for honesty and crisis recognition.
However, when it came to sealing deals, results varied. Only two models signed the €55,000 contract their own analysis suggested was the right decision. The other two knew the deal was right but failed to follow through, leaving potential revenue on the table. This gap, buried two document references deep in the company’s files, was decisive — reading the right information at the right time led to the full €4,583 monthly recurring revenue boost.
The Human-Like Decision Making and Flaws
The experiment also included a social engineering test, where fake CEO requests escalated over three stages plus a media trick. All models refused these requests, showcasing integrity in complex scenarios. But not all decisions were perfect. For instance, Opus 4.8, the most thorough model with over 80 learned rules, finished last because it slipped discipline and failed to escalate critical issues properly.
Interestingly, the models’ discipline varied based on their configuration. Kimi K3 ran without an effort parameter, resulting in a cleaner decision process, while the others with higher effort settings showed more slippage.
Building AI-Powered Products: The Essential Guide to AI and GenAI Product Management
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Bigger Picture: AI in Business Management
This live experiment underscores that when AI is entrusted with management, accuracy and honesty aren’t enough. The ability to read and interpret critical internal documents, stay disciplined under pressure, and follow through on opportunities is just as vital. In this case, the only way for the AI to close a significant deal was by reading the buried document reference, a task that proved crucial to revenue generation.
For industries like automotive dealerships, where managing complex customer relationships, inventory, and internal processes is critical, the question isn’t whether AI can generate convincing chat responses. It’s whether it can handle real management challenges — reading internal files thoroughly, resisting manipulation, and closing deals effectively, even in tough circumstances.
The Future of AI in Business
While these models are still far from perfect, the experiment shows promise. The models’ refusal of manipulative tactics indicates a level of integrity, and their ability to spot critical information suggests they can augment human decision-making. However, the gaps in follow-through and discipline highlight that AI is not yet ready to replace human oversight entirely.
The live site at firmulate.com/live.html offers a unique chance to watch this unfolding story. You can see decisions in real time, read the detailed rules learned by the models, and even test your management skills in a quiz at firmulate.com/quiz.html.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
