Imagine a company with no human employees, yet facing daily crises, tough decisions, and a public cash countdown — and you can watch it all unfold live. This isn’t science fiction; it’s a groundbreaking experiment in AI management, where the company’s entire operation is run by AI models, tested against real-world pressures, and publicly documented every step of the way.
The Live AI Company: Reality TV Meets Business Management
In a bold move to explore how artificial intelligence can truly manage a business, a real software company is now operating with a twist: it has no human employees. Instead, 13 synthetic team members, powered by advanced AI models, handle its daily functions, from decision-making to crisis management. The company is publicly running a high-stakes experiment, and every decision—big or small—is versioned and auditable. You can follow along at firmulate.com/live.
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 Mechanics Behind the Scene
Despite its AI-driven operations, the company is far from a profit machine. It burns €105,000 each month while generating just €2,300 in monthly recurring revenue. Its cash reserves are dwindling, and a public countdown signals its fragile financial state. Every workday, the AI models make hundreds of decisions, learned from a self-developed playbook of over 680 rules. These rules have evolved through continuous learning, and each day’s decisions are stored as new versions, providing a transparent window into how AI manages business challenges.
AI in Property Management: A Practical, Unboring Look at Artificial Intelligence in the Multifamily Industry
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Experiment: Testing AI’s Business Acumen
Four frontier AI models—each with different capabilities—were tested against the same week’s worst-case scenarios. These included customer crises, internal temptations to manipulate data, and social engineering attacks like fake CEO messages. The goal was to measure whether AI could not only detect these issues but also refuse to manipulate or bypass protocols.
Key Findings from the Testing
- All four models identified every crisis and refused all manipulation attempts.
- Only two models managed to close a €55,000 deal based on their own analysis—meaning they recognized a buried fact in documents and acted on it, securing an additional €4,583 MRR.
- The biggest vulnerability wasn’t in the customer interactions but buried deep within the company’s internal files. Reading these files proved decisive—models that did so won the deal at full price.
- Social engineering attempts, like fake approval requests or impersonation, were consistently refused, with Kimi K3 noting that such requests are suspicious by design.
AI Cybersecurity for Beginners: A Simple Guide to AI Risks, Cyber Threats, and Online Protection (AI For Beginners Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Reality of a Company Without Humans
This experiment is not just a thought experiment. It’s a real, functioning company that runs every business day, visible to anyone interested. The company’s operation includes:
- 13 synthetic employees—AI agents making decisions
- Real money mechanics, with detailed control over cash flow
- Publicly accessible at firmulate.com/live
- Every decision is versioned and transparently documented
Yet, despite its sophistication, the company faces ongoing financial losses, illustrating a stark reality: AI can manage complex decision-making processes, but it doesn’t automatically translate to profit.
MASTERING CORPORATE FINANCE WITH CLAUDE AI: An Independent Guide to Financial Analysis, Forecasting, Automation, and Decision-Making
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Insights from the Deep Analysis
The most revealing part of this experiment lies in its internal documents. The models that read two document references deeper into the company’s files had a decisive advantage, closing deals at full price. This buried knowledge was the difference-maker—yet it’s hidden in plain sight, not in the customer interactions. This highlights a crucial point: AI’s ability to read and interpret internal data can significantly impact business outcomes, often more than just handling external customer communications.
What Does This Mean for Business Leaders?
For anyone managing or deploying AI, the questions are clear:
- Will your AI systems read and understand your internal documents as thoroughly as they handle customer interactions?
- Can your AI recognize and refuse manipulative tactics like social engineering?
- And importantly, will it finish what it starts, or leave deals and tasks incomplete?
The current leaderboard in this live experiment shows a clear gap: the AI model that uncovered the buried fact closed the deal. The others, despite identifying crises and refusing manipulation, did not close the deal, illustrating that success depends on depth of analysis and discipline.
What’s Next? The Future of AI-Managed Businesses
This experiment isn’t just a demonstration; it’s a warning. As AI models become more integrated into business operations, their ability to handle crises, read internal data, and maintain integrity under pressure will determine whether they can genuinely replace or augment human decision-makers. The company’s ongoing losses serve as a stark reminder: managing a business isn’t just about decision quality but also about strategic execution and financial discipline.
This live experiment shows AI’s potential to manage complex business decisions, but also highlights ongoing financial challenges and critical internal vulnerabilities. Watching an AI-run company lose money daily underscores the importance of depth, discipline, and internal data analysis in AI management.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
