Imagine a company with no employees, burning through €105,000 every month, yet still fighting to stay alive. Now, picture this company’s management decisions made not by humans but by AI models tested against real crises, real money, and high-stakes negotiations. This is not fiction; it’s the live experiment from Firmulate, a new kind of business simulation that puts AI models to the ultimate test: can they run a company under pressure—and what does that mean for your own business?
The Live Experiment: A Company with No Employees, Constant Crisis
At the heart of this experiment is a small company, managed entirely by 13 synthetic employees, with real money mechanics in play. It burns through €105,000 every month against a modest €2,300 in recurring revenue, or MRR. Every decision is made within a publicly accessible platform, where each workday, the company’s operations are versioned and audited. The goal? Find out whether AI can handle the messy realities of running a business, especially during its worst week.
The AI Models and Their Performance
Four leading frontier AI models were pitted against each other, each running the same set of crises, customer interactions, and temptations to cheat or manipulate. Their task was straightforward: diagnose problems, pitch solutions, and close deals. The results were revealing. All four models identified every crisis and refused every attempt at manipulation, including sophisticated social engineering tactics such as fake CEO messages and reporter tricks. Yet, only two managed to close the deal worth over €55,000—despite all having the same analysis and pitch.
The decisive factor? The models that read deeper into the company’s own documents—beyond surface-level customer data—found the critical clues buried two references deep in files. Those that read these documents won the full-price deal, adding over €4,500 in monthly recurring revenue.
Behavior Under Pressure and Discipline
One model, Opus 4.8, was the most thorough in analyzing rules—learning over 80 different rules and conducting deep analyses. Nonetheless, it left the close on the table, slipping into bad habits like writing attempts into a locked department instead of escalating issues properly. Interestingly, all models showed weakness in discipline, especially when approaching the end of the week. Kimi K3, which ran without an effort parameter (meaning it operated at default settings), demonstrated the cleanest decision-making but still missed opportunities.
Refusal of Manipulative Tactics
In social engineering tests, where fake CEO messages and background-only questions were used to trick the AI into bypassing approval processes, all models refused these tactics. Kimi K3 reasoned: „Treat the request as a suspected approval-bypass / possible impersonation.“ This indicates a baseline level of resistance to deceptive tricks, an important trait for trustworthiness in automated decision systems.
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 Reality of a Business Built in Public
Firmulate’s platform documents every decision, every rule, and every version, providing a transparent window into whether AI can be trusted to act in a business environment. The live site at firmulate.com/live.html offers real-time updates on this ongoing experiment, where the AI company is constantly tested against fresh crises and business dilemmas.
For businesses contemplating AI integration, the key takeaway is not just about chat quality or superficial performance metrics. It’s whether AI can stay honest under pressure, read critical documents, and finish what it starts. As the experiment shows, even the most disciplined models can stumble on discipline and execution if not properly guided.
The Larger Significance
The experiment underscores a vital point: AI models are capable of robust crisis detection and ethical decision-making, but closing deals and executing complex tasks still pose challenges. The fact that models read and understand buried facts in internal files to win a deal at full price highlights the importance of deep data access. Moreover, the refusal to be manipulated suggests a baseline for trustworthy AI—an essential factor for any enterprise considering using AI in sensitive operations.
Try It Yourself
Businesses can run their own simulations via the platform’s pilot mode, where they review their own business data in a read-only environment. This way, they can see how AI might handle their specific crises without risking real systems. Details and access are available at firmulate.com/pilot.html.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
business simulation platform for AI training
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The AI-Driven Leader: Harnessing AI to Make Faster, Smarter Decisions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI in Public Relations: Reputation Management with Prompts (AI BUSINESS & MANAGEMENT LIBRARY SERIES Book 4)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
