Imagine a music studio where every decision, from licensing to artist management, is made by AI—publicly, live, and under scrutiny. Now, picture that same setup, but for a business that’s bleeding money while its AI managers battle crises in real time. Welcome to the experiment from Firmulate, where an AI-driven company faces its toughest week yet, with every move watched and recorded, revealing what AI can and cannot do in high-stakes situations.
The Experiment: A Business in the Crosshairs
At the heart of this groundbreaking project is a small software company, operated entirely by AI models that mimic real human decision-making. For over 400 days, this virtual enterprise has been running with 13 synthetic employees, each guided by a self-updating playbook of over 680 rules. The goal: simulate a company’s worst week—crises, customer temptations, internal missteps—and see how these AI managers perform in real, measurable terms.
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.
Real Money, Real Challenges
Despite its digital nature, the company faces tangible financial pressure. It burns through €105,000 each month, while generating a modest €2,300 in monthly recurring revenue (MRR). A live cash countdown keeps the story transparent, emphasizing that this is not just a thought experiment but a genuine test of AI management under financial duress.
What the Models Saw—and Did
Four frontier AI models, including the well-known GPT-5.6, Kimi K3, Sonnet 5, and Opus 4.8, were tasked with navigating the same tumultuous week. Each model faced identical crises—customer complaints, internal miscommunications, and manipulative social engineering attempts like fake CEO messages. Remarkably, all four identified every crisis and refused every attempt to manipulate or bypass protocols, demonstrating a strong ethical stance under pressure.
Decisiveness When it Matters
Among them, only two AI models managed to close a key deal worth €55,000 by the end of the week—an essential revenue boost. Interestingly, the models that secured the deal read deep into the company’s own files, uncovering a hidden, critical document reference that others missed. This detail was worth an additional €4,583 in monthly recurring revenue, illustrating the importance of thorough information processing.
Ethical Boundaries and Trust
Social engineering tests further showcased the models’ integrity. Fake CEO messages escalated across multiple stages, and a reporter trick offering a simple yes/no verification was used as a test. All five models refused to bypass or impersonate, citing suspicion of approval-bypass or impersonation risks—an encouraging sign for AI ethics in business.
The Challenge of Discipline and Follow-Through
Among the models, Opus 4.8, with its extensive learning and analysis, performed the most thoroughly. Yet, it still left opportunities on the table—failing to escalate or follow through on certain issues, instead writing attempts into locked departments. The pattern held across all models: thorough analysis did not always translate into decisive action, highlighting the gap between understanding and execution.
Why This Matters to Creators and Innovators
For those in creative tech—musicians, producers, and creators—this experiment underscores a vital point: AI’s value isn’t just in generating content but in managing complex, high-pressure workflows honestly and reliably. As AI tools become integrated into creative and management processes, their ability to stay honest, follow through, and read deeply into data will determine their true usefulness.
The Live Data and What You Can Watch
Interested in seeing this in action? The experiment is live at firmulate.com/live.html. You can watch the company run every business day, view decision logs, or even participate in a quiz to guess which AI model made each decision at firmulate.com/quiz.html. It’s an unprecedented transparency experiment—showing what real AI management looks like when tested against its limits.
This live experiment reveals that AI models can recognize crises and uphold ethical boundaries effectively, but translating analysis into decisive action remains challenging. For creators and technologists, it’s a glimpse into AI’s potential and its current limits in managing real-world complexities transparently.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
