Imagine watching a company struggle to stay afloat, not in a Hollywood film, but in real life — with no human employees and AI models making every decision. This is the radical experiment by Firmulate, where a simulated startup is living through its worst week, and you can observe every crisis, every choice, and every failure unfold daily. It’s a glimpse into the future of AI-driven management and the importance of trust, discipline, and transparency in automated decision-making.
The Live Experiment: An AI-Managed Company Under Siege
At the heart of this groundbreaking trial is a company with 13 synthetic employees, operating with real money mechanics. Each workday, the company burns through €105,000, while generating just €2,300 in monthly recurring revenue. Visible at firmulate.com/live, the experiment shows a company in continuous crisis, with decisions made by advanced AI models facing the same challenges as any real startup — customer issues, crises, and even attempts at manipulation.
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.
Testing AI Decision-Making in Crisis Conditions
Four frontier AI models, each with distinct characteristics, were tasked with running this virtual company through its worst week. They faced the same customers, the same crises, and the same temptations to cut corners or manipulate outcomes. The models were designed to be fully auditable, with every decision versioned and transparently documented.
The results were revealing: all four models identified every crisis, refused every manipulation attempt, and maintained integrity under pressure. Yet, only two managed to close a €55,000 deal they had identified as achievable — demonstrating an important gap between diagnosis and execution. Interestingly, the decisive advantage was hidden in a document a few layers deep in the company’s files. Models that read and analyze those files successfully closed the deal at full price, adding €4,583 in monthly recurring revenue.
The Challenge of Trust and Discipline
One of the most striking tests was social engineering: fake CEO messages, escalating in stages, and even a reporter’s subtle background request. All five models refused to be manipulated, with one specifically treating suspicious requests as potential impersonation attempts, according to Kimi K3, one of the models. This robustness underscores a critical goal: building AI systems that can resist deception and manipulation.
The Performance and Learning of the Models
Among the models, OPUS 4.8 was the most thorough, learning over 80 rules and performing in-depth analysis. However, it finished last, leaving a close deal unexecuted and slipping into disciplinary lapses, such as writing attempts into a locked department instead of escalating issues. Meanwhile, Kimi K3, running without an effort parameter, and others with high effort settings, showed different levels of discipline and success.
The leaderboard, based on scores from 0 to 100, reveals the stark differences: GPT-5.6 scored 95, Kimi K3 scored 93, Sonnet 5 scored 88, and Fable 5 scored 77. The scores reflect not just crisis detection but also the ability to close deals, avoid manipulations, and maintain discipline.
The Real-World Implication
This experiment isn’t just a tech showcase. It underscores the importance of AI models that do more than generate convincing chat. In real business, AI agents may interact with your customer relationships, support systems, or forecasting tools. The key questions are: will they complete tasks reliably? Will they read critical information before acting? Will they stay honest under pressure? And what does their work truly cost?
The experiment also highlights a crucial, often-hidden weakness: the full potential of AI decision-making hinges on reading and understanding complex internal files, not just surface conversations. Models that access this buried information can unlock opportunities to close deals and add value, even in high-stakes environments.
Watch It Live and Test Your Own Management Skills
For those curious, the entire experiment is live at firmulate.com/live. You can observe the small, cash-strapped company in action, see how AI models handle crises, and even run your own simulations against your business data in a read-only mode at firmulate.com/pilot.
Additionally, a quiz lets you guess which model is making each decision, revealing how transparent and understandable these AI systems really are at firmulate.com/quiz.
The Bigger Picture
This extreme build-in-public experiment offers a rare window into the future of AI in management. It demonstrates that AI can recognize crises, resist manipulation, and close deals — but only if designed with discipline, transparency, and a deep understanding of internal data structures. As businesses increasingly integrate AI into their workflows, the question isn’t just about language generation but about trust, reliability, and ethical decision-making under pressure.
This live experiment shows AI models can manage crises, resist manipulation, and close deals, but only if they read internal data and stay disciplined — crucial for future AI-driven management.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
