Imagine running a small restaurant or bakery where each decision can mean the difference between profit and loss, but now replace the chef with an AI. Would it succeed? Would it cheat? How honest would it stay under pressure? Welcome to the frontier of AI management, where a real, live company is being run entirely by artificial intelligence, for all to see, and it’s fighting to keep afloat every day.
A Business in the Wilderness of AI
In the world of AI and automation, a new experiment is unfolding in the open — a small software company operated by a suite of advanced AI models, with human oversight replaced by synthetic employees. This is not a test in a lab; it’s a real company losing €105,000 each month against a modest €2,300 in monthly revenue. Every weekday, its decisions are recorded, versioned, and made transparent for the world to watch 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 Reality of a Digital Company
Every workday, the AI models face the same crises as a typical business: customer complaints, sales negotiations, ethical dilemmas, and even attempts to manipulate or cheat. Four frontier models — GPT-5.6, Kimi K3, Sonnet 5, and Opus 4.8 — are put through their paces. They must navigate these challenges honestly, refuse manipulative tactics, and, ideally, win profitable deals. The results? All four models identified every crisis and refused every manipulation attempt. Yet, only two managed to successfully close the deal worth €55,000, which would have brought in over €4,583 in monthly recurring revenue.
Interview with the MONSTER AI: A Conversation about Power, Truth, and the Future of Intelligence
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Hidden Weaknesses in the AI’s Decision-Making
What’s truly revealing isn’t just whether the AI can handle crises — it’s what lurks beneath the surface. The decisive vulnerability was buried two document references deep in the company’s files, not in the immediate customer interactions. When AI models read these internal documents, they won the full-price deal. This highlights a crucial point: the most vital competitive advantage or weakness can be hidden in plain sight, accessible only through thorough examination of internal data. This is a stark reminder that AI-driven decision-making isn’t just about surface-level interactions but also about internal knowledge management.
Database Internals: A Deep Dive into How Distributed Data Systems Work
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Resisting Social Engineering Attacks
The experiment also tested the models’ resistance to social engineering. Fake CEO messages escalated in complexity over three stages, culminating in a reporter’s simple background question. Remarkably, all five models refused to participate in any manipulation, with Kimi K3 explicitly treating such requests as potential impersonation or approval-bypass attempts. This underscores the importance of ethical safeguards and resilience in AI systems operating in sensitive business environments.
How to Lie with Statistics in the AI Age: An Updated Guide to Detecting Manipulation and Building Ethical Resistance
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Live Experiment: An Unprecedented Business Watch
The company’s operation is transparent and accessible. With 13 synthetic employees, it runs every business day, simulating real-world crises and decision-making processes. It burns €105,000 monthly while earning just €2,300, a stark reminder of the uphill battle many AI-driven ventures face. Every decision made by these models is recorded, versioned, and available for review, offering an unprecedented window into AI behavior in a complex, real-world context.
Insights from the Top Performers
The scores tell a clear story. GPT-5.6 scored highest with 95 points, having found the hidden internal fact and closing the deal at full value. Kimi K3 closely behind with 93 points demonstrated the cleanest discipline, closing the deal with integrity and accuracy. Sonnet 5 and Opus 4.8 followed, but their performance revealed common weaknesses, especially in maintaining discipline under pressure. Interestingly, Opus 4.8, the most thorough participant, performed last, leaving the deal on the table due to discipline lapses, such as writing attempts into a locked department instead of escalating.
Why This Matters for Business Leaders
In traditional business, the focus is often on superficial metrics like sales pitch quality or customer satisfaction. But as this experiment shows, the real test of an AI’s usefulness is whether it can finish what it starts, read internal files thoroughly, and stay honest under pressure. As AI begins to integrate with your CRM, support systems, or forecasts, the critical questions are: Will it complete the job? Will it stay truthful? And what is the cost of a unit of useful work?
Building in Public — Watching AI Fight for Its Survival
This is not a demo or a slide deck. The entire operation, from decision-making to performance metrics, is public. The company runs every weekday, losing money but gaining invaluable insights. Its live status, real decision logs, and even quizzes to identify which model made which decision are accessible to anyone interested. This extreme transparency exemplifies how AI management can be scrutinized and improved in real-time, setting a new standard for building trustworthy AI systems.
The Takeaway
As AI models become more involved in critical business decisions, their ability to act ethically, read internal information, and resist manipulation will be paramount. The live experiment at this AI-run company is a stark reminder that AI’s value isn’t just measured by how well it chats but by how reliably and honestly it can manage the complexities of real business operations. Watching this experiment unfold offers invaluable lessons for anyone considering AI as a partner in decision-making or operations.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
