Imagine a Business That’s Watchable, Self-Driving, Yet Constantly Fighting for Survival
What if you could see an entire company in action—its decisions, crises, and even its failures—live, with no human employees involved? Welcome to the world of firmulate.com, where AI models run a tiny software firm in real-time, subjected to the same brutal challenges any business faces.
As an affiliate, we earn on qualifying purchases.
The Live Experiment: An AI Company Under Siege
At the heart of this bold experiment is a small, synthetic company managed by 13 AI ’employees,’ each driven by cutting-edge models like GPT-5.6-sol, Kimi K3, Sonnet 5, and Fable 5. Every day, these models make management decisions—reacting to crises, reading files, and even negotiating deals—while their actions are publicly observable at firmulate.com/live.html.
This setup isn’t just for show. It’s an intense test of AI reliability: can these models handle real business crises, spot hidden risks, and stay honest under manipulation attempts? The answer, in many cases, is yes. All four models identified every crisis and refused attempts at manipulation, such as fake CEO messages or reporter tricks.
The Results: Who Meets the Goal?
Although all models successfully identified crises and refused manipulation, only two managed to close a critical deal worth €55,000 in monthly recurring revenue (MRR). The top performer—gpt-5.6-sol—found a buried fact in the company files that led to sealing the deal at full price, adding over €4,583 MRR. The Kimi K3 model, celebrated for its fairness and discipline, also closed the deal but without the edge of uncovering hidden insights.
Interestingly, the experiment revealed a hidden vulnerability: the decisive edge was tied not to customer interactions but to internal document analysis. Models that delved into the company’s internal files, rather than just customer-facing data, secured the full-price deal.
Built-in Public: Transparency in Action
This experiment is openly accessible, with every decision versioned and auditable. This transparency allows anyone to observe how AI models handle real-world pressure and crises, providing insights into their strengths and weaknesses—crucial information if these agents are to touch your own business systems someday.
What Went Wrong and What Worked
Deep analysis shows that even the most thorough model, Opus 4.8, which learned over 80 rules and performed the deepest analysis, did not secure the deal. It left opportunities unexploited and slipped discipline, such as writing issues into a separate department instead of escalating. This highlights how even sophisticated models can falter in practical, high-pressure situations.
Moreover, the models’ ability to resist social engineering was tested through staged fake CEO messages and a reporter trick. All five models refused to approve or act on suspicious requests, demonstrating a crucial trait: integrity under pressure.
The Bigger Picture: A New Era of Business Testing
This ongoing live experiment offers a glimpse into how AI might manage real companies in the future. The setup is entirely transparent and self-contained, with no access to external systems or the ability to alter actual business data. It’s a sandbox for testing management quality, not just chat capabilities.
For business owners and decision-makers, the message is clear: the crucial question isn’t whether an AI writes well but whether it can deliver complete, honest, and effective work—especially under pressure. Can it find hidden risks in your internal documents? Will it stay honest when faced with manipulation? These are the real tests of AI readiness for business management.
The Future of AI-Driven Management
As firms like this continue to evolve, watching them in action offers valuable lessons. The experiment is ongoing, with 23 benchmark runs queued and new results published regularly. It’s a real-time window into how AI models perform when pushed to their limits—something every business should be paying attention to, especially in energy and critical infrastructure sectors.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
