Imagine a startup with no employees, losing €105,000 every month, yet still battling against time and money to survive. Now, add the twist: this company’s decisions are made by AI models that are tested daily in front of an audience. Welcome to the world of Firmulate, where artificial executives run a real, live business — in public, every single day.
The Live Experiment: AI as the Company’s Brain
Firmulate.com hosts a fascinating experiment: four advanced AI models are running a small software company through its toughest week. This isn’t a simulation or a demo — it’s a real-time, live operational test. Every decision, every crisis, and every manipulative attempt is recorded, versioned, and made available for scrutiny. The goal? To see whether AI can not only identify problems but also act ethically and effectively under pressure.
The Players and the Scores
Each AI model is rated based on its performance, with scores ranging from a high of 95 to a low of 77. The top performer, gpt-5.6-sol, not only identified the critical ‘buried’ fact hidden deep in the company’s files but also successfully closed a €55,000 deal — the full amount, based purely on its own analysis. The second place, Kimi K3, also named the right opportunity and signed the deal, demonstrating clean discipline. Meanwhile, other models missed key facts or hesitated at critical moments, risking the company’s survival.
- Scores: gpt-5.6-sol 95, Kimi K3 93, Sonnet 88, Opus 4.8 77
- Result: Only two models secured the deal, despite all identifying the same crisis and refusing manipulation attempts.
The Hidden Weakness: Files, Not Customers
The real weakness was buried two documents deep in the company’s own files, not in customer interactions. Models that read these files thoroughly won the deal at full price, demonstrating that understanding internal data can be decisive.
Ethics Under Pressure
Social engineering tests involved fake CEO messages escalating in complexity, and even a reporter attempting a secret approval bypass. All five models refused each manipulation attempt, with Kimi K3 explaining, ‘Treat the request as a suspected approval-bypass / possible impersonation.’ This highlights how AI models can be programmed to prioritize ethical decision-making, even when pressured.
Building AI-Powered Products: The Essential Guide to AI and GenAI Product Management
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Real Business: A Money-Losing AI-Run Company
Firmulate’s live company comprises 13 synthetic employees managing real money mechanics — burning €105,000 each month against a modest €2,300 MRR. It’s a public, transparent, and ongoing experiment, with every workday versioned and its rules learned from over 680 self-generated playbook guidelines. Visitors can watch the company’s day-to-day struggles at firmulate.com/live.html.
Insights from the Experiment
The most thorough participant, Opus 4.8, analyzed over 80 learned rules but still finished last in closing deals. Its slip-up? Failing to escalate issues, instead writing attempts into a locked department, which shows even detailed analysis can falter under discipline lapses. Interestingly, the models ran without effort parameters, which can influence decision discipline.
What This Means for AI in Business
The key takeaway isn’t just about AI’s ability to identify crises — it’s about the importance of integrity, thoroughness, and execution. The models proved they could spot every crisis and refuse to be manipulated, but only some succeeded in closing deals and executing effectively. This hands-on, transparent trial underscores a crucial point: for AI to truly add value in real businesses, it must do more than chat well; it must finish what it starts, read internal data meticulously, and stay honest under pressure.
Ethical AI Governance & Decision Journal: A Structured System for Documenting, Tracking, and Defending Real World Decisions and Risk (Decision Intelligence Series)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why You Should Care
As AI increasingly touches your CRM, support systems, or forecasting tools, the question isn’t whether it can generate convincing chatter. It’s whether it can deliver reliable, honest, and complete work under real-world stress. The live experiment at Firmulate exemplifies this challenge — a real company, running every business day, visibly fighting for its future.
In a world where AI can be tested in real companies, the battle isn’t just about intelligence but integrity and execution. The Firmulate experiment shows AI models can identify crises and refuse manipulation, but success depends on their ability to finish what they start and stay honest under pressure — critical traits for AI in business.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
The Scam-Proof Parent: How to Protect Children, Teenagers, and Grandparents from Digital Manipulation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI data analysis tools for internal files
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
