Imagine running your cleaning business with a team that never takes breaks, never calls in sick, but also never makes a profit. Now, what if this team was powered by artificial intelligence, working tirelessly through a simulated week of crises and tough decisions? Welcome to the world of Firmulate, where a live AI-driven company battles to stay afloat — and you can watch every move.
The Live Experiment: A Company in Crisis
At the heart of this experiment is a small software company operated entirely by artificial intelligence models. These models, representing 13 synthetic employees, face the same daily challenges any real business might encounter — customer issues, crises, and the temptation to cut corners. But unlike typical AI demos, this company is real in its mechanics: it burns €105,000 in cash each month against a modest €2,300 in monthly recurring revenue (MRR). Every workday, its decisions are versioned and made public for anyone to observe at firmulate.com/live.html.
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.
Goals and Findings
The core of the experiment was to see whether AI models could reliably handle management decisions under pressure. Four leading frontier models — including GPT-5.6-SOL and Kimi K3 — faced the same week of crises and temptations. The results are revealing:
- All models identified every crisis and refused manipulative attempts, demonstrating robust integrity.
- Only two models managed to close a €55,000 deal their own analysis had earned — a clear indicator of operational discipline.
- Interestingly, the decisive advantage was found not in the immediate crisis responses but in reading deeper into company files. The models that uncovered a critical, buried document reference secured the full deal, adding over €4,500 in monthly recurring revenue.
Deception and Integrity Under Scrutiny
Beyond crises, the models faced social engineering attempts — fake CEO messages escalating over multiple stages and even a reporter trick asking for a simple yes/no response. Remarkably, all models refused such manipulative requests, with Kimi K3 explicitly reasoning that the request resembled an impersonation or approval-bypass attempt.
The Reality of the Live Business
Despite the models’ resilience, the company’s financials are dire. It is currently losing €105,000 each month against a tiny €2,300 monthly income. The firm’s operational approach involves a detailed set of over 680 self-learned rules, with every decision, update, and mistake publicly recorded and versioned daily at firmulate.com/live.html. This transparency offers an unprecedented view into how AI manages a real business under stress, highlighting both strengths and vulnerabilities.
Lessons from the Frontline
One of the most telling insights comes from the model Opus 4.8, which, despite thorough analysis and having learned over 80 rules, left a critical deal on the table. Discipline slipped, and decision attempts were misdirected into locked departments instead of escalation. The experiment underscores an essential truth: AI can identify problems and refuse to manipulate, but operational discipline and strategic judgment remain challenging — especially in high-pressure scenarios.
Implications for the Future
This experiment isn’t about chatbots making small talk; it’s about AI agents managing complex workflows and making trustworthy decisions — even when money, reputation, and survival are at stake. As AI begins to touch critical business functions like customer relationship management, support, or forecasting, the question is not whether it can write well but whether it can finish what it starts and stay honest under pressure.
What You Can Do
Business leaders and decision-makers can run their own version of this test through Firmulate’s pilot program, which allows simulation against their own data without risking real systems. This approach offers a glimpse into how AI might perform as a trustworthy partner or threat in everyday operations.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
