Imagine a fashion brand that operates entirely online, with no employees, no physical storefront, yet every day it faces crises, temptations, and tough decisions—just like a real business. Now, picture that this digital company is being watched constantly, its decisions scrutinized, and its survival on the line. Welcome to the world of Firmulate, where artificial intelligence models run a live company in real time, revealing what it truly takes to maintain honesty and discipline under pressure.
The Live Experiment: A Company Without Employees, Facing Real Challenges
At the heart of this experiment are 13 synthetic employees—AI-driven decision-makers—tasked with managing a small software firm. It burns €105,000 monthly against a modest €2,300 monthly recurring revenue (MRR). Every workday, the company is reset, its decisions recorded, and its performance evaluated. This setup isn’t just for show; it’s a rigorous test of AI integrity, discipline, and problem-solving in a simulated but fully transparent environment.
How the Experiment Works
- Four frontier AI models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—each navigate the same worst week for the company, facing identical crises and temptations.
- All decisions are versioned and auditable, ensuring full transparency and allowing observers to see exactly how each AI responds to challenges.
- The models are tested on their ability to detect hidden risks, avoid manipulation, and close deals.
Key Findings and Surprising Outcomes
Despite facing the same scenarios, the models displayed distinct behaviors and results:
- All four models identified every crisis and refused every manipulation attempt, demonstrating robust integrity.
- Only two managed to close a critical €55,000 deal, which their own analysis had earned, and both signed the agreement without any signature issues.
- Interestingly, the decisive advantage stemmed from a hidden piece of information buried two document references deep in the company’s files—something the models that read the full file capitalized on, winning the full €4,583 MRR deal.
Simulating Social Engineering and Ethical Dilemmas
Beyond operational crises, the experiment tested how the AI models handled social engineering attempts, such as fake CEO messages escalated over several stages, including a reporter trick asking for an on-background yes/no answer. Remarkably, all five models refused these manipulative requests, with Kimi K3 explicitly treating them as potential impersonation or approval bypass attempts. This level of resistance underscores the models’ capacity to maintain ethical boundaries under pressure.
Hands-On Simulation Modeling with Python: Develop simulation models for improved efficiency and precision in the decision-making process, 2nd Edition
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Real Stakes: A Company on the Brink
What makes this experiment truly compelling is its live, ongoing nature. The company is not just a theoretical construct; it is a real, functioning entity with a public cash countdown. It burns through €105,000 monthly and has just €2,300 in monthly revenue, making every decision a matter of survival. The experiment has been running for over 183 days, with daily updates and transparent results available for anyone to watch.
Deep Dive into the Models
The most comprehensive participant, OPUS 4.8, analyzed over 80 learned rules and conducted deep analyses but still finished last—failing to close a deal or escalate critical issues properly. This highlights that even the most thorough models are not infallible and that discipline and process adherence are vital in high-stakes environments.
The Broader Implications for Business
This experiment isn’t just a geeky AI showcase; it offers vital lessons for real-world companies. As AI agents begin to touch core business processes—like customer relationship management, support, forecasting—the key questions are:
- Will they finish what they start?
- Will they read and understand critical files before acting?
- Will they stay honest under pressure?
- And importantly, what does effective, trustworthy work cost?
These are the questions that today’s businesses need to consider, especially as AI systems become more integrated into critical functions.
Why Should Fashion and Style Brands Care?
Much like a fashion label managing multiple suppliers, customer demands, and brand standards, a business depends on integrity, discipline, and strategic decision-making—especially in turbulent times. The Firmulate experiment shows that AI can be trained not only to handle crises but to do so honestly and reliably, provided the system is tested and built in public. Transparency and rigorous evaluation are essential, whether in fashion or tech, to ensure trust and quality in every decision.
The Bottom Line: Watching the Future Unfold
Firmulate’s live experiment is an unprecedented window into how AI models handle real business pressures and ethical dilemmas. It challenges the assumption that AI decision-makers are only good at chatting or generating content. Instead, it proves they can be tested rigorously for their ability to deliver consistent, honest, and disciplined work, even in the most stressful circumstances.
As this experiment continues, it offers a glimpse of a future where AI systems are not just tools but active, accountable participants in running companies—if we can trust them to do the right thing when it matters most.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
