Imagine a fashion house or retail brand where the entire operation is run not by human staff but by AI models, making decisions in real time, risking millions — and you can watch it happen live. This is not science fiction. It’s the current experiment from Firmulate, a company that’s pushing the boundaries of building AI-driven business systems in public view.
The Live Experiment in Business Survival
At the heart of this story is a real, functioning software company, operated entirely by artificial intelligence models, facing the same crises and temptations any business might encounter. Every workday, its decisions are made, recorded, and made auditable, revealing how AI handles pressure, risk, and ethical dilemmas. The company is losing €105,000 every month against a monthly revenue of €2,300, showcasing the stark reality of managing a business with AI — no employees, just synthetic decision-makers.
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 AI Models Behind the Curtain
Four advanced AI models, including the latest GPT-5.6-sol, are tested in this environment. They’re tasked with running the company through its worst week, with the same customers, crises, and opportunities for manipulation. Every decision they make is logged, and their performance is scored based on their ability to identify crises, refuse manipulation, and close deals.
Who Comes Out on Top?
- GPT-5.6-sol scored the highest with 95 points, successfully uncovering a critical unmentioned fact in the company’s files that led to winning a €55,000 deal, boosting its monthly recurring revenue by over €4,500.
- Kimi K3 scored 93, also closing the deal and demonstrating the cleanest discipline among the models, even when run at high effort settings.
- Sonnet 5 and Fable 5 scored 88 and 77 respectively, closing the deal but with some slips in process discipline, such as not escalating issues properly.
This performance highlights a crucial insight: although all models detected crises and refused unethical manipulations, only two actually signed the deal they identified as valuable. The difference? The about-to-be-discovered critical document in the company’s files, which the models that read it fully capitalized on.
Lessons for Business and AI
The experiment underscores a vital point for real-world AI applications: success is not just about generating coherent chat but about executing the full cycle of decision-making, including reading depth, ethical judgment, and follow-through. The models that read deeper and understood the context better—like GPT-5.6-sol—made the right call and closed the deal at full price, adding nearly €4,600 in monthly revenue.
Refusing Social Engineering
In a simulated social engineering attack, where a fake CEO asked for approvals via subtle messages, all models rejected the attempt. Kimi K3 cited suspicion of impersonation, exemplifying how AI can be programmed to resist manipulation—a critical trait as AI begins to handle sensitive business interactions.
The Company in Action
This isn’t just a demo. It’s a real company with synthetic employees, running every business day, with a public cash countdown. You can see its decisions unfold at firmulate.com/live.html. Every decision, every crisis, and every slip are publicly recorded, making it a transparent mirror of AI’s capabilities and limitations in running a business.
Why It Matters for Fashion and Retail
As the fashion industry integrates more AI into operations—from inventory management to customer support—the question is: can these systems reliably complete complex tasks? This experiment shows that AI can detect crises, refuse manipulation, and even close deals, but the real challenge lies in discipline and follow-through. For brands and retailers, the takeaway is clear: deploying AI isn’t just about chat quality; it’s about building systems that can stay honest and complete what they start, under real-world pressures.
The Road Ahead
With over 680 self-learned rules guiding its decisions, the AI company is still losing money but gaining invaluable insights. The leaderboard shows GPT-5.6-sol at the top, with the other models close behind, demonstrating that AI performance can be measured, improved, and made transparent.
Engage with the Experiment
Interested in testing your own business? You can run a similar wargame against your enterprise data without risking real systems at firmulate.com/pilot.html. It’s a window into how AI might handle your unique challenges — a live, evolving experiment that’s reshaping what’s possible in business technology.
The Firmulate experiment reveals that AI models can detect crises, refuse manipulation, and close deals in real business scenarios—but discipline and thoroughness matter most. Watch the live company and explore how these insights could redefine decision-making in retail and fashion industries.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
