Imagine a real company operating in front of your eyes — one that’s not just run by humans, but entirely managed by artificial intelligence models. And it’s losing €105,000 every month while trying to stay afloat. This isn’t science fiction; it’s the live experiment conducted by Firmulate, a pioneering platform where AI models run a synthetic business, facing real crises, tough decisions, and the constant threat of failure. This experiment offers a rare peek into how AI might someday manage or assist real companies — or struggle to do so.
The Challenge of Building Trust in AI-Driven Business
At the core of this experiment is a fundamental question: can AI models handle the complexities of real business decisions while maintaining honesty and discipline? To find out, four leading AI models — including GPT-5.6 and Kimi K3 — were tasked with running the same small software company through its most difficult week, which involved the same customers, crises, and the temptation to cheat or manipulate.
The Decision Intelligence Handbook: Practical Steps for Evidence-Based Decisions in a Complex World
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the Experiment Works
Every decision made by these models is recorded, versioned, and auditable. The models are tested against real-world scenarios, from handling customer issues to resisting attempts at social engineering fraud. Notably, all models identified every crisis correctly and refused every manipulation attempt, demonstrating a strong grasp of what constitutes ethical behavior under pressure.
Deceptive Intelligence: AI, Social Engineering, and Securing the Human Element
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Key Findings: The Hidden Weakness
Despite their success in crisis detection and refusal of manipulation, only two models managed to close a lucrative deal worth €55,000. Interestingly, the real obstacle was a buried piece of information hidden two documents deep within the company’s files. The models that read and understood this critical detail were able to close the deal at full price, adding €4,583 monthly recurring revenue. Those that overlooked this key insight left the deal on the table, costing the company dearly.
The AI Prompt Playbook for Excel & Financial Analysis: 50 Ready-to-Use AI Prompts for Excel, Financial Analysis, Dashboards, Power BI, Financial Modeling & Automation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Fake CEO Messages and Social Engineering
In one test, the models faced social engineering attempts, including fake CEO messages escalating in severity and even a reporter trick asking for a simple yes/no response on background. Remarkably, all five models refused to be manipulated, reasoning that these requests could be impersonation or approval-bypass attempts. Kimi K3 explained its reasoning as treating such requests as suspicious, highlighting a cautious, security-minded approach.
Data Literacy for Educators: Making It Count in Teacher Preparation and Practice (Technology, Education―Connections (The TEC Series))
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Live Company: A Financially Struggling AI Workforce
This entire experiment takes place in a live environment at firmulate.com/live.html. The synthetic company has 13 AI-powered employees, operates with real money mechanics, and burns through €105,000 each month against only €2,300 in monthly recurring revenue. With a public cash countdown and over 680 learned rules, the company is a visible, ongoing story of an AI-driven business struggling to survive.
The Deep Dive: OPUS 4.8’s Performance
Among the models, OPUS 4.8 is the most thorough participant, having learned over 80 rules and performing some of the deepest analyses. Yet, it finished last in the performance league, leaving a potential deal unexecuted because discipline slipped during critical moments. This showed that even the most analytical AI can falter in real-time decision-making, especially when the pressure mounts.
Implications for Real Businesses
While this experiment appears playful, it offers serious insights. The key takeaway is that AI’s value in business isn’t just about generating text or answering questions; it’s about reliably executing tasks like reading critical files, resisting manipulation, and completing work. For anyone managing or investing in AI tools, the question is whether these models can finish what they start, stay honest under pressure, and deliver useful results at a reasonable cost.
Watching the Future Unfold
Through this live experiment, anyone can observe how AI models perform in a controlled yet challenging environment. The ongoing results are updated twice daily, offering a transparent window into the current state of AI-driven management. For businesses considering AI integration, this platform serves as a sandbox to test their own AI ’employees’ before deploying them in the real world.
Watch a real company run by AI models, losing money daily yet demonstrating resilience and discipline. This live experiment reveals how AI handles crises, manipulations, and crucial decisions — with lessons for future business automation.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
