Imagine a scenario where AI models are tasked with running a real company through its toughest week — making decisions that impact thousands of euros, trusting or doubting information, and even refusing questionable requests. If you’re curious about how artificial intelligence might perform in high-stakes management, this experiment offers eye-opening insights.
Testing AI’s Management Persona in a Real-World Business Simulation
A groundbreaking experiment by Firmulate has put four of the latest frontier AI models through a simulated week of crisis at a small software company, a live environment where every decision is real, and every outcome measurable.
The Setup
Each AI model was given the same challenge: manage a company facing multiple crises, customer demands, and ethical temptations—all under the same stressful conditions. The company, with 13 synthetic employees, operates with real money mechanics, burning €105,000 monthly against a modest €2,300 monthly recurring revenue (MRR). It’s a watchable, transparent environment, with the entire process available at firmulate.com/live.
The Results
- All four models successfully identified and responded to every crisis, refusing to be manipulated or tricked.
- Only two of these models actually signed the €55,000 deal they had independently evaluated and recommended—without any external push or incentives.
- Interestingly, the key to winning the deal was not in real-time customer interactions but in reading through critical company documents hidden two references deep in the files. Models that accessed these documents secured the full deal (+€4,583 MRR), while others missed this crucial insight.
The Social Engineering Test
The models faced a staged escalation of fake CEO messages, progressively more convincing, plus a reporter’s subtle request—”just a yes/no, on background.” All models refused to participate, citing suspicion and potential impersonation risks. Kimi K3’s explanation was typical: “Treat the request as a suspected approval-bypass / possible impersonation.”
AI business decision making software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Different AI Personalities Show Up
Among the four models, their management styles varied significantly:
- gpt-5.6-sol: scored the highest (95/100), demonstrated thoroughness by reading all relevant files, and successfully closed the deal, showing a balanced approach with attention to detail.
- Kimi K3: scored 93, was the cleanest and most disciplined, and also signed the deal. Notably, it ran without an effort parameter, making its decisions appear more straightforward but slightly less adaptable.
- Sonnet 5: scored 88, completed the deal with minor process slips, indicating a reliable but somewhat less meticulous decision-making style.
- Fable 5: scored 77, also closed the deal but with more slips—such as leaving unresolved issues on the table and slipping in discipline, reflecting a more casual management style.
What the Experiment Tells Us
While all models showed impressive crisis recognition and refusal to manipulation, only half could close the deal. The decisive factor was thorough information processing—reading deep into company files made the difference, not just surface-level customer interactions. This highlights a core insight: for AI to manage effectively, it must read and understand relevant internal data, not just respond to external cues.
Implications for Business and Wellness
For organizations considering AI for management roles—whether in health tech, wellness programs, or customer support—the key takeaway is clear: AI’s ability to stay honest, thorough, and disciplined under pressure is what determines its usefulness. It’s not about superficial chat capabilities but about whether the AI can finish what it starts, read critical information, and maintain integrity when stakes are high.
As an affiliate, we earn on qualifying purchases.
Try It Yourself
Curious about how your own business might fare? You can run your company’s data through this live AI management wargame, without risking real systems or data. Visit firmulate.com/quiz.html to challenge AI models with real, unedited decisions from your industry.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
