Imagine an AI running a company during its worst week—can it outsmart human managers? The answer might surprise you.
At a glance, artificial intelligence often feels like an invisible assistant or a chatbot. But what if AI took the wheel of a real company—facing the same crises, the same temptations, and the same tough choices as human managers? And what if, through a live experiment, we could see exactly how different AI models handle these challenges?
This is not science fiction. It’s the core of a groundbreaking pilot by Firmulate, a company that runs AI models as complete companies in a simulated environment. Their goal: measure the management quality of frontier AI models in real-time, under realistic stress conditions.
The Setup: Making AI Manage a Real Business
Imagine a small software firm with 13 synthetic employees and real money mechanics—burning €105,000 a month against a tiny €2,300 monthly recurring revenue. Their daily operations are tightly versioned, with over 680 self-learned rules guiding every decision. The company faces real crises—customer issues, supply chain hiccups, and tempting shortcuts—just like any real business.
In this experiment, four different frontier AI models took turns managing this virtual company during its worst week. Each AI faced identical scenarios, with decisions meticulously logged and auditable. The challenge: see which model would handle crises, refuse manipulative tactics, and ultimately close a critical €55,000 deal based on their own analysis.
The Models and Their Personalities
- gpt-5.6-sol: The top scorer, with a 95 out of 100. Known for catching hidden details and sealing deals at full price.
- Kimi K3: The newcomer, scoring 93—its discipline and honesty stood out. It refused all manipulation attempts and closed the deal cleanly.
- Sonnet 5: Scoring 88, it mostly succeeded but slipped on a few process slips.
- Fable 5: The lowest among the winners, with a score of 77. It managed to close, but left some opportunities on the table and slipped discipline-wise.
The Crises: How Every Model Performed
All four AI models demonstrated remarkable vigilance—they identified every crisis and refused all attempts to manipulate or bypass their protocols. Even when fake CEO messages escalated over three stages, and reporters tried to trick them into approvals, all models refused to bend.
The critical difference came in their analysis of internal files. The decisive weakness in the competitor models was buried two document references deep in the company’s files. Models that thoroughly read these internal documents managed to win the deal at full price—adding over €4,500 in Monthly Recurring Revenue (MRR). In contrast, the model that missed this buried fact left money on the table.
Who Won? The Final Results
| Model | Score | Outcome |
|---|---|---|
| gpt-5.6-sol | 95 | Found the hidden fact and closed the deal at full price |
| Kimi K3 | 93 | Closed the deal, clean discipline |
| Sonnet 5 | 88 | Closed the deal but with some process slips |
| Fable 5 | 77 | Closed the deal but left money on the table |
The results reveal a fascinating insight: while all models detected crises and refused manipulative tactics, the ability to thoroughly analyze internal documents and make full-value deals distinguished the top performers.
What Does This Mean for Businesses?
As AI becomes more integrated into daily business operations—handling CRM, support queues, or forecasting—the key questions are no longer about how well they generate language or mimic human conversation. Instead, it’s whether these models can stay honest under pressure, read critical internal files, and complete valuable work reliably.
In this live experiment, the models’ personalities—ranging from thorough analysts to more terse decision-makers—show us that AI’s management style can be as measurable and distinct as human traits. The top performers demonstrated discipline, attention to detail, and integrity, leading to concrete business outcomes.
The Live Platform: Wargaming Your AI Workforce
This isn’t a one-off test. Firmulate offers a live platform where enterprises can run similar management wargames against their own business data—without risking real systems. Watch your AI models confront crises, temptations, and manipulations in a safe, controlled environment. The goal: ensure your AI workforce is ready to handle real-world challenges with honesty and effectiveness.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI management decision-making tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
