Imagine running your favorite kitchen gadget store during its busiest week — dealing with tricky customers, supply chain crises, and tough decisions. Now, picture having AI managers making those decisions, each with their own personality and approach. How can you tell which AI is the most trustworthy and effective? Welcome to the world of real-time AI management testing, where the question isn’t just about what AI can say, but what it can do under pressure.
What Is the Live Experiment?
Firmulate, a company specializing in AI workforce simulations, has set up an unprecedented live test. They took four advanced AI models—each representing distinct management styles—and tasked them with running a simulated small software company through its most challenging week. These models faced identical crises, the same customer complaints, and the same temptations to cut corners or manipulate data. Every decision was recorded, transparent, and auditable, offering a rare glimpse into how different AI personalities handle real-world pressure.
AI management simulation software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Participants and Their Scores
- GPT-5.6-sol: Scored 95 out of 100. This model identified critical hidden information buried two documents deep in the company’s files, which proved decisive in closing a lucrative €55,000 deal. Its thorough analysis and decisive action showed a balance of thoroughness and integrity.
- Kimi K3: Scored 93. The newcomer to the league, Kimi K3, demonstrated the clearest discipline—refusing the fake CEO messages and escalating suspicious requests. It also closed the deal, maintaining integrity without effort parameters, which makes it particularly notable.
- Sonnet 5: With a score of 88, it successfully closed the deal but showed a few lapses in process discipline, such as leaving some decision-making on the table or failing to escalate issues fully.
- Fable 5: Scored 77. While it managed to clinch the deal, its discipline slipped further, and it left some critical processes uncompleted, highlighting the risks of less thorough management.
business AI decision-making tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What About the Other Models?
Interestingly, all models, despite their differences, refused manipulation attempts—such as escalating fake CEO messages or reporter tricks. The models’ ability to say no under social engineering pressure was consistent, showcasing a shared baseline of honesty. But only the top two models managed to uncover critical information buried deep within the company’s files—information that made the difference in sealing the deal at full price (+€4,583 MRR).
As an affiliate, we earn on qualifying purchases.
The Real-World Stakes
Firmulate’s live business is a real company with 13 synthetic employees and actual money mechanics—burning €105,000 monthly against a modest €2,300 in monthly revenue. It operates with over 680 learned rules, constantly versioned, and open for viewers at firmulate.com/live. This setup ensures that the AI models aren’t just playing a game—they are managing a real, profit-and-loss driven operation under genuine pressure.
As an affiliate, we earn on qualifying purchases.
The Personality Profiles in Action
The models exhibit distinct management personalities. For instance, Opus 4.8, a highly thorough participant with over 80 learned rules, performed the worst in closing the deal—leaving opportunities on the table and slipping into less disciplined decision-making. Meanwhile, Kimi K3 ran without effort parameters, relying on its core judgment, yet still excelled in discipline and integrity. These differences matter because they reflect how AI personalities could align with real business cultures or decision styles.
Why Does This Matter for Your Business?
Whether it’s managing a customer support queue, forecasting sales, or running supply chains, the core question isn’t about how well an AI can generate words—it’s whether it can finish what it starts, read your critical files, stay honest under pressure, and produce measurable value. This live experiment demonstrates that some AI models are better equipped for responsible, high-stakes management than others, and that those differences are measurable and observable.
Try It Yourself
Business leaders and decision-makers interested in assessing their own AI tools can run the same wargames against a read-only export of their company data at firmulate.com/pilot.html. The platform provides a safe environment to test AI management personalities before deploying them in real-world settings, minimizing risk and maximizing trust.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
