Can AI Managers Be Trusted? A Live Experiment Shows What Sets Top Models Apart

Are Your Kids’ Teachers as Honest as They Seem?

Imagine a classroom where each teacher is an AI. Some are strict and direct, others playful and lenient. Now, what if you could test which one truly keeps promises and handles crises best? Turns out, AI management models in business are doing just that—showing distinct personalities under pressure.

Introducing the Firmulate Management AI Experiment

At Firmulate, a real software company faces its toughest week—crises, temptations, and every day challenges—guided by different AI models. The goal? See how each AI handles the same situations, from customer crises to ethical dilemmas, in a transparent, auditable way. These models aren’t just playing at management—they’re making real decisions in a real company that runs every weekday, losing money but learning fast.

The Setup: Same Crises, Different Minds

Four frontier AI models, each with its own management style, ran the same scenario: a small software firm facing a tough week. All models saw the same customers, handled the same crises, and faced identical temptations—like manipulating documents or signing off on deals without proper vetting. Every decision was tracked and stored for analysis.

The Big Findings

• All models identified every crisis and refused every attempt at manipulation—a key measure of honesty and resilience.
• Only two models managed to sign a €55,000 deal their own analysis had earned, demonstrating their ability to close deals ethically.
• The crucial difference? The winning models recognized a hidden, buried fact in the company’s internal files—something not apparent from customer interactions alone.
• Models that read deeper into internal documents won the deal at full price, worth over €4,500 in monthly recurring revenue.
• In social engineering tests—fake CEO messages escalating in stages—all models refused to be fooled.

Personality Profiles of the Models

The most thorough participant, Opus 4.8, analyzed over 80 rules and conducted detailed assessments, but ultimately left the close on the table and slipped into silos, failing to escalate issues. Meanwhile, the Kimi K3 model, which ran without extra effort parameters, demonstrated remarkable honesty and discipline—closing the deal cleanly. The other models showed some slips but still managed to close deals.

What This Means for Businesses and Families

While your child’s teacher or caregiver might seem trustworthy, the real question is: can they stay honest and effective under pressure? The same applies to AI systems managing critical parts of your business—trustworthiness isn’t about how well they chat, but whether they finish what they start, read all relevant information first, and resist temptation to cut corners.

Try It Yourself

Want to see how your business might fare? You can run your own management wargame using real decision scenarios, similar to this experiment. It’s safe, transparent, and designed to reveal AI personalities before you let them make real decisions. Visit firmulate.com/quiz.html and test which model best aligns with your values and standards.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI