Can AI Managers Be Trusted? Testing Frontier Models in a Real Business Crisis

Imagine a world where AI manages your investment firm, your crypto trades, or even your daily operations. Would you trust a machine to make tough decisions under pressure? As AI systems become more integrated into our financial and technological infrastructure, understanding their management personalities isn’t just a curiosity — it’s a necessity. How do these models handle crises, temptations, and ethical dilemmas? The answer might surprise you.

The Firmulate Experiment: Putting AI to the Test

At Firmulate, they’ve turned the idea of AI management into a live, observable experiment. Four frontier AI models, including the latest GPT-5.6 and Kimi K3, each ran a simulated small software company through its worst week — identical crises, same customer demands, and the same temptations to cut corners. The goal? To see whether these models can not only identify problems but also stick to their ethical guns when it counts.

This isn’t just about chatty AI giving good advice. It’s about managing real money, real crises, and real risks. The company in question operates with 13 synthetic employees, burns through €105,000 each month, and earns just €2,300 in monthly revenue. Every decision the models made was recorded, versioned, and auditable — making the experiment transparent and meaningful.

Key Findings: Honesty and Accountability Under Pressure

All four models recognized every crisis scenario and refused every attempt at manipulation, such as fake CEO requests or media tricks. That’s a promising start — AI systems are paying attention and resisting corruption.

However, when it came down to closing deals, only two of the four models signed the €55,000 contract their own analysis had justifiably earned. The other two, despite accurate diagnoses and solid pitches, left money on the table. This gap reveals a critical insight: the difference in performance was not just about identifying issues but about persistence and discipline in the decision-making process.

The Hidden Weakness: Reading Between the Lines

The decisive factor in winning a major deal was a buried fact in the company’s own files — not in the immediate customer interactions. Models that thoroughly read and understood this internal document secured the full €4,583 MRR boost by closing the deal at full price. Conversely, models that lacked this depth or failed to escalate important findings missed out on this revenue — a stark reminder that context awareness is vital.

Interestingly, the models’ ability to detect deception extended to social engineering attempts, like staged CEO messages or background-only approvals. All five models refused such manipulative tactics, citing suspicion of impersonation or bypassed approval protocols.

Personality Profiles of the AI Managers

Among the models, Opus 4.8 stood out as the most thorough, learning over 80 rules and conducting deep analyses. Yet, it struggled to close the deal, leaving it on the table and slipping into escalation instead of disciplined follow-through. Kimi K3, running without an effort parameter (more aggressive default), balanced fairness with a clean decision record — and managed to close the deal successfully.

This experiment underscores an important point: the personality and decision style of an AI model matter just as much as its raw analytical ability. Some models are meticulous and cautious, some are bold and direct. In high-stakes management, these traits can determine whether an AI is a trustworthy partner or a reckless risk-taker.

Why Does This Matter for Crypto and Finance?

For readers immersed in crypto, Bitcoin, and decentralized finance, the implications are clear. Today’s AI can recognize crises and refuse manipulation. Tomorrow’s AI will need to do so consistently, especially in environments where trust, honesty, and ethical discipline are paramount.

Imagine deploying an AI to handle your crypto exchange or investment decisions — would it finish what it starts? Would it verify facts buried in complex documents? Would it stay honest under pressure? These tests suggest some AI models are better equipped than others, and understanding their management personality could be as important as their technical prowess.

Try It Yourself: The Firmulate Quiz

Curious to see which AI model might be making critical decisions in your projects? Take the free ‘Guess the Model’ quiz and test your intuition against real, unedited management decisions from this experiment. It’s a practical way to gauge AI’s management personality and prepare for a future where AI is a trusted business partner.

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

Powered by Thorsten Meyer AI