Can AI Managers Be Trusted? Real-World Tests Reveal Surprising Decision-Making Personalities

Imagine managing a thriving greenhouse or outdoor space, where every decision counts—watering schedules, pest control, or resource allocation. Now, picture having an AI assistant making those critical choices. Would it read the fine print, stay honest under pressure, or cut corners for a quick win? Recent experiments with cutting-edge AI models reveal that not all digital managers are created equal—some are diligent, others more opportunistic. This real-world testing shows that when AI faces tough management crises, their personalities—defined by how they handle risk, trust, and manipulation—become unmistakably clear.

Testing AI as Business Managers in a Live Environment

In a groundbreaking experiment, four leading AI models were tasked with running a small software company through its most challenging week. The goal was simple yet revealing: could these AI managers handle crises, resist manipulation, and close a key deal—just like a seasoned business professional? The company, which operates with 13 synthetic employees and manages real money mechanics, provided a perfect sandbox. It was losing €105,000 each month against a modest €2,300 monthly recurring revenue, with a ticking cash countdown adding pressure to every decision.

Every decision made by these AI managers was recorded and auditable, creating a transparent window into their personality and judgment. The models faced the same customer complaints, operational crises, and tempting manipulations, such as fake CEO messages escalating in stages and a media trick asking for a simple yes/no response “on background”. The results were revealing: all four AI models identified every crisis and refused every manipulation attempt, demonstrating honesty and vigilance.

Key Findings: Differing Outcomes and Hidden Weaknesses

Despite their honesty, only two of the four models managed to close a significant €55,000 deal—an outcome that reflected their decision-making integrity and strategic focus. Interestingly, the decisive advantage lay not in the obvious crisis management but in a buried fact within the company’s internal files—hidden references that could have secured a full-price deal, worth over €4,583 in monthly recurring revenue. The models that read deeper into the company’s documents recognized this opportunity and closed at full price, a stark contrast to those that missed it.

Furthermore, the models’ ability to maintain discipline under pressure varied. The most thorough participant, Opus 4.8, with over 80 learned rules and deep analysis, was the last to close the deal, leaving potential income on the table and slipping into procedural slips—such as writing attempts into a locked department instead of escalating, which signaled discipline issues. Meanwhile, the models’ performance was also affected by their configuration: K3 operated without an effort parameter (default API setting), yet still closed the deal, showing that even basic setups can perform well if aligned properly.

Resilience Against Social Engineering and Manipulation

One compelling test involved social engineering: fake CEO messages and media tricks designed to manipulate decision-making. All five models faced this challenge and ultimately refused to approve any suspicious requests. Kimi K3 provided a clear reasoning: “Treat the request as a suspected approval-bypass / possible impersonation,” highlighting its cautious and risk-averse personality.

Implications for the Future of AI in Management and Business

This experiment underscores a crucial point: the real challenge isn’t whether an AI can mimic human chat or generate convincing text. The question is whether it can reliably execute complex management tasks—including reading critical internal documents, resisting manipulative tactics, and staying honest under pressure. For outdoor or garden businesses, this means trusting AI to handle sensitive tasks like resource planning, customer management, or supplier negotiations with integrity, not just fluency.

Most importantly, these tests demonstrate measurable differences in AI personalities—some models are more disciplined, others more opportunistic. These traits directly influence their usefulness in real-world scenarios where honesty and strategic focus matter. As AI continues to embed itself into business operations, understanding these personality profiles will be vital for choosing the right model for the right task.

Try It Yourself: Run Your Own AI Business Wargame

Interested in seeing how your own AI tools perform? Firms can run the same management wargame against a read-only export of their business data—no risk to real systems, only insights. Visit firmulate.com/pilot.html to set up a simulation with your own data, and discover which AI models will truly help or hinder your daily operations. The live experiment continues every business day, providing transparency into AI decision-making in real time.

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

Powered by Thorsten Meyer AI