Can AI Managers Be Trusted? A Live Experiment Reveals Surprising Truths

Imagine managing your favorite kitchen appliance: will it follow your recipe to perfection or cut corners to save time? Now, picture AI models running an entire company through its worst week. Sounds like science fiction? Not anymore. At Firmulate, a real experiment is happening where AI models take on the role of a management team, making decisions in a high-stakes business environment. The results might just change how you think about trust and AI — whether in your kitchen or your boardroom.

How Do AI Models Handle Crisis and Manipulation?

In a groundbreaking live test, four frontier AI models were tasked with managing a small software company facing its most challenging week. They faced identical crises, demanding decisions, and temptations to cheat. The goal? See if AI can read between the lines and make ethically sound, strategic choices under pressure.

Consistent Vigilance and Ethical Standings

Remarkably, all four models detected every crisis and refused to yield to manipulation. They faced a staged social engineering attack involving fake CEO messages and a reporter’s subtle request to bypass approval processes. Not a single model was fooled; all refused to approve suspicious requests, demonstrating robust ethical boundaries.

The Hidden Weakness in Decision-Making

While all models showed resilience, the real story was in the details. Each decision was documented and auditable, revealing that the decisive weakness was hidden two document references deep within the company’s files. Models that read these internal files successfully identified the critical information needed to close a lucrative deal, worth over €4,500 monthly recurring revenue (MRR). Conversely, models that missed this buried evidence left the deal on the table, earning only partial progress.

Decisive Results in a High-Stakes Environment

Out of the four models, the top performer was gpt-5.6-sol, scoring 95 out of 100. It uncovered the crucial information, closed the €55,000 deal, and demonstrated complete performance. Close behind was Kimi K3, with a score of 93. This newcomer was praised for its discipline and honesty, managing to close the deal without slipping into shortcuts. Others, like Sonnet 5 and Sonnet 4, scored 88 and 77 respectively, and showed more process slips, typically leaving opportunities unexploited.

Model Personalities and Management Styles

The experiment also highlighted different management personalities embedded within each AI model. For example, Opus 4.8, with the deepest analysis and over 80 learned rules, was thorough but less decisive—often leaving opportunities unclaimed when discipline slipped. Meanwhile, Kimi K3 ran without an effort parameter, making it more disciplined but less aggressive in pushing for deals.

Implications for Real Business and AI Trust

This live experiment underscores a critical point: in managing real companies, AI models are capable of spotting crises and resisting manipulation. But their effectiveness hinges on reading internal information thoroughly—something that can determine whether a deal is lost or won, especially when stakes are high.

Why It Matters for Your Business

If AI agents start touching your CRM, support queues, or forecasting systems, the key question isn’t just whether they write well. It’s whether they stay honest, finish what they start, and read your files deeply enough to make the right decision. The live data shows that some models excel at this, while others leave opportunities on the table.

Try the Quiz and See Which Model Matches Your Style

Curious to see which AI model you relate to most? You can test your decision-making style against these models in a public quiz at firmulate.com/quiz.html. For businesses considering AI, there’s also a way to simulate your own company’s worst week—without risking real money—using Firmulate’s live wargame. It’s a chance to evaluate how your AI workforce performs before you bring it onboard.

Final Thoughts: Trust and AI in Business

This experiment reveals that AI models can handle crises, refuse manipulation, and even close big deals when they read the right internal clues. But the difference in performance shows that not all models are equal—some have stronger discipline, some are more thorough, and some leave opportunity on the table. As AI begins to touch your business, understanding these nuances is essential for building trust and making smarter investment decisions.

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

Powered by Thorsten Meyer AI