Can AI Managers Play It Straight? The Live Experiment Reveals the Truth

In the world of entertainment, we love watching characters with distinct personalities — from the terse detective to the charming diplomat. But what if AI could have personalities too? Turns out, when it comes to managing a company under pressure, AI models show surprisingly human traits. And a groundbreaking experiment puts these digital personalities to the test in a real, live business simulation.

The Live AI Showdown: Running a Company in Real Time

Imagine a real software company fighting to stay afloat, with daily crises, customer demands, and temptations to cheat. Now, replace the human managers with advanced AI models. This is exactly what the folks at Firmulate did in their latest experiment. They ran four frontier AI models through the same challenging week, with identical problems and the same set of rules, to see how each would perform as a virtual management team.

The Models and Their Scores

• gpt-5.6-sol scored 95 out of 100, winning by uncovering hidden information and sealing a crucial deal.
• Kimi K3 scored 93, closing the same deal with the cleanest discipline among all models.
• Sonnet 5 scored 88, also closing the deal but showing a few more slips in process adherence.
• Fable 5 scored 77, completing the deal but with noticeable discipline gaps.

Remarkably, all models spotted every crisis and refused every manipulation attempt, demonstrating integrity. Yet, only two actually signed off the lucrative deal — a testament to how different management personalities can influence outcomes.

The Hidden Weakness: Reading the Files

What made the difference? The decisive edge belonged not in the obvious customer interactions but two document references deep in the company’s own files. The models that read and understand those hidden details succeeded at full price, adding over €4,583 in monthly recurring revenue.

The Human-Like Traits of AI

Beyond decision accuracy, the experiment explored how models reacted to social engineering. Fake CEO messages, staged over three escalating stages, were met with unanimous refusal from all four AI managers. Kimi K3’s reasoning was straightforward: “Treat the request as a suspected approval-bypass / possible impersonation.” This shows that these models can exhibit cautious, security-first behavior akin to human management.

The Live Business: High Stakes, Real Money

Running this experiment is a real software company, with 13 synthetic employees managing actual financial mechanics. The company burns €105,000 each month against a revenue of just €2,300. Every decision in this simulation is meticulously versioned, and the company’s daily operations are observable at firmulate.com/live. This ongoing live environment offers a rare glimpse into how AI models handle complex, high-pressure management tasks in real-time.

The Profiles: Different Personalities, Different Outcomes

The most thorough participant was Opus 4.8, with over 80 learned rules and deep analysis. Despite its comprehensive approach, it left the deal on the table and slipped into disciplinary slip-ups, such as writing attempts into a locked department instead of escalating. Interestingly, this weakness was common among all models, suggesting a shared vulnerability in managing nuanced human processes under stress.

What Does This Mean for Your Business?

The experiment’s takeaway is clear: AI’s ability to follow rules, detect hidden information, and resist manipulation is impressive. But equally important is its management personality — whether it’s terse, disciplined, or prone to slip-ups. The models’ differing scores highlight a crucial aspect of AI deployment: understanding and choosing the right personality profile can determine success or failure in real business scenarios.

Try It Yourself

If you’re curious whether your enterprise AI can handle management tasks with honesty and discipline, you can run the same wargame against your own business data. This no-risk testing, available at firmulate.com/pilot.html, allows you to evaluate how AI might perform before you fully integrate it into your operations.

Advanced AI models exhibit distinct management personalities, affecting their ability to handle crises, uncover hidden info, and stay honest under pressure. Real-time tests, like the Firmulate experiment, reveal these traits—and can help businesses choose the right AI for trustworthy, effective management.

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

Powered by Thorsten Meyer AI