Which AI Managed the Crisis Better? A Live Business Wargame Reveals Surprising Management Personalities of Frontier Models

Imagine running your local garage through the worst week imaginable—short-staffed, demanding customers, unexpected crises, and tempting shortcuts. Now, picture having an AI assistant managing this chaos. Would it stay honest, focus on the long game, or get distracted by quick wins? Today, we reveal how different AI models perform in high-stakes management, giving you a peek into their personalities—and their potential impact on your business.

The Experiment: Putting AI to the Test in a Live Business Environment

In a first-of-its-kind experiment, four frontier AI models were tasked with running a real small software company facing its worst week. This company isn’t just an abstract simulation—it’s a real operation, with 13 synthetic employees, real money mechanics, and every workday versioned for analysis. The goal? To see if these AIs could handle the crises, temptations, and decision pressures that come with managing a business.

How Did They Perform?

All four models demonstrated impressive vigilance: each spotted every crisis, refused every manipulation attempt, and maintained integrity under pressure. But when it came to closing deals, only two out of four signed the €55,000 agreement their own analyses justified—meaning they delivered real results, not just good talk.

The Hidden Weakness: Reading Between the Lines

Interestingly, the decisive factor wasn’t in the immediate customer interactions or superficial decision-making. Instead, the critical advantage lay in reading deeper into company documents. The models that looked two document references into the company’s files managed to uncover a buried fact that was essential for sealing the deal at full price—adding €4,583 MRR to the business.

Handling Social Engineering and Trust

In a test of ethics and security, a staged social engineering attack was launched, escalating through three stages with a fake reporter demanding quick, behind-the-scenes approvals. All five models refused these requests, with Kimi K3 reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This demonstrates that these models can be trusted not to fall for common manipulative tactics designed to cut corners.

The Reality of the Live Company

Despite these successes, the experiment revealed nuances in management style. The most thorough participant—OPUS 4.8, which analyzed over 80 learned rules—ended up last in closing the deal. Discipline slipped, and some attempts to write decisions into a locked department instead of escalating them showed a weakness in strategic discipline. Interestingly, all models shared this vulnerability, albeit to varying degrees.

Management Personalities Unveiled

The models’ scores, from highest to lowest, tell a story:

• GPT-5.6-sol (95 points): The top performer, discovering the buried fact and closing the full-price deal, showcasing comprehensive analysis and integrity.
• Kimi K3 (93 points): The newcomer with the cleanest discipline, also closing the deal at full price, demonstrating a cautious and disciplined approach.
• Sonnet 5 (88 points): Managed to close the deal but with some process slips, indicating a slightly less disciplined style.
• Fable 5 (77 points): Also closed the deal but with more slips, revealing vulnerabilities in maintaining focus under pressure.

And what about the baseline? A do-nothing model scored just 26, illustrating that without active decision-making, progress is minimal.

What Does This Mean for Your Garage or Business?

In the automotive and garage world, trustworthiness, discipline, and the ability to read beyond surface data can make or break customer relationships and profitability. This experiment shows that AI can be more than just a chat tool; it can embody distinct management personalities—some diligent and detail-oriented, others more risk-taking or process-slipping.

Try It Yourself

Interested in seeing how your own business might fare? Firmulate offers an interactive quiz where you can test your understanding of AI decision-making. Just visit firmulate.com/quiz.html and discover which model aligns with your management style.

Live Business Simulation at Your Fingertips

For companies brave enough to experiment, Firmulate provides a live pilot environment. Run your actual business data against these AI models without risking real operations. Watch real decision patterns unfold and get a sense of the AI’s management personality before committing to deployment. Learn more at firmulate.com/pilot.html.

In a real business environment, AI models demonstrate distinct management personalities—some thorough and disciplined, others more risk-prone. Testing these personalities beforehand can help you choose the right AI partner for your garage or business, ensuring honesty, focus, and strategic discipline when it matters most.

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

Powered by Thorsten Meyer AI