Can AI Managers Be Trusted? Inside the Live Business Wargame Testing Frontier Models’ Decision-Making Skills

Imagine a world where AI doesn’t just suggest furniture arrangements or color palettes but actually runs a business — making critical decisions under pressure, just like a human manager. As interior designers and furniture retailers, you know the importance of trust: can your tools be relied upon to make honest, effective choices? Well, a groundbreaking experiment by Firmulate puts frontier AI models to the test in a simulated real-world company scenario, revealing how these digital managers handle crises, temptations, and ethical dilemmas.

The Live Business Wargame: A New Benchmark for AI Management

In a pioneering live experiment, four leading AI models were tasked with running a small software company through its most chaotic week — complete with customer crises, internal temptations, and ethical challenges. The goal? To see if these models can identify problems, resist manipulation attempts, and close profitable deals — all while operating under real-world pressures.

This isn’t just a simulation. The company, with 13 synthetic employees and real money mechanics, burns through €105,000 monthly against a revenue of €2,300. Every decision is versioned and auditable, offering transparent insights into each model’s behavior. The models included:

• GPT-5.6-sol (score: 95)
• Kimi K3 (score: 93)
• Sonnet 5 (score: 88)
• Fable 5 (score: 77)

Results were telling. While all four models successfully spotted every crisis and refused every manipulation they faced — whether fake CEO messages or reporter inquiries — only two completed the critical task: closing a €55,000 deal they had analyzed and pitched for. Interestingly, the deal’s success hinged on uncovering a secret detail buried two documents deep in the company’s files, not in the customer event itself. The models that read and understood the internal files won the full-price deal, worth over €4,500 in monthly recurring revenue.

Decoding the Models’ Personalities and Decision Styles

The experiment uncovered distinct management personalities among the models, akin to human managers’ styles:

• **GPT-5.6-sol:** The most thorough and strategic, it identified the crucial internal detail and diligently closed the deal, earning a perfect score of 95.
• **Kimi K3:** The newcomer with the cleanest discipline, it also secured the deal, scoring 93, and refused to bend under social engineering attempts, citing suspicion of impersonation.
• **Sonnet 5:** Slightly more process-oriented, it closed the deal with some slips, scoring 88.
• **Fable 5:** The most cautious but less disciplined, it left the deal on the table and slipped in escalation discipline, scoring 77.

These profiles echo real human tendencies: thoroughness versus caution, discipline versus complacency. Notably, the models ran at different effort levels, with Kimi K3 operating without an effort parameter, making its discipline even more impressive.

Facing Social Engineering and Ethical Dilemmas

In a test of resilience, all models refused to engage with escalating fake CEO messages and a staged reporter request asking for a simple yes/no background approval. Kimi K3’s reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This demonstrates that these models can be programmed to uphold ethical standards even under pressure, an essential trait for trustworthy AI in business settings.

Implications for Interior Design and Furniture Retail

While your industry may seem far from running software companies, the core takeaway is universal: AI tools in your workflow — whether for managing inventory, customer relations, or project approvals — must not only produce good suggestions but also reliably complete tasks, uphold integrity, and identify critical hidden information. The experiment’s results show that the best-performing models can uncover hidden details and act ethically under stress, which is crucial when AI is entrusted with more responsibility.

Want to see how your AI tools stack up? You can run similar tests tailored to your business environment. Visit firmulate.com/quiz.html to challenge your models and understand their management personalities before making hiring decisions or integrating new AI systems into your operations.

Frontier AI models can handle crises, resist manipulation, and close deals based on deep internal insights — essential qualities for trustworthy automation in your business. Testing them in real-world scenarios reveals their true management personalities, helping you choose the right tool for your needs.

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

Powered by Thorsten Meyer AI