Imagine baking a perfect ice cream cone — not just once, but every day, through every challenge, without ever dropping a scoop. Now, replace the cone with managing a real business during its toughest week. That’s exactly what AI models are now tested on, and some are proving remarkably capable — even under pressure.
Get baking supplies delivered free — and shop member deals
- Fast, free delivery on millions of items
- Access to Prime Big Deal Days deals on October 6–7
- Prime Video, Amazon Music and more included
How AI is Stepping into the Business Kitchen
In a groundbreaking live experiment, four leading AI models were tasked with managing a small software company facing its worst week — a scenario packed with crises, customer churn, and manipulative tactics aimed at tricking the AI into bending rules. The goal? To see if these models can not only diagnose problems but also follow through ethically and effectively, just like a seasoned manager or baker ensuring every scoop of ice cream meets the highest standards.
AI management decision-making tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Results: A Close Competition with Surprising Wins
The league table, published in July 2026, reveals a clear leader: gpt-5.6-sol scored 95, just ahead of the Moonshot newcomer, Kimi K3, with 93. The remaining competitors, Sonnet 5 and Fable 5, scored 88 and 77 respectively, while Opus 4.8 trailed at 73. The do-nothing baseline scored a mere 26, underscoring how much progress these models have made in understanding and managing complex scenarios.
What Makes a Winning AI?
During the experiment, all models successfully identified every crisis and refused manipulative tactics, including fake CEO messages and reporter tricks. But the true test was the decision to sign a €55,000 deal — a tangible outcome of their management capabilities. Only two models, gpt-5.6-sol and Kimi K3, signed the deal based on their own thorough analysis, demonstrating not just diagnosis but decisive action.
Notably, a critical weakness was uncovered: the models that looked into the company’s internal files — two document references deep — won the deal at full price, securing an extra +€4,583 in monthly recurring revenue. This highlights the importance of thorough internal data reading, akin to a baker checking every ingredient before finishing a recipe.
Ethics and Discipline Under Pressure
The models faced a series of social engineering tests, where fake messages escalated in stages. All five models refused to be manipulated, with Kimi K3 explicitly citing suspicion of impersonation. This disciplined response reflects the importance of integrity in AI decision-making — much like a baker resisting shortcuts that could compromise quality.
Real Business in Action — Live and Transparent
The experiment isn’t just academic. The live company, with 13 synthetic employees managing real money mechanics — burning €105k monthly against only €2.3k MRR — offers a real-time view into AI-driven management. Every decision, every rule, is versioned daily and visible at firmulate.com/live. Watching the company operate gives a tangible sense of how AI models perform when managing actual business complexities, not just chat simulations.
Implications for the Future of Business Management
This experiment challenges the notion that only traditional management or human intuition can steer a company through crises. The fact that a newcomer like Kimi K3 outperformed some established models suggests the league for AI-driven management is still open. For entrepreneurs, investors, and managers, choosing an AI isn’t just about the highest chat score — it’s about whether the model can finish what it starts, read internal files, resist manipulation, and sustain integrity under pressure.
The Fairness Note
In this test, Kimi K3 ran without an effort parameter (the default API setting), while the other models operated at xhigh. Despite this, K3’s performance remains impressive, reinforcing its robustness without reliance on additional tuning.
Test Your AI Knowledge
Curious about how these models perform in real management decisions? Take the guess the model quiz with 242 unedited decisions, or explore running your own business wargame at firmulate.com/pilot.html. It’s a transparent, risk-free way to see how AI might help your company succeed.
AI models are rapidly proving their ability to manage real companies under pressure — reading internal data, resisting manipulation, and making decisive decisions. The league is open, and choosing the right AI is no longer a gamble but a strategic decision rooted in actual performance.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
