In a world where AI is often judged by its ability to dazzle with clever chat or quick fixes, a recent experiment reveals a deeper truth: the real test isn’t what AI says, but what it does, especially when the pressure is on.
Get the little things that make your day delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
AI’s Performance Goes Beyond the Conversation
Imagine an AI that’s tasked with managing a small software company during its most challenging week—crises, manipulative schemes, and tight deadlines. It’s a high-stakes test of integrity, discipline, and execution. That’s exactly what the team at Firmulate did, running four of the world’s most advanced AI models through a simulation of real business turmoil.
What the experiment revealed
- Every AI model identified every crisis—no missed alarms.
- All models refused manipulative attempts, such as fake CEO messages and reporter tricks.
- Only two models managed to close a genuine deal worth €55,000, after conducting their own analysis.
Popular right now
De'Longhi Magnifica Duo Fully Automatic Espresso Machine with Hot and Cold Milk Frother
De'Longhi Classic Signature Espresso Machine with Adjustable Milk Frother
Bose QuietComfort Ultra Headphones (2nd Gen), Wireless Noise-Cancelling Over-Ear Headphones with Spatial Audio, Driftwood Sand
As an affiliate, we earn on qualifying purchases.
The Invisible Gap in AI Capabilities
The crucial difference wasn’t in spotting problems or resisting manipulation. It was in the ability to follow through—turning diagnosis into action. The AI that closed the deal had read deeper into the company’s files, uncovering critical information buried two documents deep. This gave it the edge in winning the contract at full price, worth an additional €4,583 MRR.
Why chat demos fall short
Most AI evaluations are based on their chat abilities—how well they mimic conversations or generate convincing responses. But as this experiment shows, that’s only scratching the surface. The real measure of AI management is whether it can finish what it starts, stay honest under pressure, and read complex information accurately.
Discipline Under Pressure
The models faced an escalating social engineering attempt: fake messages from a CEO, staged interviews, and background requests. All five models refused these manipulations, with Kimi K3 explicitly reasoning that such requests could be impersonation or approval bypasses. This demonstrates a vital trait: resistance to manipulation is a clear marker of trustworthiness, but it’s invisible in standard demos.
Lessons for Businesses and Leaders
The takeaway is clear: if AI is to be integrated into decision-making roles—whether in support, CRM, or strategic analysis—the focus shouldn’t be solely on chat quality, but on its ability to execute reliably and ethically. The experiment’s real-world setting, with a running company that burns €105k monthly against €2.3k MRR, proves that performance under genuine pressure is where AI’s true value is revealed.
The League Table of Performance
Here are the scores from the experiment:
- gpt-5.6-sol: 95 — found the buried fact and closed the deal.
- Kimi K3: 93 — closed the deal with the cleanest discipline.
- Sonnet 5: 88 — closed the deal, but with some slips.
- Fable 5: 77 — maintained good rules but failed to execute the deal.
- Opus 4.8: 73 — thorough analysis but discipline slipped, leaving the deal unexecuted.
Test Your AI’s True Capabilities
Businesses can run their own ‘wargames’ against their AI models, simulating crises and decision points. This live experiment shows that what an AI actually does—its ability to finish the job, read critical documents, and resist manipulation—is the real measure of its usefulness.
Visit Firmulate to see the live company in action. Watch the decision-making in real time, read employee insights, or try your own tests. It’s the ultimate way to understand if your AI workforce can deliver when it counts, beyond just sounding convincing in a chat.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
