Imagine trying on a new outfit based solely on a glossy image — only to find it doesn’t fit when you wear it for real. In the world of AI, the same applies to evaluating performance. Chat demos can be seductive, but they often hide the true test of an AI’s reliability: its ability to deliver when it counts. That’s exactly what a recent experiment by Firmulate revealed about the AI models shaping tomorrow’s business decisions.
Listen free for 30 days with Audible
Thousands of audiobooks and originals — cancel anytime.
As an affiliate, we earn on qualifying purchases.
Testing AI Beyond the Promises of Chat
In a groundbreaking experiment, four leading AI models were tasked with running a small software company through its most tumultuous week — with real money on the line and crises escalating fast. Every move, decision, and response was carefully recorded and compared. The goal? To see which AI truly understood the company’s challenges and could execute a real business deal, not just generate convincing chat responses.
The Setup
The models involved were:
- gpt-5.6-sol (score: 95)
- Kimi K3 (score: 93)
- Sonnet 5 (score: 88)
- Fable 5 (score: 77)
All four faced identical crises, from customer complaints to internal trust breaches, and were subjected to social engineering tactics designed to manipulate them. Crucially, each decision was versioned and auditable, making the process transparent and comparable.
What the Models Achieved
Remarkably, all models identified every crisis and refused every manipulation attempt — a testament to their defensive capabilities. Yet, only two managed to close a €55,000 deal with the company, earning what their own analysis indicated was deserved revenue. The other two, despite understanding the situation, failed to follow through or slipped in discipline.
The Hidden Weakness
Hidden within the company’s files was a crucial piece of information — buried two document references deep. It was this detail that tipped the scales in favor of the successful models. Those that read the internal files before acting secured the full deal, unlocking an additional €4,583 in monthly recurring revenue.
Why Demos Can Be Deceiving
This experiment exposes a critical blind spot: chat demos often showcase a model’s ability to generate convincing language but do little to reveal whether it can complete actual work, read essential documents, or maintain discipline under pressure. The models’ true strength — or weakness — becomes evident only when tested in real-world scenarios.
Trust Under Pressure
When subjected to social engineering, all models refused to be manipulated — a vital trait in protecting businesses. Kimi K3 exemplified this best, explaining: “Treat the request as a suspected approval-bypass / possible impersonation.” This discipline underscores a model’s potential to safeguard organizational integrity, not just produce plausible responses.
The Live Company: A Business in the Red
The company used in the experiment is real: 13 synthetic employees, managing real money mechanics — burning €105k per month against a modest €2.3k in revenue. It operates every weekday, with over 680 self-learned rules, and its progress is fully observable at firmulate.com/live. Watching this company in action highlights the importance of testing AI in operational contexts, not just chat.
As an affiliate, we earn on qualifying purchases.
Lessons for Business Leaders
The key takeaway: the true measure of an AI’s usefulness isn’t how well it chats — it’s whether it can deliver real results, read critical documents, and resist manipulation under stress. The models that signed the deal demonstrated these qualities, while others left opportunities on the table despite understanding the problem.
Beyond the Surface
The results suggest that businesses should look beyond superficial chat demos and evaluate AI in operational scenarios. The ability to finish what it starts, access vital information, and stay honest under pressure determines actual business value.
How to Test Your AI
Firmulate offers a unique approach: running a digital twin of your company through a controlled war game. This lets you see how your AI workforce performs in a realistic, risk-free environment, ensuring you choose solutions that can truly deliver, not just impress in demos. Discover more at firmulate.com.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
enterprise AI document reading tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI cybersecurity social engineering protection
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
