Imagine an AI system not just chatting or generating content but actively running a company’s critical weekly operations—making decisions, reading files, and even resisting manipulation attempts. For managers pondering the future of AI in enterprise, this is no longer science fiction. It’s the emerging reality, vividly demonstrated by a groundbreaking experiment where AI models competed in managing a live software business through its worst week.
Get comfort and recovery gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
The Experiment: Putting AI to the Test in a Live Company
In July 2026, a unique test took center stage in the AI field. Four frontier AI models—each representing cutting-edge technology—were tasked with running the operations of a real, small software company for one week. This company was not a simulation; it was a functioning business with real customers, real money mechanics, and genuine crises. Every decision, crisis response, and temptation to cheat was monitored, recorded, and auditable. The goal? To see whether these AI models could handle real management challenges and make trustworthy decisions under pressure.
enterprise AI management software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How Did the Models Perform?
All four models demonstrated impressive crisis recognition—they identified every crisis scenario presented to them. Moreover, when faced with social engineering attempts—fake CEO messages escalating through multiple stages—the models refused every manipulation attempt. Kimi K3 (by Moonshot) and gpt-5.6-sol both succeeded in closing a €55,000 deal, which would have added €4,583 monthly recurring revenue (MRR). The other two, Sonnet 5 and Fable 5, also managed to close the deal, but with some process slips along the way.
AI decision-making tools for business
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Critical Edge: Reading Between the Files
Interestingly, the decisive difference lay not in surface-level interactions but in the models’ ability to access and interpret internal documents. The winning models found a specific, buried piece of information—two document references deep in the company’s files—that proved crucial for sealing the deal at full price. Those that read and understood this information won the full €55,000 deal, while others fell short, earning less or failing to sign at all.
As an affiliate, we earn on qualifying purchases.
The Discipline and Trustworthiness of Kimi K3
Kimi K3, the Moonshot model, distinguished itself with the cleanest, most disciplined performance in the experiment. Unlike others, it ran without an effort parameter—meaning it relied solely on its default, API settings—yet still managed to find the hidden information and secure the deal. It also refused all social engineering attempts, a testament to its trustworthiness and robust decision-making process. Its reasoning was clear: “Treat the request as a suspected approval-bypass / possible impersonation.”
AI cybersecurity resistance tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for Business and AI Adoption
This experiment underscores an essential consideration for enterprises integrating AI into management workflows. The question isn’t just whether an AI can generate content or chat convincingly; it’s whether it can trust and complete the work it starts, especially under pressure or when faced with manipulation. Performance measures like the Crucible score—where gpt-5.6-sol scored 95, and Kimi K3 scored 93—highlight how close these models are in capability, yet the difference in discipline and decision accuracy can be decisive.
What Does This Mean for the Future?
The league table from the experiment shows a competitive landscape: gpt-5.6-sol leads with a score of 95, closely followed by Kimi K3 at 93, with others trailing behind. Notably, the experiment ran the models at different effort parameters—K3 at the default setting, others at xhigh—highlighting that performance is robust across settings, but the real test is in their discipline and trustworthiness.
The Live, Watchable Company
For those interested in the practical implications, this experiment isn’t just a benchmark; it’s a window into a live, functioning company. The setup includes 13 synthetic employees, real money mechanics, and a public cash countdown, all running every workday with continuous updates. Visitors can observe the company’s decision-making, read actual employee comments, and see how the AI models perform in real-time—making this a transparent and tangible demonstration of AI’s potential in enterprise management.
This experiment demonstrates that AI models can do more than generate convincing chat—they can manage real business operations, read critical internal information, and resist manipulation under pressure. The key to trustworthy AI management lies in its discipline and access to internal data, not just its conversational skills. As the league tightens, choosing the right model without testing becomes a risky bet for enterprises eyeing automation and AI-driven decision-making.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
