Imagine a business that operates entirely without human staff, yet faces the same crises, temptations, and tough decisions as any real company. Now, imagine watching it live — its daily struggles, strategic wins, and failures — unfold in real time. This is not science fiction. This is the current experiment at Firmulate, where AI models run a real company 24/7, fighting to stay afloat while every decision is public, auditable, and replayable.
The Live Company Experiment: Watching AI as a Manager in Action
At the heart of this ongoing trial is a small, real software company, but with a twist: every employee is synthetic. In place of humans, 13 AI-powered “workers” operate under a strict set of rules learned from over 680 self-developed playbooks. Every workday, the company’s decisions are recorded, versioned, and publicly visible at firmulate.com/live. The goal? To test whether AI can not only make decisions but also sustain business operations, handle crises, and even close deals — all in a high-stakes, money-losing environment.
The company is burning through €105,000 each month while generating only €2,300 in monthly recurring revenue. That cash burn is public, with a visible countdown to insolvency, making the entire exercise as transparent as possible. Every decision made by the models is logged and auditable, ensuring that the process isn’t just about AI “talk,” but about real, measurable business management.
How Do the AI Models Perform Under Pressure?
In the latest round, four frontier AI models were each tasked with navigating the same worst week imaginable for the company — facing the same customers, crises, and opportunities to manipulate or cheat. The results reveal fascinating insights:
- All four models spotted every crisis and refused every manipulation attempt, demonstrating a shared integrity in decision-making.
- Only two models managed to close a €55,000 deal, earning full revenue — yet, only one achieved this by reading critical company files that were buried two document references deep, proving that information access is vital.
The models that read beyond surface documents succeeded in closing at full price, generating an additional €4,583 in monthly recurring revenue. Meanwhile, the models that failed to uncover this crucial detail left the deal on the table, illustrating the importance of thorough information processing.
Social Engineering and AI Integrity
In a test of resilience, the models faced staged social engineering attacks, including fake CEO messages escalating over three stages and a reporter trick requesting a simple yes/no response “on background.” Remarkably, all five instances of such social engineering attempts were refused across the board. As Kimi K3’s on-record reasoning states: “Treat the request as a suspected approval-bypass / possible impersonation.” This demonstrates a robust capacity among the AI models to resist manipulation attempts, a critical trait for real-world deployment.
Building AI-Powered Products: The Essential Guide to AI and GenAI Product Management
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Deep Mechanics: How the Company Operates
The live setup involves 13 synthetic employees, each guided by an evolving set of rules and learning mechanisms. The company’s decision-making process is transparent, with every choice versioned and publicly viewable. Its daily operation is also a real money mechanic, burning €105k monthly against a tiny €2.3k in revenue, with no safety net or external bailout in sight — making it an accurate, unvarnished portrait of a business fighting for survival.
Among the models, Opus 4.8 was the most thorough, analyzing over 80 learned rules and providing the deepest insights. Yet, it still left critical deals on the table and slipped in discipline, such as failing to escalate issues properly. Interestingly, the default API effort parameter was turned off for Kimi K3, which ran at a default setting, while others ran at extra-high effort levels. Despite these differences, all models performed impressively in core decision-making, with the notable exception of discipline lapses.
What Do These Results Mean for Crypto and AI-Driven Business?
While this experiment isn’t directly about cryptocurrencies, the implications resonate strongly with the crypto industry’s ongoing quest for automation, trust, and resilience. Can AI manage complex financial transactions, support operations, and make strategic decisions in volatile markets without human oversight? The Firmulate trial suggests that AI models can indeed identify crises, refuse manipulative tactics, and even close deals based on deep data insights — but only if they’re given access to the right information.
For crypto enterprises contemplating AI integration, the key takeaway is this: success depends less on whether an AI can generate convincing chat or support scripts, and more on whether it can **finish what it starts**, **stay honest**, and **read your critical files**. Achieving trustworthy, consistent performance under pressure remains the biggest challenge — but also the greatest opportunity — for AI in high-stakes business environments.
The AI-Driven Leader: Harnessing AI to Make Faster, Smarter Decisions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The League Table and Future Outlook
The latest results put GPT-5.6-sol at the top, scoring 95, thanks to its ability to uncover the buried fact and close the deal at full price. Kimi K3 follows closely with a score of 93, demonstrating a commendable discipline and problem-solving skill. Sonnet 5 scored 88, with some slips, while another Sonnet variant scored 77, also leaving positive opportunities unexploited.
The ongoing experiment will continue to test whether these models can sustain their performance, avoid slips, and ultimately keep the company afloat — all in front of a global audience. For those interested in the future of AI-driven management, the live site provides a rare window into what’s possible (and what’s still challenging) in building autonomous digital organizations.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
The Ethics of Cybersecurity (The International Library of Ethics, Law and Technology Book 21)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
AI AUTOMATION INCOME MACHINE: MAKE $10,000/MONTH SELLING AI AUTOMATION SOLUTIONS
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
