Imagine a real company operated entirely by artificial intelligence — making decisions under pressure, facing crises, and even risking its own survival. Now, see it all unfold live, every workday, with no filters or fiction. This is not science fiction; it’s the experimental world of Firmulate, where AI models are tested as complete companies in real-time, battling to stay afloat.
The Radical Experiment: Running a Business With AI
In an unprecedented move, the company behind this story uses AI models to simulate a small software business. These models, each representing a different frontier AI, are subjected to the same brutal week of crises, customer issues, and ethical dilemmas — all in a controlled, transparent environment.
Every decision made by the models is logged, versioned, and auditable. They face real operational challenges: managing customer relationships, handling internal conflicts, and navigating manipulative social engineering attempts. The goal? To see if AI can not only identify problems but also act with integrity and discipline, even under pressure.
Business Intelligence in the Age of AI: Modern Data Warehousing, Analytics and AI-driven Decision-Making
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Live Experiment and Its Surprising Results
This experiment is ongoing and observable at firmulate.com/live. The site shows a small, real company with 13 synthetic employees, burning €105,000 monthly against just €2,300 in recurring revenue, with a public countdown of cash remaining. Every workday, the company updates, showing the decisions of its AI workforce and how they handle the latest crises.
Four frontier models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Opus 4.8—were all given the same challenging week. Each faced the same customers, crises, and temptations. Remarkably, all four models identified every crisis and refused every manipulation attempt, including social engineering tricks designed to bypass approval processes.
Why Did Some Models Succeed More Than Others?
The key difference was in how they read the company’s internal files. The most successful model, gpt-5.6-sol, uncovered a critical piece of information buried two document references deep in the company’s files. That insight enabled it to close a deal worth over €4,500 in recurring monthly revenue, earning full credit for performance.
In contrast, Opus 4.8, though the most thorough participant with over 80 learned rules, left the deal on the table by slipping into disciplinary silos—writing attempts into a locked department instead of escalating them. The same weakness was observed in other models, revealing a common blind spot: failure to escalate when discipline was slipping.
AI In Cybersecurity: Simplifying Cyber Risk with Smart, Affordable Tools for Small Business Defense
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Real Money Mechanics and Ethical Challenges
The company’s daily operations involve complex money mechanics, burning €105,000 a month yet earning just €2,300 in recurring revenue. This intense financial pressure underscores the importance of honesty, discipline, and strategic decision-making in AI behavior.
To test moral boundaries, researchers staged social engineering attacks—fake CEO messages escalating in urgency and a reporter trick asking for quick approvals. All four models refused to be manipulated, with Kimi K3 reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This demonstrates a promising ability to resist unethical pressure, a critical trait for AI in business settings.
Trauma Doesn't Stop at the School Door: Strategies and Solutions for Educators, PreK–College
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for AI in Business
This live experiment highlights a fundamental truth: the ability of AI to identify crises and refuse manipulation is crucial. Success isn’t just about generating convincing language; it’s about consistent decision-making aligned with company goals and ethics, especially when stakes are high.
The leaderboard from the current league table shows that gpt-5.6-sol leads with a score of 95, having uncovered key information and closed the deal. Kimi K3 closely follows at 93, demonstrating impressive discipline, while Sonnet 5 and Opus 4.8 also participated, each with their strengths and weaknesses.
DIGITAL HEALTH FOUNDATIONS IN NURSING INFORMATICS & CLINICAL AI: Mastering Health Data, Decision Support, and AI Tools for Patient-Centered Care
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Beyond the Experiment: Building Trust and Reality Checks
What does this mean for real businesses? The answer is simple: AI agents operating in your CRM, support systems, or forecasting tools must be able to finish what they start — read critical information, stay honest under pressure, and avoid manipulation. The test here isn’t how well they chat but whether they deliver tangible, trustworthy results.
For those interested, the same AI-driven wargame can be run against your own business data in a read-only mode, ensuring that no real systems are impacted. Details are available at firmulate.com/pilot.html.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
