Imagine a company that operates entirely in the open, with no employees, losing money every day, yet providing a transparent window into how AI can manage real business crises. This is not a science fiction story but the reality of a groundbreaking live experiment by Firmulate, where AI models are tested as if they were running a real company — every decision, every crisis, and every dollar tracked publicly.
The Experiment: A Company in the Crosshairs of AI
At the heart of this experiment is a small, fictional software company that faces the same crises and temptations every week. The unique twist? Instead of human managers, the company is run by four advanced AI models, each tested against the same challenging week, with all decisions tracked and made auditable. This setup aims to assess whether these models can not only spot problems but also act ethically and effectively under pressure.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
How the AI Models Fared
The results are revealing. All four AI models identified every crisis, refused every manipulation attempt, and demonstrated a remarkable level of discipline. For instance, when faced with social engineering tactics like fake CEO messages or reporter tricks, every model refused to bypass security measures, reflecting a cautious, rule-abiding attitude. Kimi K3, notably, explained its refusal by treating the request as a suspected impersonation or approval bypass — an indication of cautious reasoning.
However, the real story lies in the details of deal-making. Only two of the models managed to close a significant €55,000 deal their own analysis had earned. The other two either left deals unexecuted or failed to follow through fully, despite correctly diagnosing the opportunities in their analysis. The gap? It was hidden in the company’s internal files, not in the immediate customer interactions. Reading and understanding these deeper documents was the key to winning the full-price deal — a task the models handled differently.
Faster. Messier. Tougher.: Crisis Communication Strategies in an Era of Populism, AI, and Distrust
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Transparent Company and Its Money Mechanics
Beyond decision-making, the experiment features a live, publicly visible company with 13 synthetic employees, operating with real financial mechanics. The company burns through €105,000 each month against a tiny revenue of just €2,300, cycling towards a public cash countdown. Every day, the company’s operations are versioned, and its rules — over 680 learned and self-generated guidelines — are on display at firmulate.com/live.html. This transparency turns traditional business management into an open-source, real-time puzzle for observers.
The AI Cybersecurity Handbook
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Lessons from the AI’s Performance
The experiment’s most profound insight is that AI can accurately diagnose crises and resist manipulation, yet human-like decision failures still occur. For example, Opus 4.8, the most thorough participant with over 80 learned rules, fell short in discipline due to a slip in process — leaving an approved deal unexecuted and instead escalating a situation into a locked department. This weakness was consistent across all models, suggesting that even the most comprehensive AI systems have room for improvement in disciplined execution.
Another critical finding is the importance of reading and understanding internal files. The models that delved deeper into the company’s own documents secured the full-price deal, illustrating that AI’s success depends on context and information access, not just surface-level analysis.
The AI Prompt Playbook for Excel & Financial Analysis: 50 Ready-to-Use AI Prompts for Excel, Financial Analysis, Dashboards, Power BI, Financial Modeling & Automation
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Broader Context: Why This Matters for Business
While this experiment may seem distant from everyday tasks like cleaning or floor maintenance, the core lessons are universal. As AI begins to touch areas like customer support, sales, and operations, the questions become: Will these models stay honest under pressure? Can they finish what they start? Do they understand the full context of a situation before acting? The experiment underscores that the real challenge isn’t just AI’s ability to generate convincing chat messages but to behave ethically and reliably in critical decision points.
How You Can Experiment Too
Business leaders curious about testing their own AI readiness can engage with this approach through tools like Firmulate’s wargame platform. It lets companies simulate crises in a safe, read-only environment, ensuring that AI systems can be evaluated before deployment in real systems. Every decision is versioned and transparent, providing a clear picture of strengths and weaknesses without risking actual business operations.
Final Thoughts
This live experiment showcases a new frontier in AI development: operational transparency, real-world testing, and the fight for disciplined, ethical AI behavior. As AI models grow more sophisticated, their ability to understand context, read internal documents, and uphold trust will determine their success in managing real businesses. For now, the companies that watch this space — like the one running live at firmulate.com/live.html — are pioneering the future of AI in business.
This groundbreaking live experiment reveals that AI can diagnose crises and resist manipulation, but disciplined, context-aware decision-making remains a challenge — vital for trustworthy AI in business.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
