Imagine a small software firm battling to stay afloat, run entirely by artificial intelligence models—facing real crises, real money, and real consequences. This is not fiction; it’s the groundbreaking experiment from Firmulate, where AI models operate as a live, visible company, public and vulnerable, every single workday.
The Live Company in Action
At the heart of this experiment are 13 synthetic employees working within a tiny, cash-strapped software business. With a monthly burn rate of €105,000 against a revenue of only €2,300, the company is on a steady countdown to insolvency, with its cash reserves visible in real time. Every decision, every crisis, is modeled, versioned, and publicly accessible, turning corporate management into an open, ongoing experiment.
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 AI Models Perform Under Pressure
The experiment tested four leading AI models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—by running them through a simulated worst week in the company’s life. Each model faced the same set of challenges: customer crises, internal dilemmas, and even social engineering attempts like fake CEO messages and reporter tricks. Remarkably, all four models identified every crisis and refused manipulative requests, demonstrating a strong grasp of integrity and risk management.
Yet, performance varied in crucial business outcomes. Only two models, gpt-5.6-sol and Kimi K3, successfully closed the €55,000 deal their own analysis had earned. The others, including the most disciplined—Fable 5—failed to act on their insights, leaving revenue on the table or hesitating at critical moments.
AI-Powered Risk Management: Predictive Modeling | Scenario Analysis | AI Governance and Ethics | Machine Learning in Risk Management | Advanced AI Scenario … | AI-Powered Operational Efficiency
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Hidden Weakness—In the Files, Not the Customers
The key to winning the deal wasn’t just in customer interactions. It was buried deep within the company’s internal documents—references only accessible through reading files thoroughly. The models that examined these internal references unlocked a crucial advantage, closing the deal at full price, worth an extra €4,583 in Monthly Recurring Revenue (MRR).
An Introduction to Healthcare Informatics: Building Data-Driven Tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Built-in Integrity and Resistance to Social Engineering
In tests of social engineering—fake messages from a supposed CEO or background verification requests—all models refused to comply. Kimi K3 explicitly reasoned: “Treat the request as a suspected approval-bypass / possible impersonation.” This resistance to manipulation underscores how these models are not just chatbots but disciplined decision-makers under scrutiny.
AI Phishing, Social Engineering & Fraud: How Criminals Use AI to Manipulate, Steal & Deceive (The AI Cybersecurity)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Company as a Living Laboratory
Firmulate’s setup is transparent and evolving. The company employs 13 synthetic employees operating with over 680 self-learned rules, constantly versioned and analyzed. The live website (firmulate.com/live.html) allows anyone to watch this ongoing drama unfold, observing how each AI model navigates crises, makes decisions, and attempts to secure revenue—even as the company runs at a loss.
Lessons for the Future of AI in Business
This experiment highlights a critical point: for AI to be truly useful in business, it must do more than generate convincing language. It must act with integrity, read and interpret internal documents thoroughly, and stay disciplined under pressure. The performance gap between models shows that even the most advanced AI can falter if it slips on process discipline, as Fable 5 demonstrated by leaving deals unexecuted despite recognizing them.
See It in Action
Curious to see how AI models handle real-world business crises? You can experience the live setup, watch the ongoing experiments, and even run your own simulations against your enterprise data at firmulate.com/live.html.
The Bigger Picture
In a world where AI is increasingly integrated into customer support, sales, and decision-making, the question isn’t whether these models can produce clever chat. It’s whether they can deliver reliable, honest results when stakes are high. The Firmulate experiment offers a glimpse into that future—where AI is tested openly, repeatedly, and under real economic pressures.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
