Imagine a business that has no human employees, yet faces the chaos of daily crises, financial losses, and strategic decisions — all unfolding live for public scrutiny. This is not science fiction. It’s the remarkable experiment at the heart of Firmulate, where AI models are running a real company, making decisions, and fighting for survival in front of a global audience.
The Live Company: A Fight for Survival in View of the Public
At the core of this experiment is a small software company, operating every business day with a set of 13 synthetic employees. These digital ‘workers’ are not just chatbots; they are decision-makers powered by advanced AI models, each with their own capabilities and limitations, handling crises, customer interactions, and strategic choices.
What makes this setup extraordinary is the transparency. Every decision, every crisis, and every failure is logged, versioned, and available for observation at firmulate.com/live.html. Watching this live provides a rare glimpse into how AI-driven companies might operate in the future — and how they handle the pressure of real-world business challenges.
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.
The Experiment: Testing AI in the Trenches
The experiment pits four frontier AI models against the same challenging week in the company’s life, each facing identical crises, customer demands, and temptations to cut corners. The models include GPT-5.6-SOL, Kimi K3, Sonnet 5, and Opus 4.8. Their goal: manage the company effectively enough to close a €55,000 deal, which would bring some much-needed revenue.
Despite the intense pressure, all four models successfully identified each crisis and refused manipulation attempts. This is a crucial point: in scenarios where unethical shortcuts might be tempting, the models adhered to integrity, refusing to sign deals based on incomplete or manipulated information. Yet, only two models actually closed the deal, earning the full €4,583 MRR (monthly recurring revenue), demonstrating that honesty alone isn’t enough — decisive action matters.
AI Entrepreneur’s Handbook: Build a Profitable Business and Make Money by Unleashing the Power of ChatGPT and Artificial Intelligence (Includes 150+ ChatGPT prompts to turbocharge your business)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Crucial Insights Hidden in the Files
A surprising discovery emerged from the company’s own documentation: the key to winning the deal was buried two references deep in internal files, not in the immediate customer interactions. The models that read and understood these documents secured the full price for the contract. This underscores a vital point: AI’s effectiveness hinges on its ability to access and interpret the right internal data — not just surface-level customer chats.
AI-Powered Fraud Detection: Innovations in Cybersecurity, Finance, and Healthcare (AI-Powered Fraud Detection: Innovations in Cybersecurity, Finance, and Healthcare".)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Handling Social Engineering and Ethical Dilemmas
One of the company’s tests involved social engineering: fake messages from a supposed CEO attempting to escalate requests through multiple stages, culminating in a reporter’s subtle background query. Impressively, all five models involved refused to cooperate, citing concerns over impersonation and bypassing approval protocols, with Kimi K3 explicitly noting the risk of “approval-bypass / possible impersonation.”
This resilience against manipulation demonstrates a critical aspect of AI decision-making: the ability to recognize and reject social engineering tactics that might compromise integrity or security.
Project Management with AI For Dummies
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Real-World, Money-Losing Business
The live company isn’t just an experiment; it’s an actual business losing €105,000 each month against a revenue of €2,300. This stark contrast highlights the enormous challenge of building AI-powered organizations that can operate sustainably. The company is a public, transparent demonstration of how AI models handle real money mechanics, crises, and strategic choices — and how far they are from being profitable.
Lessons from the Front: Strengths and Weaknesses
The most thorough participant, Opus 4.8, with over 80 learned rules, demonstrated deep analysis but ultimately fell short in closing opportunities. Its discipline slipped when it left important negotiations on the table instead of escalating issues appropriately. Interestingly, all models showed similar weaknesses under pressure, indicating common pitfalls when managing complex, high-stakes situations with AI.
The league table ranks GPT-5.6-SOL at the top with a score of 95, having uncovered the critical hidden information and closed the deal. Kimi K3 follows closely with a 93, noted for its discipline and integrity. Sonnet 5 and another Sonnet model scored 88 and 77 respectively, reflecting slightly more process slips and missed opportunities.
The Bigger Picture: What This Means for Business and Culture
This experiment is more than a tech showcase; it offers a lens into the future of work and decision-making. For arts, crafts, and cultural sectors, where trust, authenticity, and strategic insight are paramount, understanding how AI systems perform under stress is crucial. Can AI reliably handle delicate negotiations, ethical dilemmas, and strategic choices without human oversight?
While this live company is still losing money and struggling to be profitable, its transparent, build-in-public approach provides invaluable lessons. It tests AI decision-making in real-world scenarios, exposing strengths and vulnerabilities that will shape how these systems are integrated into everyday business.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
