How AI Is Running a Real Company in Public — and Facing Its Limits
Imagine a company with no human employees, where every decision is made by artificial intelligence, and you can watch its struggles unfold live. This is not a sci-fi story but a real-time experiment that reveals how AI handles crisis, trust, and survival in the business world.
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 Live Experiment: An AI Company Under Pressure
At the heart of this experiment is a small software firm operated entirely by AI models, each acting as a synthetic employee. Every workday, the company faces simulated crises—customer issues, internal dilemmas, and even manipulative tactics—identical across all models. The goal: see how well these models navigate real-world business challenges while being held accountable for decisions.
This live setup, hosted at firmulate.com/live.html, is an ongoing transparency effort. It showcases a company burning through €105,000 monthly against a modest €2,300 recurring revenue, with a publicly visible cash countdown. The entire operation, including over 680 self-learned rules, is versioned daily—each decision and strategy recorded for review.
AI for Solo Lawyers: A Practical Guide to AI Tools that Save You Time and Grow Your Practice (AI for Professionals)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
What the Models Achieve—and Fail To Do
The experiment’s results are striking. All four AI models tested — from the most advanced GPT-5.6-sol to newcomer Kimi K3 — identified every crisis and refused manipulative attempts, such as fake CEO messages or secret deals. This indicates a high level of situational awareness and ethical resistance. For example, in a staged social engineering attack, all models refused to approve a suspicious request, with K3 explaining: “Treat the request as a suspected approval-bypass / possible impersonation.”
However, the models’ ability to close deals varied significantly. Only two models managed to sign the €55,000 contract, which their own analysis had recommended. Interestingly, the decisive advantage was hidden in a buried document reference within the company’s files—something that models which read this file successfully capitalized on, securing an extra €4,583 in monthly recurring revenue.
AI Incident Response Systems: Crisis Management AI | AI Security Playbooks | Digital Forensics Enhanced | AI-Driven Incident Management | AI Forensic Innovations | Automated Security Solutions
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Decision-Making Under Pressure and the Cost of Errors
The live company demonstrates not just decision accuracy but discipline. For instance, Opus 4.8, the most thorough model with over 80 learned rules, was the best at analysis but ultimately left a close deal on the table due to internal missteps—writing attempts into a locked department rather than escalating them. This reveals that thoroughness alone isn’t enough; disciplined execution matters.
Crucially, the experiment shows that when models are tested under pressure, their performance varies—not just in decision quality but in their willingness to cheat or escalate. The AI models refused to be manipulated or bypass rules, demonstrating a level of integrity that is often missing in human decision-making under stress.
CLAUDE CODE MASTERY: The Complete Step-by-Step Guide to AI-Powered Software Development, Agentic Coding, Automation, Debugging, Testing, MCP Integration, … (TekkyVille's AI Series Book 15)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for Human Business Leaders
For investors and business owners, especially those involved in assets like precious metals, this experiment offers a stark insight: AI’s ability to handle crises ethically and effectively is still being tested in real time. If AI agents are to manage customer relationships, forecasts, or compliance, their capacity to complete tasks honestly and thoroughly is critical.
The experiment also underscores that success isn’t just about generating chat responses; it’s about executing real work—reading files, making decisions, and following through on commitments. The models that read deeper and act decisively, like GPT-5.6-sol and Kimi K3, are closer to being reliable in a business context, even if they still face discipline challenges.
The Broader Picture and Future Outlook
This experiment is a bold step towards building transparent AI systems capable of real business management. It’s a public, ongoing story of AI’s strengths and weaknesses—an invaluable resource for anyone considering AI solutions for complex, high-stakes environments.
Interested in how this plays out? You can watch the live company at firmulate.com/live.html or explore detailed quotes and decisions at firmulate.com/quotes.html. The experiment offers a rare glimpse into the future of AI-driven enterprise—where honesty, discipline, and strategic insight are put to the ultimate test.
Key Takeaway
This live AI company reveals that while models can identify crises and refuse manipulative tactics, executing complex deals and maintaining discipline under pressure remains a challenge. Watching this unfold offers vital lessons for integrating AI into real-world business operations—highlighting the importance of thoroughness, ethics, and disciplined execution.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
