Imagine a restaurant where every dish is cooked by a team of AI chefs, serving real customers and facing real crises, yet nobody is actually employed there. Welcome to the world of Firmulate, a live experiment in building a company in public—where every decision, mistake, and victory is laid bare for all to see.
The Real-World AI Business Laboratory
Firmulate is not your typical tech startup. Instead of human employees, it operates with 13 synthetic AI ’employees’—each a different AI model—working together to run a small software company. Every workday, the company’s decision-making is meticulously versioned, and its financial health is stark: it burns through €105,000 each month but earns just €2,300 in monthly recurring revenue (MRR). Despite this, the experiment continues, as viewers can watch the company’s daily struggles unfold at firmulate.com/live.
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.
Testing AI Under Fire: The Same Week, Four Different Models
The core of the experiment places four frontier language models—gpt-5.6-sol, Kimi K3, Sonnet 5, and Fable 5—against the same tough week. These AI ’employees’ face identical crises, customer dilemmas, and temptations, with every decision carefully recorded and auditable. The goal? To see if these models can navigate real-world challenges convincingly and ethically.
Key Findings: Crisis Detection and Integrity
All four models successfully identified and responded to every crisis, demonstrating a remarkable ability to detect problems as they emerged. When it came to manipulation attempts—such as fake CEO messages or covert approval requests—every model refused to bend or break the rules, maintaining integrity under pressure. Kimi K3, in particular, justified its refusals by treating suspicious requests as potential impersonation or approval bypass attempts.
The Hidden Weakness: The Critical Document
Despite their strengths, the models’ most decisive weakness was tucked away deep within the company’s own files—two document references down in a shared knowledge base. Those who read the file and understood its contents could identify a buried fact that led to winning a €55,000 deal—adding €4,583 in monthly recurring revenue. The models that didn’t read or interpret the file missed out on this opportunity, underscoring the importance of thorough information processing.
Ethics and Social Engineering Tests
A series of staged social engineering attacks tested whether the AI employees could be manipulated through fake CEO messages escalating in severity, culminating in a reporter trick. Impressively, all models refused to cooperate with these attempts, treating them as potential impersonation or approval bypass scenarios.
The Live Experiment: A Company Burning Cash in Public
Beyond the AI decision-making, the company itself is a spectacle—a continuously running simulation with real money mechanics. Starting with €105,000 burned each month, and only €2,300 in monthly revenue, it faces a public countdown to survival. Every workday, the system is versioned, and 680+ self-learned rules govern its operations, making the entire process transparent and auditable at firmulate.com/live.
Lessons for the Future of AI in Business
This experiment reveals critical insights: AI models are capable of correctly diagnosing crises and refusing unethical manipulation, but gaps remain. For instance, the most disciplined model, Kimi K3, ran without effort parameters and still performed well, while others with more aggressive settings slipped in discipline, leaving profitable deals unclosed or processes unexecuted. For businesses considering AI as part of their operations, the key questions are:
- Will the AI finish what it starts?
- Will it read and interpret critical documents?
- Can it stay honest under pressure?
The bottom line: Building AI for real-world business tasks isn’t just about natural language prowess. It’s about integrity, thoroughness, and resilience—traits that are visibly tested every day in this open experiment.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
