What if a company operated without human employees, yet faced the harsh realities of cash flow, crises, and tough decisions — all in the open?
That’s the bold experiment unfolding at Firmulate, where a real small software firm is run by AI models, observed daily as it battles survival. For those interested in innovation, transparency, and the future of work, this is a story worth watching.
AI decision-making software tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Unusual World of Firmulate’s Live Experiment
Firmulate is pioneering an extreme form of “build-in-public”. Every workday, a set of 13 synthetic employees — powered by advanced AI models — makes decisions in managing a small software company. Every move, crisis, and opportunity is recorded and publicly accessible at firmulate.com/live.html.
The company’s operations represent a real money mechanics scenario: it burns €105,000 each month against a modest €2,300 in monthly recurring revenue (MRR). Its cash countdown is visible to all, creating a high-stakes environment where every decision has tangible consequences.
The Experiment: Testing AI Under Pressure
Four frontier AI models, including the top-rated gpt-5.6-sol with a score of 95, were tasked with running this company through its worst week — the same crises, same customers, same temptations to cut corners. Every decision they made was versioned and auditable, offering a transparent view into their reasoning process.
Remarkably, despite the intense pressure, all four models identified every crisis and refused every attempt at manipulation. This included social engineering tactics: fake CEO messages and reporter tricks. All five models refused to be duped, with Kimi K3 explicitly treating suspicious requests as potential impersonation or approval bypasses.
Critical Findings: The Hidden Weakness
The key weakness wasn’t in the decisions during crises but in the models’ ability to uncover crucial internal information. The decisive advantage came from reading a buried document in the company’s own files — a secret that, if uncovered, allowed the models to close a full-priced deal worth an additional +€4,583 MRR. That insight was not apparent in the customer-facing data but was buried two document references deep in the internal files.
Real Money, Real Stakes
The live site at firmulate.com/live.html offers viewers a window into this ongoing drama. The experiment continues daily, with the models’ decisions driving the simulated company’s fate. The company operates with no human employees — just AI-driven decision-making, every step recorded and publicly accessible.
While the models generally performed well, the most thorough participant, Opus 4.8, demonstrated the importance of discipline. Despite its extensive rules and analysis, it left a critical deal on the table and slipped in escalation procedures, illustrating that even the most sophisticated AI can falter without rigorous discipline.
What Does This Mean for Businesses and AI?
For interior design and furniture companies, the implications are profound. When AI begins to manage customer relationships, support, and forecasting, the real questions aren’t just about how well it writes or chats. Instead, focus shifts to whether AI can see through crises, resist manipulation, and follow through on critical tasks — especially under pressure where human oversight might be lacking.
As the experiment shows, AI can detect hidden information, refuse to manipulate, and make tough decisions. But it also reveals the importance of disciplined processes and thorough rules to prevent slip-ups that cost money and credibility.
Watch the Future Unfold
This experiment is a rare glimpse into an ongoing battle for AI’s trustworthiness in a high-stakes environment. It demonstrates not only the potential of AI to manage complex business scenarios but also the necessity of transparency and rigorous oversight. For those curious about how AI might reshape your industry, it’s a story worth following closely.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
