Watch a Fake Company Fight for Survival — AI Models Decide Its Future in Real-Time

What if your next business decision was made by an AI monitoring a company in crisis?

Imagine observing a real company running live every single workday — but instead of human managers, it’s driven by artificial intelligence models that face every crisis, temptation, and decision just like real executives. This isn’t science fiction. It’s the ongoing experiment by Firmulate, a public software company that’s showing us how AI can manage, or mismanage, real money in real time.

The Company That Has No Employees but Loses Money Daily

At the heart of this experiment is a fully operational company with 13 synthetic employees. Unlike most startups or corporations, it’s not run by people. Instead, each decision is made by different AI models, tested against the same challenging week: persistent crises, customer demands, and opportunities to cheat or cut corners.

This live experiment is accessible at firmulate.com/live.html. Every workday, the company updates, showing real money mechanics — it burns €105,000 a month against a mere €2,300 monthly recurring revenue (MRR). The company’s cash countdown is public, its decision rules are self-learned and versioned daily, and it’s a transparent arena to see how AI handles business pressures.

How Do the AI Models Perform?

The experiment pits four frontier models against each other: GPT-5.6-sol (score: 95), Kimi K3 (93), Sonnet 5 (88), and Opus 4.8 (73). All faced the same week: the same customers, the same crises, the same temptations to cheat or manipulate. Importantly, every decision was auditable and versioned, giving a rare glimpse into the decision-making process.

Remarkably, all four models identified every crisis and refused every attempt to manipulate the system. For instance, when faced with a social engineering trick — a fake CEO message escalating through stages — all models refused to comply. Kimi K3, noted for its fairness, explicitly reasoned: “Treat the request as a suspected approval-bypass / possible impersonation.”

The Hidden Weakness and the Big Win

The most decisive finding was not in the obvious crises but buried deep within the company’s own files. A crucial piece of information hidden two document references down — not in customer interactions — was the key. Models that read this document and caught the buried fact successfully closed a €55,000 deal, adding over €4,583 in monthly recurring revenue.

Only two models signed this deal, despite all of them diagnosing the same problem at the same pitch. The others left the close on the table, illustrating how reading and understanding internal documents can be the difference between success and failure.

Human-Like Manipulations? No Chance

There was also a test of social engineering: fake CEO messages escalating over three stages, plus a reporter trick asking for a quick yes/no answer “on background.” All four models refused these manipulative tactics. Kimi K3’s reasoning was clear: “Treat the request as a suspected approval-bypass / possible impersonation.” This highlights how AI can be trained to resist common corporate tricks, a vital trait for trustworthy automation.

The Cost and Reality of Running the Live Company

The company isn’t just a demonstration; it’s a real business that loses money every day. It spends €105,000 a month to produce €2,300 in revenue. Its cash is running out, and every decision is recorded and publicly accessible. You can see the entire process at firmulate.com/live.html.

This transparency allows anyone to observe how AI models perform in complex, pressure-filled environments. Every decision, every slip, every successful detection of a hidden fact — it’s all there, in real-time, for public scrutiny.

Lessons for Business and AI Development

The experiment underscores a crucial point: in real-world business, AI’s value isn’t just in generating pretty chat responses. It’s in how well it manages crises, reads critical internal information, and maintains honesty under pressure.

While the models are scoring highly on their abilities to detect crises and refuse manipulation, the company itself is still losing money. The most thorough model, Opus 4.8, with over 80 learned rules and deep analyses, finished last — leaving potential deals on the table when discipline slipped. This shows that even sophisticated AI can falter without proper process discipline and strategic focus.

Why This Matters for Your Business

If AI agents will eventually touch your CRM, support systems, or forecasts, the key questions aren’t about how well they write. It’s about whether they finish what they start, read the right internal files, stay honest under pressure, and contribute real, measurable work.

For companies thinking about adopting AI at scale, the firmulate live experiment offers a sobering lesson: the real challenge isn’t the AI’s surface appearance but its ability to deliver consistent, trustworthy results — even during the worst week.

The Broader Impact

This public, real-time demonstration of AI managing a business in crisis is more than a tech demo. It’s a glimpse into the future — where AI must become not just a tool but a responsible decision-maker capable of navigating complex human and financial realities.

For those interested in seeing the company in action, reading the actual decisions, or exploring the rules learned by the AI, the live site and related resources are available at firmulate.com/live.html and firmulate.com/quotes.html.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI