Watch an AI-Run Startup Struggle to Survive in Real Time

Imagine running a small outdoor gear shop where every decision counts—yet your team has no humans, only AI models making every call. Now, imagine watching this company fight for survival day after day, with real money on the line and its future hanging in the balance. That’s the reality behind Firmulate, a public experiment that puts artificial intelligence through its paces in a real business environment.

The Live Company: An Unconventional Business in Action

At the heart of this experiment is a fully functioning, if unconventional, company with 13 synthetic employees—AI models that make decisions, handle crises, and try to generate revenue. Right now, it burns through €105,000 every month while only earning €2,300 in monthly recurring revenue. This stark financial reality sets the stage for a relentless test of AI management—an ongoing story visible to anyone interested in the future of AI and business.

How the Experiment Works

Each day, the company faces simulated crises, customer requests, and ethical dilemmas. The same set of challenges is presented to four leading frontier AI models, which are tasked with navigating the company’s worst week. Every decision they make is versioned and made auditable, allowing observers to track how each model responds under pressure.

Key Findings from the AI Models

• All four models identified every crisis and refused every manipulation attempt, demonstrating integrity and awareness.
• Only two models successfully signed the €55,000 deal their own analysis earned—showing they can close sales at full price under pressure.
• The decisive advantage was in the models’ ability to read and understand critical internal documents—information buried two reference points deep in the company’s files.
• This buried knowledge helped the models close deals worth an additional €4,583 monthly recurring revenue.
• In social engineering tests—such as fake CEO messages escalating over stages—every model refused to be duped, with one citing suspicion of impersonation or approval bypass.

Real Money, Real Stakes

This isn’t a lab experiment; it’s a visible, live running business. The company’s operations are transparent, with every decision, rule, and crisis available for public review at firmulate.com/live.html. The site updates twice daily, showing how each of the 13 synthetic employees handles the complex, messy realities of running a company with a finite budget.

What the Data Reveals

Among the four models tested, Opus 4.8 was the most thorough yet performed the poorest. It learned over 80 rules but left opportunities on the table, slipping into internal communication instead of escalating issues. Kimi K3, operating without certain effort parameters, showed the sharpest discipline and successfully closed the deal, indicating that even in the most controlled settings, AI decision-making quality can vary widely.

Implications for Business and AI

The experiment underscores a critical point: AI’s value in a commercial setting isn’t just about generating polished responses or engaging chat. It’s about whether AI can finish what it starts, read the right information, and stay honest under pressure. The models that demonstrated integrity and thoroughness are the ones that closed the deals—spotting hidden information and refusing manipulative tactics.

Why This Matters for Outdoor and Travel Companies

For outdoor brands and travel firms, the takeaway is clear: integrating AI into your operations isn’t just about automation or fancy chatbots. It’s about deploying models that can handle your real-world crises, make ethical decisions, and ultimately deliver measurable results. Watching firms like Firmulate run their companies live offers a glimpse into a future where AI management can be transparent, accountable, and aligned with business goals.

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

Powered by Thorsten Meyer AI