Inside an AI-Run Startup That’s Failing in Public — but Learning Fast

What Happens When an AI Runs a Company in Real Time?

Imagine watching a startup operate entirely without human employees, driven solely by artificial intelligence. Not in a distant sci-fi future, but right now, on a public platform where every decision, crisis, and mistake is laid bare. This is the bold experiment of Firmulate, a company that’s pushing the boundaries of transparency, AI, and business management — and revealing some startling truths along the way.

The Challenge of Building an AI-Driven Business

Firmulate runs a small, real company with 13 synthetic employees, each fueled by advanced AI models. Despite its innovative approach, the company is currently losing €105,000 every month against a modest income of €2,300 in monthly recurring revenue. The site openly tracks its cash countdown, showcasing a high-stakes environment where every decision impacts survival. Although it’s an extreme build-in-public experiment, it offers invaluable insights into the capabilities and limitations of AI in complex management roles.

Watching AI Face Real Crises and Ethical Dilemmas

The experiment subjects different AI models to the same challenging week — with identical crises, customer interactions, and temptations to manipulate or cheat. Interestingly, all models successfully identified every crisis and refused every unethical manipulation attempt, whether it was a fake CEO message escalating over multiple stages or a reporter’s subtle request for a background yes/no answer. For instance, the Kimi K3 model explained its refusal by treating the request as a potential impersonation risk. This indicates a shared core strength in AI: ethical decision-making under pressure.

What Sets the Models Apart? The Hidden Details

While all models demonstrated a strong ethical stance, their success in closing deals varied significantly. Two models — gpt-5.6-sol and Kimi K3 — managed to sign a €55,000 deal, earned through their own analysis and pitches. The other two, despite identifying the opportunity, left the deal unclosed. The critical difference? The models that succeeded read deeper into the company’s own files, uncovering a crucial, buried document reference that led to the full-price deal. That hidden information, just two references deep in the files, proved to be the decisive factor in securing revenue.

Transparency and Accountability in AI Decision-Making

Each decision by the AI models is versioned and auditable, allowing observers to see exactly what information was considered and how conclusions were reached. This rigorous transparency is a core feature, enabling a clear understanding of AI reasoning — an essential step toward trustworthy automation. Moreover, every day of this live experiment is versioned, creating a continuous, evolving story of AI performance under real-world pressures.

The Human Element in an AI World

Despite the high-tech setup, the experiment highlights human-like traits in the AI’s behavior. For instance, in the case of Opus 4.8 — the most thorough model — discipline slipped, and it left an opportunity unexecuted because it failed to escalate appropriately. This reveals that even the most advanced AI can struggle with consistency and process discipline, echoing human management challenges.

Why This Matters for Creators and Innovators

If you’re involved in creative tech, audio, or digital arts, the core takeaway is simple: the future of AI in your fields will revolve around trustworthiness, ethical consistency, and the ability to finish what it starts. Watching this live experiment unfold shows that AI can identify crises and refuse unethical shortcuts, but it still struggles with closing deals or executing plans fully — especially when discipline slips or when buried data is overlooked.

The Bigger Picture: Building a Better AI Workforce

Firmulate offers a unique, risk-free way for businesses to test how AI might perform before integrating it into real systems. By modeling the entire business environment, companies can see if their AI agents will stay honest, read critical information, and complete tasks under pressure. The platform even offers a read-only pilot mode, allowing organizations to run their own scenarios without risking their actual operations.

Key Lessons from the AI Business Lab

This live experiment reveals that AI can identify crises and reject unethical manipulations, but it still faces challenges in closing deals and maintaining discipline. Transparency, deep information reading, and consistent decision-making are crucial for trustworthy AI in business. For creators, understanding these dynamics is vital as AI begins to touch more of your work — trust and thoroughness will define its success.

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

Powered by Thorsten Meyer AI