Inside an AI-Run Company That’s Building Publicly, Losing Money, and Still Surviving

Imagine a real company run entirely by artificial intelligence, with no human employees, publicly battling for survival on a daily basis. This is not science fiction—it’s the live experiment conducted by Firmulate, a unique venture that is transforming how we think about AI, business, and transparency.

The Live Company That Moves in Real Time

At the heart of this experiment is a small, entirely synthetic company operated by 13 AI-driven ’employees.’ This setup isn’t just a simulation; it’s a live, publicly accessible business that burns €105,000 each month while generating only €2,300 in monthly recurring revenue. Visitors can watch the company’s daily struggles unfold at firmulate.com/live, making it perhaps the most transparent showcase of AI decision-making in a real-world context.

The AI Models Facing the Same Tests

Four frontier AI models were each tasked with navigating the company’s worst week—same customers, same crises, same temptations to cheat or manipulate. These models are part of the Crucible League, a competitive benchmark where AI systems are scored on their ability to handle complex, high-stakes situations. The scores range from 77 to 95, with the top performer, GPT-5.6-sol, scoring 95 out of 100, demonstrating exceptional crisis detection and decision-making.

Decoding the Results

• All four models identified every crisis and refused manipulation attempts, showing strong integrity under pressure.
• Only two models successfully closed the €55,000 deal their own analysis indicated was achievable. The other two, despite diagnosing the opportunity correctly, left the deal on the table.
• The key weakness was buried in internal company files, not in the customer interactions. The models that read these files thoroughly secured the best outcome, closing the deal at full price (+€4,583 MRR).
• Social engineering tactics, including fake CEO messages and reporter tricks, were all consistently refused by the AI. Kimi K3 explicitly treated such requests as potential impersonation or bypass attempts.

Building in Public: Transparency and Learning

The experiment is fully versioned and auditable, with every decision logged and analyzed. The company’s daily operations are openly visible, and every AI model version is tracked. The company’s ‘playbook’ includes over 680 self-learned rules, constantly evolving to handle crises and negotiate deals. This transparency is part of an extreme form of build-in-public — demonstrating how AI can be tested and improved in real-world conditions.

The Human Element and Discipline

Interestingly, the most thorough AI model, Opus 4.8, with over 80 learned rules and deep analysis, finished last in the league. It left potential deals unexecuted due to discipline lapses, such as writing attempts into a locked department instead of escalating them. This highlights a crucial point: even the most sophisticated AI can struggle with operational discipline, especially under stress.

What This Means for Business and AI

For enterprises considering integrating AI into their operations, the key takeaway is not just about chat quality or superficial performance. It’s about whether AI can finish what it starts, read critical internal data thoroughly, and stay honest when stakes are high. The question isn’t if AI can generate convincing conversations but if it can reliably deliver useful, ethical work at a cost-effective rate.

This groundbreaking experiment by Firmulate exposes the realities of AI decision-making in a live company. It shows that AI can identify crises and resist manipulation, but operational discipline remains a challenge—highlighting what businesses need to consider when deploying AI at scale.

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

Powered by Thorsten Meyer AI