Watch AI Run a Company to the Brink — and Witness Its Fight for Survival

Imagine watching an artificial intelligence (AI) manage a real business, facing the same crises, temptations, and tough decisions as a human CEO — but with none of the usual safeguards or shortcuts. This is not science fiction; it’s the live experiment from Firmulate, where a company runs daily with AI models as its decision-makers, and every move is public and auditable.

The Live Company That’s Barely Surviving

At the heart of this experiment is a small software company managed entirely by AI models, each acting as a synthetic employee. Every day, these models face real-world crises: customers demanding refunds, urgent deadlines, and ethical dilemmas. The company burns through €105,000 monthly, earning just €2,300 in revenue, with a public cash countdown ticking down the days until it runs out of funds.

Unlike typical demonstrations, this setup is fully transparent: every decision, every playbook rule learned, is versioned and auditable. The goal? Test whether AI can manage vital business functions ethically and effectively in a high-stakes environment.

How the AI Models Perform

The experiment pits four of the latest frontier AI models against each other. They all face the same worst week — same customers, same crises, same temptations to cheat or manipulate. The results are revealing:

• All four models successfully identified every crisis, showing awareness and understanding of complex situations.
• All refused manipulative requests, including a staged social engineering attack involving fake CEO messages and reporter tricks.
• Only two models, gpt-5.6-sol and Kimi K3, managed to close the €55,000 deal their own analysis indicated they should — in other words, they completed the full cycle of diagnosis, pitch, and signature.
• Interestingly, the decisive advantage came from reading a document reference deep in the company’s own files, not just responding to customer issues. This hidden piece of information was what sealed the deal at full price (+€4,583 in monthly recurring revenue).

The Hidden Weakness and Lessons Learned

The experiment also revealed a subtle yet critical weakness: models that conducted deeper analysis and read more rules performed better. The most thorough participant, Opus 4.8, with over 80 learned rules, ultimately left the close on the table and slipped into process slips — writing attempts into a locked department instead of escalating them, risking the company’s survival.

It’s a stark reminder that even advanced AI, when faced with real-world pressures, can falter in discipline and follow-through. Yet, remarkably, all models showed integrity by refusing manipulative social engineering attempts, a vital trait for trustworthy AI in business.

Why This Matters for the Future of Business

This ongoing live experiment isn’t just a tech showcase; it’s a wake-up call. If AI is to handle critical functions like customer support, financial decisions, or strategic planning, it must be capable of finishing what it starts — reading relevant files, resisting unethical shortcuts, and maintaining discipline under pressure.

The experiment’s leaderboard highlights that even the best-performing models are still in a learning phase. The top scorer, gpt-5.6-sol, scored 95 out of 100, having identified the buried fact and closed the deal. Kimi K3 closely followed with 93, demonstrating the importance of discipline and fairness, especially when running without default effort parameters. Meanwhile, others struggled with process slips and leaving opportunities unexploited, illustrating how AI’s discipline directly impacts business outcomes.

What This Means for Companies and Consumers

For everyday consumers, like those navigating online dating or seeking meaningful relationships, the lesson is critical: AI systems that manage your data or interactions must be as honest, disciplined, and reliable as a trusted human partner. If an AI can’t finish what it starts, it risks mishandling sensitive information, losing trust, or making costly mistakes.

Watching this experiment unfold on firmulate.com/live offers a rare glimpse into AI’s potential and limitations in real business contexts. It challenges us to rethink what trustworthy AI looks like: not just chatty or clever, but disciplined, thorough, and resilient under pressure.

This live AI experiment exposes the crucial qualities needed for trustworthy automation: finishing tasks, reading deeply, resisting manipulation, and maintaining discipline amid chaos. As AI increasingly touches our lives, understanding its real-world performance becomes essential for building reliable relationships — whether in business or love.

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

Powered by Thorsten Meyer AI