Can AI Run a Business Without Human Help—and Survive Its Own Crises?

What if an AI Could Run a Company—and Keep It Honest?

Imagine watching an artificial intelligence manage a real company day by day, facing crises, making decisions, and even losing money all in public view. This isn’t science fiction; it’s the ongoing live experiment from Firmulate, where AI models are tested as the sole managers of a small software firm in a high-stakes, transparent trial.

The Live Company: A Front-Row Seat to AI Management

At the heart of the experiment is a simulated company staffed by 13 synthetic employees and governed by an intricate web of over 680 self-learned rules. This company is real enough to burn through €105,000 each month — more than it earns, with a monthly recurring revenue (MRR) of just €2,300. Every decision, crisis, and temptation is versioned and publicly available for scrutiny at firmulate.com/live.html.

The Challenge: Same Week, Different Models

Four frontier AI models — including the latest and most thorough, Opus 4.8 — each managed the same worst-week scenario, facing identical customer issues and ethical temptations like social engineering. These models were tested in a brutal environment designed to reveal their true management capabilities, not just their chat skills.

Critical Findings: Crisis Detection & Ethical Integrity

All models identified every crisis, from technical failures to customer complaints, and refused every manipulation attempt, including social engineering tactics like fake CEO messages or reporter tricks. For instance, five out of five models rejected a staged fake request asking for a quick approval — a testament to their built-in safeguards.

The Hidden Weakness: The Power of Internal Files

The decisive advantage was lurking within the company’s internal files, just two document references deep. Models that read these files thoroughly closed the deal at full price, adding +€4,583 in monthly recurring revenue. Those that missed this buried fact left the deal unclosed, losing potential profit.

Real Money Mechanics & Public Accountability

This experiment isn’t just about AI prowess; it’s about managing real money in a transparent environment. The firm operates with a public cash countdown, burning €105,000 monthly to chase a tiny €2,300 MRR, with every decision openly versioned and available for review.

The Deep Dive: Different Models, Different Outcomes

Opus 4.8, the most thorough participant with over 80 learned rules, showcased the depth of analysis but still fell short on closing deals — often leaving opportunities on the table due to discipline lapses. Meanwhile, the newer Kimi K3, which ran without an effort parameter, managed to close the deal with the cleanest discipline, highlighting the importance of setup and parameters in AI management.

Management Under Pressure: Ethical and Strategic Decisions

The experiment also tested social engineering resilience. All models refused to be manipulated through staged requests, underscoring their built-in resistance against dishonest tactics. This is crucial for real-world applications where AI might face such manipulations in customer support or sales.

Why You Should Watch & Care

For investors and business leaders, the core question isn’t whether AI can generate polished chat responses — it’s whether AI can reliably finish tasks, read critical internal information, and act ethically under pressure. These are the qualities that determine if AI can truly manage vital aspects of a business or support system.

The League Table & Performance Scores

• gpt-5.6-sol: 95 — Found the buried fact and closed the deal, delivering full performance.
• Kimi K3: 93 — Managed to close the deal with the cleanest discipline.
• Sonnet 5: 88 — Closed the deal but with some process slips.
• Fable 5: 77 — Maintained best rules discipline but failed to execute the deal.

These results underscore that even the most capable models can falter without proper discipline or strategic setup, emphasizing the importance of continuous oversight and transparency.

Learn More & Watch Live

If you’re curious about the potential of AI to manage your business processes or want to see this experiment in action, visit firmulate.com/live.html to watch the live company in operation. Additional insights, detailed scores, and the decision quiz are available at firmulate.com/quotes.html.

Key Takeaway: AI Management Is Transparent, Challenging, and Still Flawed

This ongoing live experiment reveals that while AI models can recognize crises and resist manipulations, they still struggle with closing deals and maintaining discipline. For business leaders, it highlights the importance of transparency, rigorous testing, and continuous oversight when deploying AI in critical roles—especially in environments where money and trust are on the line.

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

Powered by Thorsten Meyer AI