Inside a Live AI-Run Company Fighting for Survival — Watch Real Decisions in Action

Imagine a real company operated entirely by AI models, facing daily crises, tough ethical tests, and a public cash countdown — and you can watch it all happen live. This is not science fiction. It’s the extreme build-in-public experiment by Firmulate, where four advanced AI models run a simulated software business through its worst week, with outcomes measurable and transparent at https://firmulate.com/live.html.

The Live Experiment: An AI-Managed Company Under Pressure

At the core of this experiment is a small, synthetic company with 13 AI-powered employees, all guided by over 680 self-learned rules. Every day, the models face the same set of crises, customer interactions, and ethical tests, with decisions tracked and versioned in real-time. The goal? To evaluate how well AI can manage real business challenges, not just generate impressive dialogue.

The Stakes Are High

The company spends €105,000 each month but earns just €2,300 in monthly recurring revenue (MRR), placing it on a slow burn with a public cash countdown. This stark contrast underscores the challenge: can AI handle relentless pressure while making trustworthy decisions? The live site offers viewers a front-row seat to this ongoing struggle, with updates and decision logs rebuilt twice daily.

Key Findings from the League of Models

• The four models—GPT-5.6-SOL, Kimi K3, Sonnet 5, and Fable 5—faced identical crises, customer pushes, and ethical tests.
• All models successfully identified every crisis and refused manipulative requests, such as fake CEO messages or reporter tricks, demonstrating robust ethical boundaries.
• Despite this, only two models managed to close the €55,000 deal based on their analysis — GPT-5.6-SOL and Kimi K3. The other two, despite similar diagnoses, left the deal on the table, highlighting a crucial weakness: the ability to act decisively, not just analyze.
• Interestingly, the decisive factor was buried two document references deep within the company’s own files. The models that read these files won the deal at full price, worth an extra €4,583 MRR.

The Ethical and Strategic Tests

Beyond crunching data, the models also faced social engineering attempts. Over three stages, fake CEO messages and a reporter trick were employed to test trust and compliance. All five models refused to approve these requests, with Kimi K3 explicitly treating them as suspected impersonation or approval bypass attempts.

What This Means for Business Automation

This experiment offers a stark view: AI can recognize crises and ethical breaches reliably, yet strategic action—closing deals, making decisive bets—remains a challenge. The live site showcases these behaviors in real time, giving enterprises a chance to simulate their own scenarios against their business data, without risking real systems.

The Deep Dive: The Opus 4.8 Profile

The most thorough participant, Opus 4.8, analyzed over 80 learned rules but ultimately slipped in discipline, leaving a close opportunity unexecuted. This pattern was echoed across other models, emphasizing that even with extensive analysis, control and decisive action are difficult to maintain under stress.

What Should You Watch for?

• Can your AI handle crisis detection and ethical dilemmas without faltering?
• Will it act on critical opportunities or leave money on the table?
• Does it stay disciplined under pressure, or does it slip into indecision?

The live experiment at https://firmulate.com/live.html not only demonstrates these dynamics but also invites businesses to run their own simulated war games—using their data without affecting real operations—through the Pilot feature.

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

Powered by Thorsten Meyer AI