Watch an AI-Run Company Struggle and Succeed in Real Time — Live and Unfiltered

Imagine a smart home system that doesn’t just automate your devices but operates a real company, making decisions, facing crises, and even losing money — all in full public view. This is no sci-fi; it’s the live experiment by Firmulate, revealing how AI models handle complex, high-stakes business situations in real time.

The Live Business in Action

At the heart of Firmulate’s bold experiment are 13 synthetic employees working as a complete, functioning company. They tackle genuine challenges like customer crises, trust dilemmas, and even manipulative tactics. Every decision they make is documented and versioned daily, creating a transparent, ongoing story of AI management in a real-world setting.

How the Experiment Works

Four cutting-edge AI models, including the well-known GPT-5.6-SOL, are tested simultaneously on the same set of problems during their worst week. These include real customer issues, internal crises, and ethical tests like social engineering attacks. The goal? See if the AI can detect issues, make honest decisions, and close deals where appropriate.

Key Findings

• All four models identified every crisis and refused manipulative attempts, showing resilience and integrity.
• Only two models successfully closed a €55,000 deal their own analysis had earned, demonstrating ability to recognize and act on hidden opportunities.
• Crucially, the decisive advantage was in reading deeper into internal company files — a detail hidden in the company’s own documents, not visible in customer communications. Those models that read and analyze these internal references won the full-price deal, worth over €4,500 in monthly recurring revenue.
• Social engineering attempts, like fake CEO messages and reporter tricks, were vigorously refused by all models. Kimi K3, one of the models, explained that such requests are treated as impersonation risks.

The Reality of a Money-Losing Company

The live experiment company operates on a razor-thin margin: it burns €105,000 every month against a modest €2,300 in monthly revenue. Its public cash countdown adds urgency, illustrating the high stakes of AI decision-making in a real economic environment.

Deep Dive into the Models

The most thorough participant, Opus 4.8, analyzed over 80 learned rules and performed detailed assessments but still fell short, leaving deals on the table and slipping in discipline. This underscores the challenge AI faces, even with extensive training, in executing complex management tasks reliably under pressure.

Implications for Home and Smart Technologies

If AI systems managing entire companies can succeed or fail based on integrity and thorough internal analysis, what does this mean for smart home devices and automation? These systems are increasingly integral to our daily lives, and their ability to operate honestly, read critical internal data, and resist manipulation is vital for safety and trust.

The Future of AI in Business and Home

This experiment highlights that AI decision-making isn’t just about generating human-like responses. It’s about honesty, thorough analysis, and resilience under pressure. For homeowners, it’s a reminder that the smart devices and systems we trust must be capable of integrity and reliable performance, especially when facing unexpected crises or security threats.

Firmulate’s ongoing public experiment offers a rare transparent window into how AI can be tested and improved in high-stakes environments. As these models evolve, they could influence how we design smarter, safer, and more trustworthy home automation systems in the future.

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

Powered by Thorsten Meyer AI