Inside a Live Experiment: Can AI Run a Company and Keep Its Wallet Open?

What if your workout partner was an AI that never took a break, but kept losing money every day?

Imagine a fitness regimen where the trainer is an AI, guiding you through routines, spotting crises, and making decisions—all in real time. Now, replace the gym with a tiny software company managed by AI, and you get a glimpse of a radical experiment happening right now. This isn’t science fiction; it’s a live demonstration of AI’s capacity to run a business, with all its messiness and realities, right in front of your eyes.

The Live Business Arena: An AI Company Under Fire

At the heart of this experiment is a virtual company, operating with 13 synthetic employees and real money mechanics. Every day, it burns through 105,000 euros—more than its monthly revenue of just 2,300 euros—while its digital team navigates crises, customer demands, and ethical dilemmas. This setup, public and transparent, is designed to test how well AI models can handle the chaos of real business operations.

Testing AI’s Crisis Management and Integrity

The experiment challenges four different AI models, including the leading GPT-5.6-sol, Kimi K3, Sonnet 5, and Opus 4.8, to guide the company through its worst week. Each model faces identical scenarios—same customers, same crises, and the same temptations to cheat or manipulate. The decisions are fully versioned and auditable, ensuring transparency in how each AI responds.

Remarkably, all four models identified every crisis and refused every attempt at manipulation, including social engineering tactics like fake CEO messages and reporter tricks. For instance, when presented with a staged approval bypass, all models refused, citing suspicion of impersonation or bypass risk.

Findings That Matter Beyond the Demos

Despite their competence, only two models actually closed a €55,000 deal they had earned through careful analysis—Kimi K3 and the more thorough Opus 4.8. The critical factor? The models that read deeper into company files uncovered a hidden, decisive document reference. This buried fact led to the full-price deal—a difference of over €4,500 in monthly recurring revenue.

This illustrates a vital point: AI’s effectiveness isn’t just about surface-level chat or superficial decision-making. It’s about digging into the details that matter, the stuff hidden in plain sight, and making choices based on comprehensive understanding.

The Built-in Public Showdown

This isn’t a controlled lab experiment; it’s a real, running company that burns 105,000 euros each month against a tiny income. The live site at firmulate.com/live.html offers a window into this ongoing story where every decision, every crisis, and every attempt at manipulation is recorded and public.

The Limitations and Lessons

One of the notable insights came from Opus 4.8, which used over 80 learned rules for analysis but still left opportunities on the table — like not escalating critical issues properly. Meanwhile, the K3 model, running without a default effort parameter, performed well but was slightly less disciplined. These nuances highlight that AI performance depends heavily on the rules, parameters, and discipline built into it.

What This Means for Business and Fitness

For fitness enthusiasts, the takeaway is clear: success isn’t just about how smart or fast your trainer is. It’s whether your trainer consistently handles crises, reads the fine print, and stays honest when under pressure. Similarly, in business, AI’s value lies in its ability to finish what it starts, read your files thoroughly, and remain trustworthy—especially when temptation and pressure mount.

The Future of AI-Managed Companies

Today’s experiment shows that AI can recognize crises, refuse manipulation, and even close deals based on detailed insights. But it also reveals weaknesses—like lapses in discipline or missing hidden clues—that can be critical in real-world applications. The question isn’t whether AI can write well, but whether it can deliver reliable, honest work day after day.

As these AI models continue to evolve and companies become more comfortable testing them in live settings, the line between virtual and real business management blurs. The experiment at Firmulate is a bold step into that future—one where AI’s capacity to operate ethically, effectively, and transparently is put under the same scrutiny as your personal fitness routine.

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

Powered by Thorsten Meyer AI