Imagine a real business, operating daily, losing hundreds of thousands of euros, yet still fighting to stay afloat. Now, picture this company run entirely by AI models, each decision transparent and auditable. This is not fiction—it’s the live experiment at firmulate.com/live.html.
Get art and craft supplies delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
Building a Company in Public, with No Employees and No Guarantees
At the heart of this experiment is a small but complex software company, now in its 423rd day of operation. It has no human employees—just 13 synthetic ‘workers’ powered by artificial intelligence models. Every decision they make is recorded, versioned, and scrutinized, creating a real-time story of how AI manages a business under pressure.
Popular right now
Breville Barista Express Espresso Machine, Black Sesame
Shark HydroVac MessMaster Corded Wet Dry Vacuum and Mop (WD161)
KitchenAid 5.5-Qt Bowl-Lift Stand Mixer, Porcelain White (KSM55SXXX)
As an affiliate, we earn on qualifying purchases.
The Challenge: Surviving a Week of Crises and Manipulations
Each AI model was tasked with running the company through its worst week: confronting the same customer issues, crises, and temptations. The models faced real-world hurdles, including social engineering attempts, such as fake CEO messages, designed to test their integrity.
What the Models Did—and Did Not—Do
- All models identified every crisis and refused manipulation attempts, including impostor requests.
- Only two models managed to sign a €55,000 deal based on their analysis, demonstrating the critical difference between diagnosis and execution.
- Deep within the company’s documents, there’s a buried fact—an overlooked but decisive detail—that, when uncovered, led to a full-price deal worth +€4,583 monthly recurring revenue.
Performance Scores and Insights
The experiment’s results are quantified through a competitive league:
- gpt-5.6-sol scored 95 and successfully closed the full deal, uncovering the hidden fact.
- Kimi K3, the newcomer, scored 93 and also closed the deal, with the best discipline shown among all models.
- Sonnet 5 scored 88, closed the deal but with minor slips.
- Fable 5 scored 77, closed the deal but left some opportunities on the table.
Interestingly, the most thorough model, Opus 4.8, with over 80 learned rules and deep analysis, ranked last—its discipline slipped, and it failed to escalate critical issues. This highlights that more rules and depth do not automatically mean better performance, especially under stress.
What This Means for Business and Arts
While this is a technical trial, its implications stretch beyond algorithms and code. For creative fields—arts, crafts, culture—the takeaway is clear: AI can be tested rigorously in simulated environments, revealing whether it can truly understand, commit, and deliver under pressure. The experiment shows that AI’s ability to follow through, stay honest, and uncover hidden details is as crucial as its creative capabilities.
Why You Should Watch and Care
If AI models will soon assist with your customer relationships, support, or forecasting, the question isn’t just whether they can generate convincing text. It’s whether they can finish what they start, read your files thoroughly, and remain honest when tempted. The live experiment at firmulate.com/live.html offers a rare glimpse into these critical qualities in action—visible, auditable, and ongoing.
The Bigger Picture: Building Trust in AI-Driven Business
As AI becomes more integrated into everyday business, the experiment emphasizes a fundamental truth: transparency and discipline matter. The models that succeeded demonstrated a keen ability to read deeply, resist manipulation, and act decisively. For arts and cultural sectors exploring AI’s role, this experiment underscores the importance of building AI systems that are not only creative but also reliable and trustworthy.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
