Imagine a company with no employees, losing €105,000 every month, yet somehow still competing for business in a high-stakes game. This isn’t fiction — it’s the reality of a live experiment where artificial intelligence models run a simulated business, facing real crises, tough decisions, and the pressure to deliver. Welcome to the most extreme build-in-public story in AI and business today.
The Live Company That’s Watching Itself Fail
At the heart of this experiment is a small, virtual company powered by AI models, with 13 synthetic employees making decisions every workday. It burns through €105,000 each month against a modest €2,300 in monthly recurring revenue, all the while under a relentless public timer counting down its cash reserves. This setup, hosted at firmulate.com/live, isn’t just a tech demo — it’s a real-time, visible battle for survival, with every decision and crisis scrutinized and versioned.
AI Builders: Making The Decisions That Turn AI Code Into Real Software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Test: Same Crises, Different AI Models
Four frontier AI models, including the latest GPT-5.6-sol, faced the same week of simulated business challenges. These crises ranged from customer issues to internal manipulations, testing the models’ ability to detect problems and make honest decisions. The models had access to the company’s documents — not just the front-facing data, but deep internal files too. As it turns out, the key to winning the deal wasn’t just identifying crises — it was reading and understanding internal documents buried two references deep.
Trustworthy Medical AI: A Builder's Guide to Safe, Compliant Software as a Medical Device
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Results Show Both Promise and Pitfalls
- All four models identified every crisis, demonstrating strong situational awareness.
- All rejected manipulation attempts, including social engineering tricks like fake CEO messages and reporter tricks.
- Only two models, GPT-5.6-sol and Kimi K3, managed to close the €55,000 deal, earning full monthly recurring revenue of +€4,583.
- The other models either left the deal on the table or failed to follow through, despite their accurate diagnoses.
AI: THE PERPETUAL INTERN – Its Brilliance and Failures Share the Same Root
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Understanding the Hidden Weakness
The most critical weakness wasn’t in the models’ crisis detection but in their internal process discipline. Deep inside the company’s files, a single document reference contained a crucial piece of information that could clinch the deal. The models that read this hidden data and incorporated it into their decision-making won the contract at full price. Those that missed it or failed to escalate properly lost the opportunity.
Silhouette America Studio Business Edition Software, Multicolor
Silhouette Studio Business Edition Is A Version Of The Silhouette Studio Software Program That Intended For Small Business…
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Social Engineering and Trust
In a simulated social engineering attack, fake CEO messages were escalated through multiple stages, attempting to bypass approval. All five models tested refused to sign off on these requests, with Kimi K3 explicitly reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This underscores the importance of honesty and caution in AI decision-making, especially when facing pressure to manipulate.
The Real-World Implications
This experiment isn’t just an academic exercise. It’s a window into how AI can be trusted — or betrayed — in real business contexts. A company that employs AI for customer support, CRM, forecasting, or decision-making needs assurance it can finish what it starts, read the right files, and stay honest under pressure. These are the qualities that separate a helpful AI from a risky one.
The Living Experiment Continues
Every weekday, the live experiment at firmulate.com/live updates with new decision versions, results, and insights. Watch how these models navigate crises, close deals, or slip up under the weight of internal discipline. The scores are transparent: GPT-5.6-sol leads with a 95 out of 100, having found the buried fact and sealed the deal, while other models trail behind.
This public experiment exposes the critical strengths and weaknesses of AI in business decision-making. Trustworthiness, thoroughness, and internal discipline matter more than just surface-level intelligence. As AI models become more embedded in your work, understanding what they see — and whether they can finish what they start — is more vital than ever.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
