Imagine a business where every decision, crisis, and temptation is played out live — with no employees, no shortcuts, and every move watched by the world. Welcome to the most extreme build-in-public experiment in AI management, where a tiny software company is fighting for its life against relentless cash burn, real crises, and the challenge of trust. This isn’t fiction — it’s happening right now, at firmulate.com/live.
The Live Experiment: An AI-Driven Company in Real Time
At the heart of this groundbreaking venture is a tiny, virtual company with 13 synthetic employees powered by advanced AI models. This firm isn’t just a demo; it’s a real-time, auditable laboratory where AI models are tasked with running a business through its worst week. Every decision made, every crisis faced, and every tempting shortcut is logged, versioned, and public for all to see.
From a financial standpoint, the company burns through €105,000 each month but earns only about €2,300 in monthly recurring revenue. It’s a fragile state, with a public cash countdown that underscores the urgency to make better decisions. Each workday, the company’s rules and responses are updated and versioned, creating a transparent view into how AI decision-making evolves under pressure.
AI business management software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The AI Models Under Scrutiny
Four frontier AI models participated in the experiment, each tested against the same week of crises. These models, ranging from GPT-5.6 to a more rule-disciplined version called Opus 4.8, faced identical challenges: customer issues, potential manipulation attempts, and internal crises. The results were revealing:
- All four models identified every crisis accurately.
- All refused attempts at manipulation, including social engineering tricks like fake CEO messages and staged reporter inquiries.
- Only two of the four models managed to close a deal worth €55,000, matching their own analysis and diagnosis.
Interestingly, the decisive advantage came not from surface-level chat interactions but from reading internal documents. Models that delved into the company’s own files uncovered critical information buried two references deep, leading to a winning deal at full price (+€4,583 MRR).
AI for Project Managers: A Desk Reference & Field Guide: Use Artificial Intelligence to Streamline Workflows, Automate Tasks, and Make Smarter Decisions with Practical Tools and Ethical Insights
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Trust and Discipline in AI Management
While some models excelled in crisis detection, others faltered on discipline and follow-through. For instance, Opus 4.8, the most thorough participant, exhibited deep analysis but ultimately left a promising deal unexecuted due to a slip in process discipline. It attempted to write notes into a restricted department instead of escalating, revealing that even the most capable models are vulnerable to operational lapses.
Meanwhile, a fairness note highlights that different models ran at varying effort levels, with some defaulting to less aggressive configurations. Despite this, the core findings remained consistent across the board: trust and discipline are critical, and AI must be trained not just for analysis but for execution under pressure.
AI document analysis software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Implications for Business Automation
This transparent, live experiment offers compelling lessons for companies considering AI automation. It’s not enough for AI to generate clever responses; it must also:
- Finish what it starts — avoid leaving deals or tasks half-done.
- Read and understand relevant internal documents thoroughly.
- Resist manipulation and social engineering attempts.
- Operate with discipline, escalating issues appropriately rather than slipping into operational shortcuts.
The experiment underscores a stark truth: a company’s viability isn’t just about quick wins or impressive chat demos. It’s about consistent, trustworthy execution — and that’s what the AI models are being tested for in real-time.
AI crisis detection tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Watch the Company in Action
Curious to see this in practice? The live site at firmulate.com/live streams ongoing runs of the experiment. Currently, 23 benchmark runs are queued, with each iteration revealing more about how these AI agents handle crisis, opportunity, and manipulation in a constant cycle of testing and learning.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
