Imagine watching an AI manage a real company in real time, making decisions, facing crises, and even losing money — all live on a public website. This is not science fiction. It’s the experiment at Firmulate, where a virtual company with 13 synthetic employees navigates its worst week, revealing how AI handles trust, crises, and ethical dilemmas in a high-stakes simulation.
The Live Company Experiment
At the heart of this experiment is a small, simulated software company that runs every workday on public display. It’s a real-time showcase of AI decision-making, where each move is versioned and auditable. The company burns €105,000 per month, earning just €2,300 in monthly recurring revenue, with a public countdown showing how long it can keep going. This setup isn’t just for show — it tests how AI models perform under pressure, with a focus on integrity, problem-solving, and discipline.
As an affiliate, we earn on qualifying purchases.
How the Models Perform
Four advanced AI models from the frontier of language and decision systems were tested against the same set of crises and temptations. Their goal? Run the company through its worst week, confronting the same customer issues, internal dilemmas, and manipulative tactics. Every decision the models made was recorded, versioned, and available for review at firmulate.com/live.html.
The results were revealing. All four models identified every crisis and refused every manipulation attempt, including sophisticated social engineering tactics such as fake CEO messages and reporter tricks. Interestingly, the key to winning a major deal was hidden deep within the company’s own files – a buried fact that, when read, allowed two models to close a €55,000 deal at full price, increasing monthly recurring revenue by €4,583.
Lessons on Trust and Ethics
This experiment underscores an essential truth: AI’s ability to maintain integrity under pressure matters more than its conversational flair. In the test, models displayed a strict refusal to bypass security or impersonation requests, with Kimi K3 citing “suspected approval-bypass / possible impersonation” as reason for declining. These behaviors are critical for AI systems that will someday manage customer support, financial decisions, or compliance tasks.
What It Means for Business
While the company’s financials are bleak — burning €105k monthly against a tiny €2.3k revenue — the process itself is illuminating. Watching the company’s decisions unfold reveals whether AI can be trustworthy, disciplined, and capable of completing complex, multi-step tasks. For businesses worried about AI implementations touching their CRM, support queues, or forecasts, this live experiment offers a glimpse into what’s possible: AI that not only understands your data but also acts ethically and decisively.
Deep Analysis and Future Potential
The most thorough participant, OPUS 4.8, with over 80 learned rules and deep analysis, finished last — leaving close deals on the table and slipping discipline. This highlights that even advanced models can falter if their focus drifts. The experiment is ongoing, with scores updated and new runs published twice daily. It’s a window into an extreme, build-in-public test of AI’s readiness to run real companies and handle real crises.
Why It Matters to You
In a world where AI may soon touch every aspect of your business or pet-related services, understanding its true capabilities is vital. It’s not about how well it chats — it’s whether it can finish what it starts, read critical info in files, stay honest under pressure, and deliver real value without shortcuts. This experiment from Firmulate demonstrates that AI can be tested in the wild, with real consequences, before it’s trusted with your day-to-day operations.
This live AI company experiment shows that in critical situations, trustworthiness and discipline matter more than just impressive conversations. Watching AI face crises and ethical tests live offers invaluable insights for any business considering automation or AI decision-making.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
