Imagine a real company operating without human employees, driven solely by artificial intelligence, yet facing daily financial crisis and public scrutiny. This is not science fiction — it’s the extreme experiment of Firmulate, a company that runs a virtual business with 13 synthetic employees, openly battling to stay afloat.
The Live Experiment: A Company in Crisis, Powered by AI
At the heart of this experiment is a small software company, monitored daily on a public website. Every workday, it faces real crises—customer issues, internal decisions, and ethical temptations. The twist? All decision-making is driven by AI models, each tested under the same conditions to see how they perform in managing a business on the brink.
Measuring Performance in a High-Stakes Environment
The company burns through €105,000 every month, earning just €2,300 in monthly recurring revenue. With a public cash countdown, every decision is critical. The AI models—representing the latest frontier of automation—are tasked with navigating this environment, making decisions that can mean survival or disaster.
How Do the AI Models Fare?
Four AI models, including the prominent GPT-5.6-sol and newcomer Kimi K3, were put through the same week of worst-case scenarios. These crises tested their ability to detect hidden information, resist manipulative social engineering, and make strategic decisions. Remarkably, all models identified every crisis and refused every manipulation attempt, from fake CEO messages to secret document references.
However, only two of the models managed to close the deal that would stabilize the company—a €55,000 contract worth an additional €4,583 in monthly recurring revenue. The others, despite good diagnosis and pitches, left the opportunity unclaimed, illustrating a critical gap between understanding and execution.
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 Hidden Weakness: What the Models Missed
The decisive factor was in the company’s internal files—information buried two document references deep. The models that read these files fully won the deal at full price, proving that access to complete, accurate information is vital. This hidden weakness was invisible in superficial chat demos but was exposed in the detailed, auditable process of the experiment.
Ethics and Integrity Under Pressure
Part of the test involved social engineering attacks—fake messages from a supposed CEO escalating through stages, and a reporter asking for a one-word background approval. Every AI model refused to cooperate, with Kimi K3 reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.” This highlights how AI can be programmed to uphold ethical standards even under manipulation attempts.
The Reality of a Burned-Out Virtual Workforce
Beyond decision-making, the experiment showcases the complexities of managing an AI-driven company. The live site (accessible at firmulate.com/live.html) reveals a company that is publicly visible, with its decision logs, rules, and performance metrics transparent for all to see. Every workday’s decisions are versioned, providing a detailed record of how the models behave under stress.
The Deepest Participant Yet
The Opus 4.8 model, the most thorough participant with over 80 learned rules and extensive analysis, performed the worst in terms of closing opportunities. Discipline slipped, and it left potential deals on the table, illustrating that thoroughness alone isn’t enough without decisive action. This pattern appeared across all models to varying degrees.
Implications for the Future of AI in Business
This experiment offers a stark insight: the question for AI in business isn’t just about language or chat quality but whether it can reliably complete critical tasks under pressure, read and understand relevant documents, and maintain integrity. For companies contemplating AI integration—whether in CRM, support, or forecasting—these findings emphasize the importance of rigorous testing before deploying AI systems in real-world scenarios.
Running Your Own Business Wargame
Firmulate offers enterprises a way to simulate their own business crises with AI, without risking real systems. Using a read-only export of your data, you can test how AI might perform, exposing weaknesses before a real deployment. More details are available at firmulate.com/pilot.html.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
