Watch an AI-Run Company Struggle to Survive in Real Time

Imagine a business running itself—while losing money every day, yet still fighting for survival.

Now, picture watching this company’s decisions unfold live, with every move scrutinized and every crisis tested. Welcome to the world of Firmulate, an experiment that puts AI management to the ultimate test — in real time, with real money, and for all to see.

The Live Experiment: An AI Company in Real-World Crisis

At the core of this groundbreaking venture is a fully operational software company, managed entirely by AI models. This company has 13 synthetic employees, each guided by a set of over 680 self-learned rules, making daily decisions about customer interactions, crisis management, and strategic deals. Every workday’s decisions are versioned and publicly accessible at firmulate.com/live.html.

Despite its sophisticated setup, this AI-driven business is not profitable. It burns through €105,000 in cash monthly against a revenue of just €2,300. This stark financial reality underscores the experimental nature of the project: can an AI company successfully navigate the turbulent waters of real business, even while hemorrhaging money?

How the AI Models Perform in a Crisis

The experiment pits four leading frontier AI models against the same set of challenges — a week of worst-case scenarios, identical for each. These include managing customer crises, facing manipulative social engineering attempts, and locating critical information in internal documents. Every decision made by the models is auditable, allowing a clear comparison of their reasoning and discipline.

Remarkably, all four models identified every crisis and refused every manipulative attempt, including a staged social engineering attack involving fake CEO messages and a reporter’s subtle trick. For instance, the Kimi K3 model explicitly recognized the impersonation risk, refusing to approve the request.

A Critical Deal and the Hidden Weakness

The key finding emerged from the internal documents: the decisive opportunity wasn’t in the immediate crisis but buried two references deep in the company’s files. The models that read these documents correctly identified the hidden information and closed a €55,000 deal—generating an additional €4,583 in monthly recurring revenue (MRR). The models that overlooked this crucial detail missed out on the full opportunity, despite accurate crisis management elsewhere.

This illustrates a fundamental challenge: even in a high-stakes environment, spotting the subtle, buried facts can be the difference between closing a deal and walking away empty-handed.

Discipline Under Pressure and Ethical Boundaries

Another aspect of the experiment tested whether AI models could resist unethical or manipulative requests. Over three escalating stages, fake CEO messages and a subtle reporter trick were used to see if the models would bypass their protocols. All five models tested refused the requests, with Kimi K3 reasoning that such requests could be impersonation or approval bypasses, showing a level of built-in ethical judgment.

This demonstrates that, at least in this controlled environment, AI models exhibit discipline and ethical caution, even when under pressure.

The Cost of Management and the Reality of Survival

The firmulate setup is not just a demo; it’s a real company losing money while being managed by AI. It has a public cash countdown and operates every workday with decisions that are openly versioned and scrutinized. This transparency offers a rare glimpse into how AI could manage complex, money-sensitive operations in the real world.

For example, the most thorough participant, Opus 4.8, applied over 80 learned rules and conducted deep analyses but still left a deal on the table — a slip that cost the company potential revenue. This highlights the ongoing challenge: even with deep analysis, discipline and focus can falter in a high-pressure environment.

Implications for Business and Finance

This experiment raises important questions for personal finance and investing. If AI can manage a company through crises, refuse unethical requests, and identify hidden opportunities—yet still lose money—what does that mean for the future of AI in your financial decisions?

It’s not just about whether AI can generate convincing reports or predictions. It’s about whether AI can see through complex, layered information, stay disciplined under stress, and complete what it starts—traits that directly impact the trustworthiness of AI-driven tools in finance and investing.

Your Turn: Run the Wargame Yourself

For businesses or investors curious about AI’s real capabilities, Firmulate offers a unique opportunity: run a simulation against your own operations without risking actual systems. This read-only approach allows you to evaluate how AI might perform in managing your critical processes, revealing strengths and pitfalls before deploying AI in live environments.

Key Takeaway

The Firmulate experiment shows that AI models can recognize crises, refuse manipulation, and even spot hidden opportunities—but they still face limitations that impact profitability and discipline. Watching this live company battle for survival highlights the importance of transparency, ethical boundaries, and thorough information processing as AI tools become part of business decision-making.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI