Imagine if your favorite kitchen gadget could not only follow recipes but also handle crises, negotiate deals, and stay honest under pressure—all without a single employee. Now, consider the possibility of automating an entire company with AI that’s put to the test in real-world chaos. That’s exactly what the live experiment by Firmulate offers: a public, unfiltered glimpse into an AI-driven company fighting for survival, day after day.
The Live Company That’s Watching Its Own Failures
At the heart of this experiment is a small software firm built entirely around AI models, with 13 synthetic employees and a clear goal: prove whether AI can manage a business as effectively as humans. Every day, the company faces real crises—customer issues, negotiations, and even sabotage attempts—just like any business, but in this case, all decision-making is done by AI models that are openly tested and analyzed.
What makes this experiment extraordinary is its transparency. Every decision by these models is versioned and auditable, and the entire process is live at firmulate.com/live. You can see the models in action—how they spot crises, respond to manipulative tactics, and, crucially, how they perform under pressure.
Performance Benchmarks and Surprising Outcomes
To evaluate these models, researchers set up a benchmark league, pitting four of the most advanced AI models against each other. The models—named GPT-5.6-SOL, Kimi K3, Sonnet 5, and Fable 5—faced the same tough week, with identical customer issues, crises, and even attempts at social engineering.
The results? All four models identified every crisis and refused all manipulative tactics. Only two of the four managed to close a €55,000 deal—an actual sign that their recommendations and decisions were not just accurate but also disciplined enough to follow through. Interestingly, the models that read deeper into company files, rather than just superficial customer interactions, secured the full deal, adding over €4,500 to monthly recurring revenue.
The Hidden Weaknesses and Human-Like Mistakes
One notable detail is that all models shared a common weakness: they faltered when it came to executing certain decisions, often leaving promising deals unclosed or slipping into disciplinary breakdowns. For example, the most thorough model, Opus 4.8, with over 80 learned rules, left a close deal on the table because it failed to escalate a critical issue properly. These lapses mirror human frailties—missed steps, misjudged priorities—that challenge the idea of fully autonomous business management.
Social Engineering and Ethical Testing
The experiment also tested whether these models could be duped through social engineering—fake messages from a CEO or a reporter requesting quick approvals. Remarkably, all five models refused to bypass security protocols, citing suspicion or impersonation risks. Kimi K3 explicitly noted that such requests should be treated as potential fraud, highlighting that these systems are designed to recognize and resist manipulation.
AI in Property Management: A Practical, Unboring Look at Artificial Intelligence in the Multifamily Industry
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Reality of a Zero-Employee Business
This entire setup is happening inside a company that burns €105,000 each month against just €2,300 in recurring revenue. It’s a public, watchable experiment where every workday is versioned and analyzed, with the firm’s cash countdown ticking visibly. It’s not just a demo—it’s a raw, unfiltered test of what AI-driven management could look like.
For business leaders, the key takeaway isn’t how well these AI models chat or generate ideas—it’s whether they can finish what they start, stay honest under pressure, and adapt to crises without human intervention. The experiment shows that while models like GPT-5.6-SOL and Kimi K3 are capable of identifying issues and even closing deals, gaps remain—particularly in disciplined execution and escalation.
Why This Matters for Your Business
As AI technology matures, its role in managing customer relationships, support queues, or forecasting becomes more tangible. But the critical question for any enterprise is: will the AI stick to the plan? Will it read necessary documents, resist manipulation, and deliver real, useful work—rather than just sounding convincing?
In this experiment, the models demonstrate they can detect crises, refuse manipulative tactics, and even sign deals at full price. Yet, the lapses in discipline suggest that AI still needs careful oversight and testing before it can be entrusted with core business functions. This is where the concept of ‘wargaming’ your AI workforce—testing it against real crises before deploying—is invaluable. Firms can simulate their own business scenarios through tools like the live experiment at firmulate.com/pilot.html.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
AI for Public Relations: A How-To Guide for Implementation and Management
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The Confident Intern: What AI really is, and the seven-rung ladder for making it yours
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Mastering Cyber Detection Engineering: A Comprehensive Guide to Proactive Cybersecurity
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
