Imagine managing a cleaning or floor maintenance company where every decision could be your last — and you can watch that struggle unfold live online. Today, AI technology is pushing into the world of business operations in a way that’s both transparent and startling. This isn’t a simulation; it’s a real company, run entirely by artificial intelligence models, facing the harsh realities of a daily cash burn, crises, and tough decisions. For anyone in maintenance or cleaning industries, or simply curious about AI’s limits, this is a front-row seat to the future of work.
Open a free Amazon Business account
Business pricing, bulk buying and tax-exempt orders.
As an affiliate, we earn on qualifying purchases.
The Live Experiment: An AI Company in Real Time
At the heart of this story is a publicly accessible live experiment conducted by Firmulate, a company that runs AI models as if they were full-fledged businesses. The company’s setup is straightforward but intense: 13 synthetic employees, running a real money mechanic that burns €105,000 per month against a modest €2,300 in monthly recurring revenue (MRR). Every workday, the AI models face the same crises, customer requests, and temptations a real company would encounter, with decisions carefully tracked, versioned, and published for all to see.
This experiment is designed to test the very core of AI management: can these models operate with honesty, discipline, and strategic acumen under pressure? The answer, as it turns out, is both promising and sobering.
Superior Crisis Detection and Integrity
All four frontier AI models tested through the same worst-week scenario successfully identified every crisis — from customer complaints to operational emergencies. They also refused every manipulation attempt, including social engineering tactics like fake CEO messages and reporter tricks, which five out of five models declined. Kimi K3, one of the models, explicitly treated suspicious requests as possible impersonations, highlighting an awareness of trust issues that’s crucial for real-world management.
The Hidden Weakness — Reading the Files
Despite their strengths, the models’ weaknesses became apparent in their decision-making processes. The decisive advantage was found not in the immediate customer interactions but two document references deep within the company’s own files. The models that read these files uncovered critical information that led to sealing a €55,000 deal — worth over €4,583 in additional monthly revenue. These models closed the deal at full price, demonstrating that reading and understanding internal data is vital for effective AI management.
Every Decision Counted — Versioned and Auditable
The experiment also involved rigorous versioning. Every decision made by the models was recorded, allowing observers to trace how strategies evolved day by day. Interestingly, the most thorough participant, Opus 4.8, with over 80 learned rules and deep analyses, finished last in the deal-making, due to slips in discipline and decision execution. This highlights that knowledge alone doesn’t guarantee success; disciplined application and process adherence are equally vital.
Real Money, Real Losses
The live company’s financials tell a stark story: burning €105,000 each month against just €2,300 in income. A public countdown shows how long the company can sustain this burn rate before running out of cash — a real-time visualization of an AI-managed business fighting for survival. Watching the company’s daily decisions, and reading its public logs, offers a unique window into the operational challenges and potential of AI management tools.
AI for Modern Finance: The Complete Guide to AI-Assisted Research, Stock Analysis, Investing, Trading, Risk Management & Financial Decision Making
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Implications for Industries like Cleaning and Maintenance
While the experiment involves a software company, the lessons are strikingly relevant for fields like cleaning, floor care, and maintenance. AI models that can identify crises, refuse manipulative tactics, and read internal documents could revolutionize how these industries manage operations, customer relationships, and compliance. But the key takeaway is that AI’s true value lies not in generating pretty chat responses but in completing critical work with honesty and consistency — qualities that are often tested in the hardest moments.
Why It Matters
As AI begins to touch more aspects of business, the question isn’t whether it can produce convincing conversations. It’s whether it can follow through, stay honest, and deliver useful results under pressure. This live experiment underscores that distinction clearly. The models that succeeded not only diagnosed problems but also closed deals, all while maintaining integrity in the face of social engineering and internal complexities.
Looking Ahead
For companies eager to evaluate AI before integrating it into their operations, Firmulate offers a unique tool: a read-only wargame that lets you run your own business scenarios against AI models, without risking real systems. Visit firmulate.com/pilot.html to see how your enterprise might stand up to this kind of rigorous testing.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Fall Picks
fall essentials
As an affiliate, we earn on qualifying purchases.
