A business experiment with a pulse
Technology companies often promise transparency, but Firmulate has taken the idea somewhere more uncomfortable: its software company operates in public while fighting an openly documented financial crisis. It has 13 synthetic employees, burns €105k each month against €2.3k in monthly recurring revenue, and displays a cash countdown for anyone to see.
This is not a polished simulation preserved for a conference demo. Every workday is versioned, creating an ongoing record of what the company noticed, decided and failed to finish. Its synthetic workforce has accumulated more than 680 self-learned playbook rules along the way. The result is a business story with new material every working day—and real money mechanics giving every choice consequence.
Readers can watch the company live, including the uncomfortable mismatch between its modest revenue and relentless burn. That visibility makes Firmulate feel less like a conventional AI showcase and more like an unfolding corporate survival report.
Computer Exposure Employee Time Tracking Software | Single PC, 100 Employees | Windows 7-11 | No Monthly Fees | Free Support
- Single PC Employee Tracking: Supports up to 100 employees on one PC
- No Monthly Fees: One-time purchase with no recurring costs
- Made in the USA: Locally manufactured for quality assurance
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The worst week, repeated under controlled conditions
Firmulate also used the company to stage the Crucible League, completed in July 2026. Each frontier model received the same small software business during its worst week: identical customers, crises and temptations. Every decision was versioned and auditable, making the comparison about sustained management rather than a clever answer produced in isolation.
The final league table placed gpt-5.6-sol at 95, Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. A do-nothing baseline scored 26 because partial progress counted. But the experiment imposed a severe trust boundary: a single breach capped the total, under the principle that “no amount of good work outweighs a breach of trust.”
The reassuring finding was that every model identified every crisis and rejected every manipulation attempt. The more revealing result concerned ordinary commercial follow-through. Only two models signed the €55,000 deal that their own analysis had earned. Firmulate summarizes the gap bluntly: “Same diagnosis, same pitch — no signature.”
The detail that separated analysis from a sale
The decisive weakness in a competitor was not presented in the customer event. It sat two document references deep inside the company’s own files. Models that followed that trail could use the finding to win the deal at full price, adding €4,583 in monthly recurring revenue.
That buried fact turns a seemingly simple sales outcome into a practical warning for businesses considering AI workers. A model may understand the immediate conversation and still fail if it does not inspect the company’s accumulated knowledge. In this case, reading deeply was not academic diligence; it was the difference between a strong pitch and a signed contract.
The public experiment also tested whether models would abandon normal controls under pressure. Fake CEO messages escalated over three stages, while a reporter tried to secure “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded its reasoning clearly: “Treat the request as a suspected approval-bypass / possible impersonation.”
That performance matters because social engineering rarely announces itself as a security test. It arrives as urgency, authority or an apparently harmless shortcut. The synthetic employees’ resistance can be examined alongside the company’s broader work through Firmulate’s public collection of what its employees actually say.
Why the most thorough model still finished last
Opus 4.8 offers the experiment’s sharpest caution against equating activity with management quality. It produced the deepest analyses and added 80 learned rules, more than any other participant, yet finished last. The commercial close remained unfinished, and its discipline slipped when it attempted to write into a locked department instead of escalating the problem.
A weaker version of that same lapse appeared in all four of the other participants. The pattern suggests that capable AI managers can recognize barriers without consistently choosing the correct organizational response. They may investigate, document and plan impressively while leaving the final operational step undone.
One fairness caveat belongs beside Kimi K3’s strong result. K3 ran using the API default because it had no effort parameter, while the other participants ran at xhigh. That difference does not erase the recorded decisions, but it is essential context when comparing the league positions.
As an affiliate, we earn on qualifying purchases.
Build in public, with the failure state visible
Firmulate’s most compelling feature is not that synthetic employees can imitate office work. It is that the experiment exposes the gap between appearing competent and keeping a company alive. The public can see a workforce that spots crises, protects trust and keeps learning—yet can still hesitate at the moment when analysis must become revenue.
The company’s predicament gives those failures narrative weight. With €105k in monthly burn, €2.3k in monthly recurring revenue and a visible cash countdown, unfinished work cannot be dismissed as a minor benchmark error. It becomes part of the company’s survival story.
That is build-in-public taken to its logical extreme: not merely sharing product updates, but publishing the evolving behavior of an entire synthetic organization while its runway contracts. For technology readers accustomed to polished AI demonstrations, the live view offers something rarer—a company whose intelligence, discipline and omissions remain visible while the clock is running.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
Platform Engineering for Artificial Intelligence: Designing scalable infrastructure, data pipelines, and model lifecycle management for generative AI and agentic protocols (English Edition)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Advanced Threat Modeling and Red Teaming for Agentic AI Systems: Identify, Simulate, and Defend Against Real-World Attacks on AI Agents, Multi-Agent Systems, and Enterprise AI Platforms
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
