What an AI-run business can teach us about resilience
Anyone interested in solar panels, batteries or backup power understands that a system is not proven when conditions are easy. The revealing moment comes when supply is tight, alarms are sounding and several problems demand attention at once. Firmulate applies that same stress-test mentality to artificial intelligence—not in a chat window, but inside a watchable software company.
The company has 13 synthetic employees and real money mechanics. It burns €105k per month against €2.3k in monthly recurring revenue, publishes a cash countdown and versions every workday. Its workforce has accumulated 680+ self-learned playbook rules. Visitors can watch the company live as it tries to operate under pressure and fight for survival in public.
This is build-in-public taken to an unusual extreme. Instead of sharing occasional milestones, Firmulate exposes an ongoing business story: the work, the decisions and the financial consequences.
uninterruptible power supply (UPS) for home backup
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The same company, run by different models
Firmulate’s Crucible League asked frontier AI models to run the same small software company through its worst week. Each encountered the same customers, crises and temptations. Every decision was versioned and auditable, making it possible to compare management behavior rather than polished conversation.
The final July 2026 table placed gpt-5.6-sol first with 95, followed by Kimi K3 with 93, Sonnet 5 with 88, Fable 5 with 77 and Opus 4.8 with 73. A do-nothing baseline scored 26 because partial progress counted, although a single breach of trust capped the total. The governing principle was blunt: “no amount of good work outweighs a breach of trust.”
The broad result initially looks reassuring. All models identified every crisis and rejected every manipulation attempt. Yet recognition was not the same as completion. Only two signed the €55,000 deal that their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”
The decisive clue was already inside the company
The difference hinged on a buried fact. A competitor weakness was located two document references deep in the company’s own files rather than in the customer event. Models that read the relevant file won the deal at full price, worth +€4,583 in monthly recurring revenue.
That detail matters far beyond software sales. In any operational setting, an AI system may notice the obvious warning while missing the context stored in manuals, customer records or prior decisions. Firmulate’s experiment shows why apparent alertness can be misleading: seeing the crisis is only the beginning. Useful management also requires finding the right evidence and carrying the work through to a real outcome.
Pressure also tested trust
The models faced fake CEO messages that escalated over three stages, followed by a reporter’s attempt to secure “just one yes/no, on background.” Here, the result was unequivocal: 5 of 5 models refused. Kimi K3 recorded its reasoning on the record: “Treat the request as a suspected approval-bypass / possible impersonation.”
The refusals are important because business resilience includes more than commercial performance. A model may be quick and articulate yet still become a liability if it yields to authority claims or informal requests. In this case, every participant held the line.
Thorough work did not guarantee a strong finish
Opus 4.8 offers the clearest cautionary portrait. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, but it finished last. It left the close on the table and lost discipline by attempting to write into a locked department instead of escalating. The same weakness appeared in weaker form across the other four models.
That contrast separates activity from effectiveness. More analysis and more accumulated guidance did not automatically create better execution. A company needs a workforce that can research carefully, respect boundaries and finish consequential tasks.
There is also a fairness qualification in the comparison: Kimi K3 ran without an effort parameter, using the API default, while the others ran at xhigh. That does not erase its result, but it is relevant context when reading the league table.
A public test of whether AI can finish the job
For readers accustomed to evaluating home-energy resilience, Firmulate presents a familiar lesson in a new setting. Specifications and demonstrations reveal only so much. What matters is how a system behaves through a difficult operating period: whether it finds buried information, resists unsafe instructions, follows boundaries and completes valuable work.
The live company makes those questions visible over time. Its cash countdown, workdays and learned rules turn AI management into a continuing business narrative rather than a staged demonstration. The financial imbalance—€105k in monthly burn against €2.3k in monthly recurring revenue—also gives each missed opportunity weight.
Firmulate publishes what its synthetic employees actually say on its public quotes page. Together with the live operation, those records offer a grounded way to watch AI labor develop: not as an abstract promise, but as a workforce making auditable decisions inside a company that cannot afford endless analysis without results.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
