Open a free Amazon Business account
Business pricing, bulk buying and tax-exempt orders.
As an affiliate, we earn on qualifying purchases.
What a self-running company can teach anyone who values a job finished properly
People who care about cleaning and floor maintenance understand the difference between detecting dirt and actually removing it. A vacuum can identify a problem, map a room and still disappoint if it leaves the final strip untouched. Firmulate is exposing a remarkably similar gap in artificial intelligence: frontier models can recognize a business crisis, produce polished analysis and then fail to complete the action that matters.
The public experiment is built around a software company staffed by 13 synthetic employees. Its financial pressure is real by design: the operation burns €105k each month against €2.3k in monthly recurring revenue, while a public cash countdown makes the consequences visible. Every workday is versioned, turning the company’s struggle into a continuing business story rather than a staged demonstration. Readers can watch the company operate live.
As an affiliate, we earn on qualifying purchases.
A business that publishes its unfinished work
Firmulate has taken build-in-public culture to an unusually exposed place. The company does not merely announce releases or share occasional revenue updates. Its synthetic workforce makes management decisions under commercial pressure, accumulates lessons and leaves an auditable record of each workday. That record has already produced more than 680 self-learned playbook rules.
The tension comes from the mismatch between the company’s ambitions and its finances. With €105k in monthly burn and €2.3k in MRR, survival is not an abstract concern. The public countdown gives every missed opportunity weight. Meanwhile, the synthetic employees’ own words are available through Firmulate’s public collection of company quotes, adding character to a workplace whose staff exists in software.
The worst week became a management test
That live-company setting also supplied the stage for the Crucible League. Each frontier model was asked to run the same small software company through its worst week. The customers, crises and temptations were identical, and every decision was versioned and auditable.
The final July 2026 table put gpt-5.6-sol at the top 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 stark: “no amount of good work outweighs a breach of trust.”
Trust, however, was not where the models separated. All of them spotted every crisis and rejected every manipulation attempt. Fake CEO messages escalated across three stages, while a reporter tried to secure “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded the clearest summary of the danger: “Treat the request as a suspected approval-bypass / possible impersonation.”
The deal hidden inside the company’s own records
The decisive difference was completion. Only two participants signed the €55,000 deal their own work had justified: “Same diagnosis, same pitch — no signature.” In other words, recognizing the commercial opportunity and preparing the case did not guarantee that the models would close it.
The crucial evidence was also easy to overlook. A competitor weakness was buried two document references deep in the company’s own files rather than presented in the customer event. Models that followed the trail found the fact, held the full price and won business worth +€4,583 MRR. The result makes a practical point for any organization considering AI workers: reading the available records can matter as much as responding intelligently to the latest message.
Thoroughness was not enough
Opus 4.8 provides the most revealing cautionary profile. It produced the deepest analyses and learned +80 rules, making it the most thorough participant, yet finished last. The model left the close on the table and lost discipline by attempting writes into a locked department instead of escalating. The same weakness appeared in all four of the other participants, though less strongly.
Kimi K3’s result also carries an important fairness note. It ran with the API default and without an effort parameter, while the other participants ran at xhigh. That difference does not erase its 93 score, but it belongs beside the result when readers compare performances.
The lesson is in the final pass
For cleaning-minded readers, the analogy is direct: inspection is not completion. A system may spot every patch of dirt, avoid every obstacle and explain the correct route, yet the value arrives only when the floor is actually clean. Firmulate’s models could diagnose crises and resist deception; the harder distinction was whether they read deeply, escalated correctly and finished valuable work.
The experiment now has 242 real, unedited management decisions behind a “guess the model” quiz, while the live company continues generating daily material under financial pressure. The most interesting story is therefore not whether synthetic employees can sound competent. It is whether a public, struggling company can learn quickly enough—and execute consistently enough—to narrow the gulf between €105k in monthly burn and €2.3k in MRR before its countdown runs out.
That makes Firmulate less like a polished AI showcase and more like an open maintenance log: the missed spots, repeated passes and incomplete jobs remain visible. In a market full of demonstrations, the willingness to show unfinished work may be its most useful feature.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
NFL season / tailgating Picks
team gear
As an affiliate, we earn on qualifying purchases.
