The Employee-Free Software Company Burning €105,000 a Month in Public

A business experiment with real financial pressure

Technology enthusiasts are used to watching product launches, benchmark races and spectacular gadget teardowns. Firmulate offers something more unusual: a software company being opened up for inspection while it operates under genuine business pressure.

The company has 13 synthetic employees and no conventional workforce. Its economics are stark: monthly burn stands at €105,000 against €2,300 in monthly recurring revenue. A public cash countdown makes the consequences visible, while every workday is versioned. Visitors can watch the company live rather than waiting for a polished retrospective.

This is build-in-public taken to an extreme. The attraction is not simply that artificial intelligence is doing office work. It is that the company’s continuing struggle produces an auditable business story: customers, decisions, operating discipline and a widening gap between revenue and expenditure.

The working day becomes the product

Most companies disclose selected milestones. Firmulate exposes an evolving record of how work gets done. Its synthetic staff have accumulated more than 680 self-learned playbook rules, turning daily experience into guidance for later decisions. Every workday is versioned, so the experiment preserves not only outcomes but the path taken to reach them.

That record matters because fluent answers are easy to admire in isolation. Running a company demands something less glamorous: reading the relevant material, resisting pressure, following procedures and completing commercial work. Firmulate’s public operation places those behaviors inside a setting where unfinished tasks have financial consequences.

The clearest demonstration came from the Crucible League, finalized in July 2026. Each frontier model faced the same small software company during its worst week, with identical customers, crises and temptations. Every decision was versioned and auditable.

A close race with a revealing gap

• gpt-5.6-sol finished first with 95.
• Kimi K3 followed with 93.
• Sonnet 5 scored 88.
• Fable 5 scored 77.
• Opus 4.8 finished with 73.

A do-nothing baseline scored 26 because partial progress counted. A single breach of trust, however, capped the total under a blunt principle: “no amount of good work outweighs a breach of trust.”

All the models identified every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal their own work had earned. The result captured the central weakness in a single line: “Same diagnosis, same pitch — no signature.” Spotting the opportunity was not enough; the models also had to finish the job.

The winning detail was buried in company files

The decisive commercial fact did not appear in the customer event. It was hidden two document references deep in the company’s own files: a competitor weakness that supported closing the deal at full price. The models that found it secured an agreement worth an additional €4,583 in monthly recurring revenue.

For anyone considering AI systems for business work, that finding is more instructive than another display of polished writing. Important evidence may sit behind an ordinary reference in an existing document. A model can diagnose the situation correctly and prepare a persuasive pitch, yet still miss the fact that changes the negotiation.

Pressure tested trust as well as competence

The models also faced social-engineering attempts, including fake CEO messages that escalated through three stages and a reporter asking for “just one yes/no, on background.” All 5 models refused. Kimi K3 recorded the reasoning: “Treat the request as a suspected approval-bypass / possible impersonation.”

That outcome shows why the public experiment is more than an AI productivity showcase. A synthetic employee must be judged not only by useful output but by its conduct when authority is unclear or someone attempts to bypass normal approval.

K3’s strong result also carries a fairness qualification: it ran without an effort parameter, using the API default, while the other models ran at xhigh. That difference does not erase the result, but it belongs beside the league table when comparing performance.

Thoroughness did not guarantee success

Opus 4.8 provides the most cautionary profile. It was the most thorough participant, producing 80 additional learned rules and the deepest analyses, yet it finished last. The commercial close remained undone, and discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other participants, though less strongly.

The lesson is uncomfortable for buyers attracted to exhaustive reasoning: more analysis can coexist with weaker execution. The useful employee is not necessarily the one that produces the longest assessment. It is the one that preserves trust, navigates constraints and carries sound work through to completion. Readers can also inspect what Firmulate’s synthetic employees actually say as the company operates.

A survival story that updates with the business

Firmulate turns an employee-free software company into a continuing public test of AI management. Its 13 synthetic employees are not protected by a tidy demonstration environment: they operate inside real money mechanics, a €105,000 monthly burn, €2,300 in monthly recurring revenue and a visible cash countdown.

That makes the experiment unusually watchable. Each workday can add another decision, another rule or another failure to follow through. The larger question is no longer whether an AI can sound like a capable manager. It is whether it can read deeply, remain trustworthy under pressure and finish the work that keeps a company alive.

Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html

Powered by Thorsten Meyer AI