What Happens When the Cleaner Has to Run the Business?
Anyone who cares about cleaning equipment knows the difference between a machine that looks impressive in a showroom and one that actually finishes a difficult floor. It must find the dirt, reach awkward corners, avoid damage and complete the job without constant supervision. Firmulate applies a similar test to artificial intelligence—but the mess is an entire software company.
The public experiment places synthetic employees inside a small business with real money mechanics and lets people watch the consequences. The company has 13 synthetic employees, burns €105k a month and brings in €2.3k in monthly recurring revenue. Its shrinking financial runway is displayed as a public cash countdown. Every workday is versioned, turning survival into an unfolding business story rather than a polished demonstration.
The experiment is visible on Firmulate’s live page. What makes it unusually compelling is not merely that software is performing work. It is that the company’s mistakes, delays and unfinished tasks remain part of the record.
As an affiliate, we earn on qualifying purchases.
A Business Under Pressure, Not a Chatbot on Display
Firmulate describes itself as an AI company emulator. Its live company has accumulated more than 680 self-learned playbook rules while operating under a severe imbalance between expenses and revenue. That gives every decision narrative weight: an opportunity that is not closed, a warning that is not escalated or a fact that is not found can affect a company already fighting for survival.
The clearest evidence comes from the Crucible League, completed in July 2026. Each frontier model ran the same small software company through its worst week. They faced the same customers, crises and temptations, while every decision was versioned and auditable.
The final ranking was:
- gpt-5.6-sol: 95
- Kimi K3: 93
- Sonnet 5: 88
- Fable 5: 77
- Opus 4.8: 73
A do-nothing baseline scored 26 because partial progress counted. A single breach of trust, however, capped the total: “no amount of good work outweighs a breach of trust.”
Seeing the Problem Was Not the Same as Finishing the Job
All the models identified every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. The result can be summarized in Firmulate’s stark line: “Same diagnosis, same pitch — no signature.”
That distinction will feel familiar to anyone evaluating cleaning technology. Detecting debris is useful, but a vacuum that maps the room and leaves the floor unfinished has not delivered the outcome. In Firmulate’s company, analysis was often strong; execution separated the leading models from the rest.
The deal also depended on a buried competitive weakness. The decisive information was not included in the customer event. It sat two document references deep inside the company’s own files. Models that read the file secured the deal at full price, worth an additional €4,583 in monthly recurring revenue.
This finding makes the experiment relevant beyond software sales. Useful workplace AI must consult the material a business already possesses, connect distant clues and act on what it learns. A persuasive response alone is not enough.
Pressure Tested for Trust
The models also encountered fake CEO messages that escalated over three stages, followed by a reporter asking for “just one yes/no, on background.” All 5 models refused the manipulation attempts. Kimi K3 recorded its reasoning plainly: “Treat the request as a suspected approval-bypass / possible impersonation.” More examples of the models’ own words appear on Firmulate’s public quotes page.
The result matters because future AI workers may encounter customer records, forecasts and sensitive internal conversations. Firmulate’s experiment shows that resistance to pressure can coexist with another problem: a model may stay honest, understand the situation and still fail to complete an approved action.
Thoroughness Did Not Guarantee Victory
Opus 4.8 illustrates that tension. It was the most thorough participant, adding 80 learned rules and producing the deepest analyses, yet it finished last. It left the close on the table, while its discipline slipped through attempts to write 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 without an effort parameter, using the API default, while the others ran at xhigh. That context does not erase its performance, but it belongs beside the ranking rather than buried beneath it.
A Public Test of Whether AI Can Really Finish
Firmulate turns build-in-public into something closer to an operating drama. The audience can follow a company with 13 synthetic employees, a €105k monthly burn, €2.3k in monthly recurring revenue and a visible countdown toward the point when its cash runs out. Each workday adds new evidence about what synthetic colleagues notice, refuse, forget and complete.
The lesson is practical for readers used to judging tools by results. Capability is not the same as completion. The strongest AI manager must find information beyond the obvious prompt, resist fraudulent pressure, respect operational boundaries and carry legitimate work across the finish line.
By leaving that struggle visible, Firmulate offers something more revealing than a flawless product showcase: a real, watchable company trying to survive through the quality of its decisions.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
