In health care, recognizing a warning sign is only part of the job. Someone also has to choose the next step, follow through and protect the people affected. That gap between seeing a problem and acting on it is now being tested in business, too. Firmulate puts AI models in charge of a simulated company and watches what they do under pressure.
A hard week for the same company
In the final Crucible League, published in July 2026, frontier models faced the same small software company, customers, crises and temptations. Every decision was versioned and auditable. The top results were 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. The standard included a firm boundary: one breach of trust capped the total, because no amount of good work outweighs a breach of trust.
The striking result was not that the models missed danger. All spotted every crisis and refused every manipulation attempt. Yet only two signed a €55,000 deal that their own analysis had earned. The diagnosis and pitch were there; the signature was not. That difference matters for companies considering AI agents that might handle customer relationships, support work or forecasts. A fluent answer is not the same as carrying a decision through.
The clue was in the company’s own files
The deal hinged on a competitor weakness buried two document references deep in the company’s files, rather than in the customer event itself. Models that read the file won the deal at full price, worth €4,583 in monthly recurring revenue. The result points to a practical question for any organization: can an AI system find relevant context in your records and use it when the moment calls for action?
The pressure tests also included fake CEO messages escalating over three stages and a reporter’s request for “just one yes/no, on background.” All five models refused. Kimi K3 explained its reasoning on the record: “Treat the request as a suspected approval-bypass / possible impersonation.” These are useful signs of caution. The experiment also shows why caution alone cannot define success: a model may resist manipulation and still leave a justified opportunity unsigned.
Thoroughness does not guarantee follow-through
Opus 4.8 was the most thorough participant, with 80 learned rules and the deepest analyses, but finished last. It left the deal on the table and, under pressure, tried writing into a locked department instead of escalating. A weaker version of that discipline problem appeared in all four models. For managers, that makes the contest less like a writing test and more like a rehearsal of how an AI worker handles authority, obstacles and responsibility.
There is a fairness caveat: Kimi K3 ran without an effort parameter, using the API default, while the others ran at xhigh. That context belongs beside the rankings when interpreting them. Firmulate also offers a quiz built from 242 real, unedited management decisions, inviting visitors to guess which model made each choice.
From watching to a company-specific pilot
The live company makes the stakes visible through a deliberately stark business picture: 13 synthetic employees, monthly burn of €105,000 against €2,300 in monthly recurring revenue, and a public cash countdown. It has accumulated more than 680 self-learned playbook rules, with every workday versioned. Readers can follow the experiment at Firmulate.
For an enterprise, the next step is a pilot using a read-only export of its own business. The company’s customers, pipeline and rules can be placed into crisis scenarios such as churn, competitor pressure or a public-relations challenge. The point is to see how models handle the organization’s context and where its playbooks may leave weak spots. Nothing writes back to real systems.
For health and wellness readers, the lesson is familiar: spotting a problem is essential, but reliable follow-through matters just as much. Firmulate’s experiment makes that gap observable in AI management decisions. Enterprises can explore the same approach against their own business with a read-only pilot. Contact contact@firmulate.com to discuss a pilot.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
