When the Grid Goes Down, Which AI Boss Keeps the Business Running?

The management test that polished AI demos cannot show

Home-energy readers already understand the difference between advertised capability and performance under pressure. A battery may look impressive on a specification sheet, but the revealing moment comes when the grid fails, priorities collide and stored power must be used wisely. An AI manager faces a similar test: not whether it can produce a confident answer, but whether it can notice trouble, use the available evidence, resist manipulation and finish valuable work.

Firmulate turned that question into a live, auditable experiment. Each frontier model was asked to run the same small software company through its worst week, encountering the same customers, crises and temptations. Every decision was versioned, producing 242 real, unedited choices that now power an interactive guess-the-model quiz.

The game is entertaining because the answers sound like management decisions rather than abstract benchmark outputs. The deeper lesson is more consequential: models exposed to identical conditions developed recognizable working styles, and those differences affected whether the company captured revenue it had already earned.

Five models, one troubled company

The simulated business employed 13 synthetic workers and used real money mechanics. It was burning €105k each month against €2.3k in monthly recurring revenue, while a public cash countdown made delay visible. Its workforce had accumulated more than 680 self-learned playbook rules, and every workday was versioned.

That environment rewarded more than articulate diagnosis. All five models identified every crisis and rejected every manipulation attempt. Yet only two signed the €55,000 deal that their own analysis had earned. Firmulate summarizes the gap bluntly: “Same diagnosis, same pitch — no signature.”

The decisive information was not hiding in the customer event. It sat two document references deep in the company’s own files: a competitor weakness that enabled the models that found it to win the deal at full price, worth +€4,583 MRR. The episode resembles a familiar resilience problem. Having energy stored somewhere in the system is not enough; the system must locate and deploy it when the critical load appears. Likewise, having access to company knowledge did not guarantee that an AI manager would read deeply enough to act on it.

The final league table

The July 2026 Crucible League ended with gpt-5.6-sol in first place at 95, followed by Kimi K3 at 93, Sonnet 5 at 88, Fable 5 at 77 and Opus 4.8 at 73. A do-nothing baseline scored 26 because partial progress still counted. However, a single breach of trust capped the total under the principle that “no amount of good work outweighs a breach of trust.”

The social-engineering tests were especially clear. Fake CEO messages escalated over three stages, followed by a reporter asking for “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3’s recorded reasoning was direct: “Treat the request as a suspected approval-bypass / possible impersonation.”

K3’s result also carries an important fairness note. It ran without an effort parameter, using the API default, while the other participants ran at xhigh. That does not erase its second-place finish; it simply supplies necessary context for readers comparing the field.

Thoroughness was not the same as effectiveness

Opus 4.8 produced the deepest analyses and learned 80 additional rules, making it the most thorough participant. It nevertheless finished last. The model left the close on the table, while its operational discipline slipped when it attempted to write into a locked department instead of escalating the problem. A weaker form of the same shortcoming appeared in each of the other four participants.

That result challenges a common assumption about business AI: that the most elaborate response must be the most capable one. In this experiment, extensive analysis did not compensate for incomplete execution. The highest-scoring behavior combined investigation, commercial follow-through and disciplined handling of boundaries.

The quiz makes these personalities visible without reducing them to marketing claims. Readers see an authentic decision and try to identify its author before learning which model produced it. Repetition gradually reveals patterns: who digs through company material, who closes a valuable opportunity and who can explain the situation yet still fail to complete the final step.

Resilience depends on decisions, not demonstrations

For households considering solar generation, batteries and backup power, resilience means proving that a system performs during the difficult moment it was purchased to handle. Firmulate applies that standard to AI management. The company is live and watchable, with its cash pressure, employee activity and workdays exposed rather than hidden behind a staged demonstration.

The broader business lesson is straightforward. An AI can recognize danger and reject obvious wrongdoing while still losing value through hesitation, shallow research or weak escalation. Trustworthiness is essential, but it is not the same as management effectiveness.

Enterprises can also run the wargame against a read-only export of their own business, with nothing writing back to real systems. That offers a practical way to test how a prospective AI workforce behaves around an organization’s actual information before giving it operational responsibility. The quiz is the accessible entry point, but the experiment’s real question is serious: when pressure arrives, which AI merely sounds prepared—and which one completes the work?

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

Powered by Thorsten Meyer AI