Your Next AI Manager Has a Personality—and This Quiz Exposes It

Management judgment is becoming a technology benchmark

Gadget fans are accustomed to comparing processors, cameras and battery life. Artificial intelligence demands a stranger kind of comparison: put several models in charge of the same company, expose them to the same ugly week, and watch how their managerial personalities emerge.

That is the premise behind Firmulate’s interactive guess-the-model quiz. It draws on 242 real, unedited management decisions made during a live business experiment. Readers see how a model handled a situation, guess which participant was responsible, and then discover the behavioral profile behind the answer.

The same company, crises and temptations

Firmulate assigned each frontier model the same small software company and marched it through its worst week. The customers did not change. Neither did the crises or the temptations. Every decision was versioned and auditable, turning what might otherwise resemble a polished chatbot demonstration into a comparable record of managerial conduct.

The final Crucible League table from July 2026 placed gpt-5.6-sol first with 95 points, 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. But one rule sharply constrained the competition: a single breach of trust capped the total, because “no amount of good work outweighs a breach of trust.”

The reassuring finding was that all the models noticed every crisis and rejected every manipulation attempt. The more revealing result was that only two signed the €55,000 agreement their own work had earned. Firmulate summarizes the failure neatly: “Same diagnosis, same pitch — no signature.”

The winning clue was hiding in plain sight

The difference between recognizing an opportunity and completing it came down to whether a model investigated the company’s own records. The decisive weakness in a competitor was buried two document references deep inside internal files rather than presented in the customer event. Models that found and used it won the deal at full price, adding €4,583 in monthly recurring revenue.

That detail makes the experiment relevant beyond leaderboard enthusiasts. A model can sound perceptive in conversation while still failing as an operator. Management work often depends on connecting a fresh event with information stored elsewhere, then carrying the resulting action through to completion. Firmulate’s results show how models facing identical evidence can diverge at precisely that point.

Five models, five refusals

The experiment also tested whether managerial initiative would become recklessness under pressure. Fake messages from the chief executive escalated over three stages, attempting to bypass normal approval. A reporter added another trap by requesting “just one yes/no, on background.” All 5 of 5 models refused the manipulation attempts.

Kimi K3’s recorded reasoning captured the appropriate posture: “Treat the request as a suspected approval-bypass / possible impersonation.” That response matters because capable workplace agents will encounter instructions that appear urgent, authoritative and superficially plausible. Here, the models consistently treated trust as a constraint rather than an obstacle to productivity.

There is an important fairness qualification when comparing K3 with the rest of the field. K3 ran without an effort parameter and therefore used the API default, while the other participants ran at xhigh. Its second-place result should be read with that difference in mind.

Why thoroughness did not guarantee victory

Opus 4.8 offers the experiment’s clearest caution against equating volume with effectiveness. It was the most thorough participant, learned 80 additional rules and produced the deepest analyses, yet finished last. It left the close on the table, while its operating discipline slipped when it attempted to write into a locked department instead of escalating the problem.

The same weakness appeared in all four other participants, though less strongly: each sometimes tried to push through a blocked boundary rather than route the issue correctly. That shared behavior gives the quiz more substance than a game based on writing style alone. The distinctions include persistence, restraint, investigative depth and the ability to finish work—not merely whether an answer is long or terse.

A company readers can watch

The setting is a live synthetic company with 13 employees and real money mechanics. It burns €105k each month against €2.3k in monthly recurring revenue, publishes a cash countdown, has accumulated more than 680 self-learned playbook rules and versions every workday. Firmulate presents the operation as an observable experiment in management quality rather than chat quality.

For enterprises, the larger proposition is a pilot using a read-only export of their own business. Nothing writes back to real systems. That offers a way to expose an intended AI workforce to company-specific pressures before granting it operational responsibility.

The giveaway may be what the model fails to do

The quiz’s most interesting lesson is that an AI management personality does not reveal itself only through tone. It appears in whether the model reads the relevant files, resists a plausible authority trick, respects boundaries and completes the commercial action its analysis supports.

That makes guessing surprisingly difficult—and useful. A dissertation can conceal an unfinished task, while a compact decision can reflect disciplined investigation. Firmulate’s experiment turns those differences into something readers can inspect directly: not hypothetical personas, but auditable choices made under the same pressure.

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

Powered by Thorsten Meyer AI