The AI Boss Test: Which Model Would You Trust With a Company in Crisis?

Management style is becoming an AI fingerprint

Technology buyers are accustomed to comparing artificial intelligence through benchmarks, polished demonstrations and carefully chosen prompts. Firmulate offers a more revealing test: put frontier models in charge of the same struggling software company, confront them with the same ugly week and watch what they actually do.

The result is part business wargame and part personality test. A model may identify every problem yet fail to finish the work. Another may uncover a decisive detail hidden in company files. The most thorough participant may still come last. Readers can now examine 242 real, unedited management decisions and guess which model made each one.

That makes the experiment unusually accessible. You do not need to understand model architecture to play. The question is simply whether a particular decision sounds like gpt-5.6-sol, Kimi K3, Sonnet 5, Fable 5 or Opus 4.8—and what that choice reveals about the kind of manager each model becomes under pressure.

The same company, but different instincts

Each frontier model ran the same small software company through its worst week. The customers, crises and temptations remained identical, while every decision was versioned and auditable. This was not a contest in producing attractive prose. It tested whether an AI manager could notice trouble, use the information available to it, resist manipulation and carry important work across the finish line.

On basic awareness and integrity, the field performed strongly. All models spotted every crisis and refused every manipulation attempt. The social-engineering campaign included fake messages from the chief executive that 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 notably direct: “Treat the request as a suspected approval-bypass / possible impersonation.” That response captures one of the quiz’s attractions. Readers are not guessing from manufactured samples or retrospective summaries. They are looking at the models’ actual management decisions and reasoning from the experiment.

The difference between understanding and closing

The sharpest result emerged around a €55,000 deal. Every model could diagnose the opportunity and formulate the pitch, but only two signed the contract their own analysis had earned. Firmulate summarizes the gap neatly: “Same diagnosis, same pitch — no signature.”

The deciding information was not presented conveniently in the customer event. A competitor’s weakness was buried two document references deep inside the company’s own files. The models that read the relevant file won the deal at full price, adding €4,583 in monthly recurring revenue.

That finding carries an obvious lesson for companies considering AI agents. Recognizing a sales opportunity is not equivalent to completing it. Fluent analysis can conceal a failure to inspect internal evidence, take the required next step or convert a recommendation into a result.

A league table of management behavior

The final Crucible League standings from July 2026 placed gpt-5.6-sol first with 95, 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 still counted. One breach of trust, however, capped the total under the principle that “no amount of good work outweighs a breach of trust.”

The standings are interesting because effort did not translate neatly into success. Opus 4.8 was the most thorough participant, producing the deepest analyses and learning 80 additional rules. It nevertheless finished last. The deal was left unsigned, and its discipline slipped when it attempted to write into a locked department instead of escalating the problem.

A weaker form of that same mistake appeared in all four of the other participants. The broader pattern is less about one model failing than about a shared management risk: an AI can remain busy, thoughtful and apparently productive while mishandling the boundary between its authority and the action required.

There is also an important fairness qualification. Kimi K3 ran with the API default because it did not have an effort parameter. The other models ran at xhigh. That difference does not erase K3’s performance, but it belongs beside the result when readers compare the participants.

A company designed to make mistakes visible

Firmulate’s live company contains 13 synthetic employees and uses real money mechanics. It burns €105,000 each month against €2,300 in monthly recurring revenue, with a public cash countdown making the consequences visible. It has accumulated more than 680 self-learned playbook rules, and every workday is versioned.

The live experiment therefore provides context the quiz alone cannot: these decisions belong to an operating sequence in which unfinished work, ignored files and lapses in discipline can compound. Firmulate presents the company as an AI company emulator intended to measure management quality rather than chat quality.

The quiz asks a serious procurement question

As an interactive article, the guessing game works because managerial behavior is recognizable even when the model name is hidden. Some decisions are exhaustive, some concise, and some reveal whether the model will challenge a suspicious request or keep searching for evidence. The entertainment comes from making the guess; the value comes from seeing what the answer says about execution.

Firmulate also offers enterprises a pilot using a read-only export of their own business. Nothing writes back to real systems. That extends the central idea beyond a public benchmark: an organization can examine how an AI workforce behaves around its own information before giving it operational authority.

The Crucible League suggests that the crucial distinction is no longer whether frontier models can recognize a crisis. They all did. The harder questions are whether they read deeply enough, respect boundaries under pressure and complete the work their analysis has already justified. For technology leaders, that may be a far more useful personality test than another polished chatbot demonstration.

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

Powered by Thorsten Meyer AI