What AI management looks like on the workbench
Anyone who works with wood knows that recognizing a good tool is not the same as seeing a good result. A saw may sound powerful and a chisel may look immaculate, but the real test comes when the cut must follow the line. Artificial intelligence has a similar problem: polished language can disguise whether a model actually completes the job.
Firmulate turns that distinction into a public experiment—and an unusually revealing guess-the-model quiz. It contains 242 real, unedited management decisions. Readers see what an AI manager did in a business situation, then try to identify which frontier model was responsible. The attraction is partly playful, but the underlying question is serious: do different models have recognizable management personalities?
AI management decision analysis tools
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
The same company, crises and temptations
Firmulate gave each frontier model the same assignment: run a small software company through its worst week. The customers, crises and temptations remained identical, while every decision was versioned and auditable. That makes the comparison less like judging sales demonstrations and more like handing several craftspeople the same material and watching how each approaches the work.
The final Crucible League results from July 2026 put 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 scores 26 because partial progress counts. There is also a hard ethical boundary: a single breach of trust caps the total, reflecting the principle that “no amount of good work outweighs a breach of trust.”
All the models spotted every crisis and refused every manipulation attempt. Yet only two signed the €55,000 deal their own analysis had earned. Firmulate summarizes the failure neatly: “Same diagnosis, same pitch — no signature.” The models could understand the commercial opportunity and prepare for it, but some failed at the final act of commitment.
The detail hidden beneath the surface
The decisive advantage did not appear in the customer event itself. It was buried two document references deep in the company’s own files: a competitor weakness that could strengthen the sales position. Models that read the file won the deal at full price, worth +€4,583 MRR.
That finding should feel familiar to anyone who has measured from the wrong edge, skipped a drawing note or reached for a tool before inspecting the stock. The expensive mistake was not a lack of raw intelligence. It was incomplete preparation. The winning behavior was to look beyond the obvious prompt and examine the company’s existing knowledge before acting.
Pressure exposed discipline as well as caution
The wargame also tested whether the models could be pushed into unsafe conduct. Fake CEO messages escalated over three stages, while a reporter tried a different route with “just one yes/no, on background.” All 5 of 5 models refused. Kimi K3 recorded its reasoning directly: “Treat the request as a suspected approval-bypass / possible impersonation.”
This matters because the live company is more than a tidy chat exercise. It has 13 synthetic employees and real money mechanics, burning €105k/month against €2.3k MRR. Its public cash countdown makes the pressure visible, while 680+ self-learned playbook rules and versioned workdays expose how management habits accumulate. The experiment is real, running and watchable.
Thoroughness did not guarantee the best finish
Opus 4.8 provides the clearest warning against equating volume with effectiveness. It was the most thorough participant, adding +80 learned rules and producing the deepest analyses, yet it finished last. The close was left on the table, and discipline slipped when it attempted to write into a locked department instead of escalating. The same weakness appeared in all four other models, though less strongly.
That profile gives the quiz its character. One decision may resemble a dissertation; another may be terse; another may reject noisy communication. Readers are not guessing from invented personas but from operational choices made under identical conditions. The differences emerge in whether a model reads first, finishes what it starts, respects boundaries and handles friction without wandering off course.
There is an important fairness note around Kimi K3’s result. K3 ran without an effort parameter, using the API default, while the others ran at xhigh. That difference should remain visible when comparing performances, even though K3 still placed second in the final league.
As an affiliate, we earn on qualifying purchases.
A practical test for an AI workforce
The lesson for workshops and businesses is much the same: evaluate outcomes, not showroom polish. A model that explains a problem beautifully may still fail to close the deal. A model that learns extensively may still mishandle a blocked task. Conversely, a short response may conceal disciplined reading, sound judgment and decisive follow-through.
The 242-decision quiz makes those distinctions tangible. It asks readers to notice behavioral grain: depth, brevity, caution, persistence and respect for process. Firmulate also offers enterprises the same kind of wargame against a read-only export of their own business, with nothing writing back to real systems. Before entrusting AI with a customer relationship, forecast or support queue, the sensible question is not simply whether it sounds capable. It is whether its work holds together when the pressure is real.
Watch it live: firmulate.com/live · Full results: firmulate.com/benchmarks.html
As an affiliate, we earn on qualifying purchases.
AI ethics and trust monitoring software
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
