Sanded to Perfection, Never Assembled: What a Woodshop Lesson Taught Me About AI Managers

Every woodworker knows one: the craftsman with the best-equipped shop on the street. Chisels sharper than anyone’s, lumber acclimated for weeks, cuts measured three times. And yet their project — the beautiful hall table started last spring — sits in pieces, unfinished, while the neighbor with a beat-up jobsite saw has a completed bench on the porch. In the workshop, diligence is necessary but never sufficient. At some point you have to glue it up, clamp it, and finish it.

It turns out AI models have exactly the same problem. And now there’s a live experiment that proves it.

The Worst Week in Business, Run Four Times

At Firmulate, four frontier AI models were each handed the same job: run the same small software company through its worst week. Same customers, same crises, same temptations to cut corners — only the model changed. Every decision was versioned and auditable, so nothing could be hand-waved after the fact. The final league table from July 2026: gpt-5.6-sol scored 95, Kimi K3 scored 93, Sonnet 5 scored 88, a second Sonnet configuration scored 77, and Opus 4.8 finished last at 73. For context, doing nothing at all scores 26 — and a single breach of trust caps the whole total, because no amount of good work outweighs broken trust.

The Star Student Who Flunked the Job

Here’s what makes Opus 4.8’s story worth telling to anyone who has ever spent a weekend flattening a board by hand: it was, by a wide margin, the most thorough participant in the entire field. Over the run it accumulated 80 self-learned playbook rules — more than any other model — and produced the deepest analyses of any participant. If the experiment graded homework, Opus would have topped the class.

It finished last anyway. Two things sank it. First, the close was left on the table: a €55,000 deal that Opus’s own analysis had earned went unsigned — same diagnosis, same pitch, no signature. Second, discipline slipped under pressure: when it ran into a locked department, it kept attempting writes instead of escalating the problem, the organizational equivalent of forcing a screwdriver into a Phillips head rather than going to find the right driver.

If that sounds like a uniquely Opus flaw, it isn’t — and to be fair, the data says so. The same weakness appeared in all four models, just weaker. The failure mode of thoroughness is universal: prepare, analyze, document… and somehow never drive the last screw home.

The Buried Fact That Separated Winners from Also-Rans

The most striking finding of the experiment is how small the gap between winning and losing actually was. All four models spotted every crisis. All four refused every manipulation attempt — including a social-engineering gauntlet of fake CEO messages escalating over three stages, plus a reporter offering an easy “just one yes/no, on background” trap. Five out of five refusals across the field, with Kimi K3’s on-record reasoning reading like a veteran security officer: “Treat the request as a suspected approval-bypass / possible impersonation.”

But only two models signed the €55,000 deal. The difference? The decisive competitor weakness wasn’t in the customer event at all — it sat two document references deep in the company’s own files. The models that actually read the file won the deal at full price, worth +€4,583 in monthly recurring revenue. The models that didn’t, didn’t.

Woodworkers will recognize this immediately. The answer to why a joint won’t close is rarely in the joint. It’s in how the lumber was milled, two steps back in the process. You have to read the whole board.

Why Any of This Matters

Firmulate isn’t a one-off demo. It runs AI models as complete companies — 13 synthetic employees, real money mechanics, a public cash burn of €105k per month against €2.3k in MRR, all watchable with a live countdown at firmulate.com/live. The self-learned playbook has grown past 680 rules, and every workday is versioned. There’s even a “guess the model” quiz built from 242 real, unedited management decisions at firmulate.com/quiz.html — a humbling exercise for anyone confident they can tell AI judgment from human judgment.

One fairness note the experimenters themselves flag: Kimi K3 ran at its API-default effort setting while the others ran at maximum effort — and still nearly won. Enterprises curious whether their own operations would survive this treatment can run the same wargame against a read-only export of their business, with nothing ever written back to real systems.

The lesson from Opus 4.8 is the lesson every workshop teaches eventually: prioritization beats volume. The most careful participant lost to models that did less analysis but finished the job — read the files, made the call, signed the deal. As AI agents move toward your CRM, your support queue, your forecast, the question isn’t whether they write beautifully or analyze deeply. It’s whether they finish what they start. A perfectly milled board that never becomes a table is just expensive firewood.

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

Powered by Thorsten Meyer AI