The purple pen problem
What a dermatology AI teaches boards about hiring CFOs.
Machine learning has a failure mode that anyone writing “sector experience essential” into a CFO brief should know about: a model that learns the wrong thing from the right data.
Medical AI has the best example I know. In 2019 a group in Heidelberg tested a CE-marked skin-cancer model, trained on 120,000 images, against 130 moles photographed twice: once clean, and once with the purple surgical pen marks a dermatologist draws around a lesion they’re worried about.
On the clean images the model’s specificity was 84%. With the pen marks it fell to 46%: it called more than half of the harmless moles melanoma.
It had learned that purple ink means cancer, because in its training set it usually did. Dermatologists circle the ones they’re worried about, so marks and melanomas came together, and the model learned the marker instead of the disease.
Crop the ink out and specificity went up to 97%.
The ink wasn’t noise. It was the trace of an experienced dermatologist’s judgement, and that judgement was usually right. The model’s mistake was learning the conclusion and skipping the reasoning, so that the moment a mole turned up marked for some other reason, or unmarked, it had nothing underneath.
Sector-specialist hiring has the same shape. The brief says “must have deep experience in [insert sector]”. What the board wants is someone who understands the business fast and makes good decisions in it. What they get is someone who has learned the sector’s purple pen.
The specialist arrives with the markers already installed: the KPIs everyone reports, the multiples it last traded at, the way contracts have always been structured.
In the sector’s past those arrived alongside good outcomes, so they look like the outcome. But they aren’t. They’re the ink.
None of this makes the sector’s accepted wisdom wrong; it was right when it was formed. The specialist has absorbed the conclusions without interrogating the reasoning, which is the only way to know when a conclusion has stopped applying.
In a finance function the false positive is the plan waved through because it carried the sector’s markers, the right KPIs, the standard contract, the usual pricing, while the economics underneath were wrong. The false negative is the idea with no ink on it, dismissed as naive, that would have worked better.
Specialists make both more often, because they trust the markers. Generalists have to look at the mole.
I work as a fractional CFO for SaaS and technology businesses with a raise, a refinancing, a listing or an exit ahead. If that is where you are, I would be glad to talk.