by Serguey Shinder
In 2024 we started using a model to help our credit team assess applications for trade accounts. A builder or a plumber applying for thirty days' credit has to be checked: company records, payment history from a credit agency, how long they have traded, who the directors are. The model read all of it and recommended a decision and a limit. A credit controller then reviewed each one and approved it, changed it or turned it down. A person always made the final decision, and we said so often, to the auditors and to ourselves.
Six months in, the figures looked excellent. Our controllers agreed with the model on ninety nine per cent of applications. I read that as evidence the model was good. Our finance director read the same report and asked a different question: how long were the reviews taking? The answer was forty seconds on average. Before the model, a manual assessment took about twelve minutes.
Then one of the accounts went bad. A groundworks firm with a ten thousand pound limit stopped paying after two months. When we looked, one of its directors had been a director of two other companies that went into liquidation owing us money. It was in the information the model had been given. The model had not weighed it heavily, and the controller who approved the account had not opened the director history at all, because the recommendation was green and the queue was long.
I do not blame her. We had built a process in which the easiest thing was to agree, and agreeing was right ninety nine times in a hundred, so it became a habit. A review that always agrees with the thing it is reviewing looks the same as a review that is not happening. Our ninety nine per cent had been measuring the model and the habit together, and we could not tell them apart.
So we changed what the controller sees. For one application in five, chosen at random, the controller now makes their own decision first, and only then sees the model's. We track how often the two differ and why. Every recommendation lists the three facts that most moved it, so disagreeing means looking at something specific. And applications the model is least sure about, or where a director has any history of insolvency, go to a controller without a recommendation at all.
Agreement on the blind sample settled at about eighty eight per cent. That gap is the part of the job our people actually do.
– Serguey Asael Shinder
Leave a Reply