by Serguey Shinder
In March our replenishment model ordered eleven months of a slow moving line. About nine thousand pounds of stock, sitting in bay fourteen, where a good deal of it will still be sitting in 2028. It was a bad order and nobody has tried to argue otherwise.
The model had been proposing the weekly buy since November, with a buyer approving or changing every proposal before it goes anywhere. The meeting about bay fourteen had nine people in it, including a director, and ran for an hour and ten minutes. By the end the proposal on the table was that we switch the thing off and go back to buying the way we used to.
I asked the question that made me unpopular for a fortnight. How often was the old way wrong, and by how much.
Nobody knew. Nobody had ever known. We have been buying stock for nineteen years and there was not a number anywhere in the company for how often a buyer's order turned out to be a mistake, because a buyer's mistake does not arrive as an event. It arrives as stock that moves a little slower than expected, and it is discounted eighteen months later by somebody in a different department, and by then it is not a mistake at all. It is the market.
So two of us built the number out of four years of purchase orders and stock movements. On the meanest definition we could all agree to, which was stock still unsold after a year, the buyers were wrong on roughly four per cent of lines. The model, in its first four months, was wrong on about three. Bay fourteen was the largest single miss either of them had made, which is precisely why we were in the room and why none of the others had ever had a meeting.
That asymmetry is the whole thing. A model's mistake is attributable, dated, sitting in a table, with an owner present. A person's mistake is spread across a market, a season, a supplier and a reorganisation, and it has no timestamp and no meeting. So we put a measured thing next to an unmeasured feeling, and the measured thing loses every time, whatever the underlying quality.
We kept it, with a cap on any proposal that exceeds a line's previous twelve months, and a person required above a threshold.
What I would now do before automating any judgement is spend the fortnight it takes to measure the judgement we already make. Not to defend the replacement. Twice now the measurement has told us to leave a process alone. But because that number is only available beforehand, and without it every incident afterwards becomes a referendum. An error rate is not a fact about quality. It is half of a comparison, and the other half is usually written down nowhere.
– Serguey Asael Shinder
Leave a Reply