A retail customer wanted AI to forecast daily sales, item by item, in every shop.
At first, the request sounded reasonable. Forecasting is one of those use cases that makes sense on paper. Better predictions, better stock levels, fewer missed sales, less waste.
Then we looked at the data before looking at the model.
Most of the catalogue sold zero units a day in any given store. A few iconic pieces sold steadily, but even those moved only a handful of units a day. At that level, the signal is thin. You cannot build a reliable daily per-item forecast on a handful of events and expect the model to invent certainty.
So we stopped trying to forecast each item.
That was the turning point.
The real question was not “How many units of every item will each shop sell tomorrow?”
The real question was: “Which items can the shop not afford to be without?”
The Decision Was Hiding in the Sparse Data
Sparse data does not mean useless data.
It means you have to ask a question the data can actually support.
The proven sellers were easy to identify even when their daily quantities could not be predicted with precision. They earned a permanent stock slot because availability mattered more than perfect forecasting.
The long tail was different. For those items, the better answer was not to pretend the store could predict demand item by item. It was to change the operating model: order in store and deliver to the customer, or offer a similar item already on the shelf.
Same data.
Different question.
The forecast everyone asked for was never going to come. But the decision the business needed was sitting there the whole time.
This is a common AI failure pattern. The model gets blamed because it refuses to answer a question the data could never support.
In reality, the project failed earlier.
At the framing step.
Forecasting Was Also a Knowledge Problem
There was another issue in that retail business.
For years, forecasting lived in two people’s heads.
Two people understood the logic well enough to run it. Everyone else depended on them. That looked like expertise, and it was. But it was also a bottleneck with two single points of failure.
The tools we built did not only sharpen the forecast.
They moved knowledge out of two heads and into the system.
That does not mean the expertise disappeared. It means the expertise became available to the operation. Someone in the business could run forecasting without years of informal training behind them. The system could prepare the order book, and the store manager could review, accept, or change it.
That distinction matters.
The calculation moved into the system.
The decision stayed human.
Production AI Changes the Job
A pilot can produce a better forecast.
Production changes who can do the job and what they spend their attention on.
At the end of each day, every shop could receive an estimated order with items and quantities already calculated. The store manager no longer had to build the order from scratch. They could study it, adjust it, and apply local judgment where the system did not know enough.
That is a better use of human attention.
Not because the human was removed.
Because the human was moved to the part of the work where judgment mattered.
This is often the best shape for AI in operations. Let the system prepare, classify, calculate, retrieve, and suggest. Keep the human in the position where context, accountability, and exception handling still matter.
The goal is not to make the business blind by automating decisions it does not understand.
The goal is to make the decision path more visible, more repeatable, and less dependent on one or two people carrying the whole thing in memory.
The Lesson
AI did not fail in that retail case.
It was asked the wrong question.
The useful work began when the question changed from perfect prediction to operational decision: which items must be present, which can be handled differently, and which knowledge should be encoded so the business is no longer dependent on two people.
That is the lesson I keep seeing in AI projects.
The model is rarely the first problem.
The first problem is usually deciding what question the business is actually trying to answer.