Data Department built a machine learning model for a beverage bottler that predicts which outlets' coolers will underperform, before the company's own evaluation cycle identifies them. Field teams could then act earlier, by redeploying coolers or targeting sales effort. The model reached 99% accuracy, and the estimated upside is about 2.35% in annual sales.
The decision
Beverage companies in FMCG (fast-moving consumer goods) place branded coolers in shops, cafés and restaurants. A cooler is an investment: it costs money to buy, place and maintain, and it earns its keep only if the outlet sells enough.
The bottler already reviewed cooler performance on a regular evaluation cycle. The problem was timing. By the time a cooler was confirmed as underperforming, months of potential sales had already been lost.
The decision we supported was simple to state: which coolers should field teams look at now, and what should they do about them? Two actions were on the table:
- Redeploy the cooler to an outlet where it is likely to sell more.
- Target sales effort at the outlet to lift its performance.
What we built
We built a machine learning model that flags, ahead of the company's own evaluation cycle, which outlets' coolers are likely to underperform. The output is a list that field teams can act on.
| Element | What it means here |
|---|---|
| Prediction | Will this outlet's cooler underperform? |
| Timing | Before the company's regular evaluation cycle |
| Users | Field teams responsible for outlets and coolers |
| Actions | Redeploy the cooler or target sales at the outlet |
The value of the model comes from timing. A prediction that only confirms what the next evaluation will show is not worth much. A prediction that arrives earlier gives field teams time to act.
Results
- 99% accuracy in flagging outlets whose coolers would underperform, ahead of the company's own evaluation cycle.
- Estimated ~2.35% annual sales upside from acting earlier on underperforming coolers. This figure is an estimate.
Lessons
Earlier is the whole point. The bottler already knew how to find underperforming coolers. What machine learning added was time: the chance to act before the next review rather than after it.
Link every prediction to an action. A flag on its own changes nothing. We framed the output around the two things field teams can actually do, redeploy or target sales, so the list maps directly to work.
Read accuracy in context. A 99% accuracy figure is strong. With any model like this, we agree with the business which errors matter most: a missed underperformer or a false alarm that sends a field visit to a healthy outlet. That choice should be made on business grounds, not left to the default settings of a model.
Estimate the upside honestly. The ~2.35% figure is an estimate of what earlier action could be worth. In general, the way to turn such an estimate into a measured result is a controlled test in which field teams act on the predictions for part of the outlet base.
This kind of outlet-level prediction is relevant across retail and e-commerce and distribution and logistics, and uses the same methods as our churn and customer analytics work: predict who is at risk, then decide what to do about it.
Frequently asked questions
What did the model predict?
It predicted which outlets' coolers were likely to underperform, before the bottler's own evaluation cycle would identify them. Field teams could then redeploy coolers or focus sales effort earlier.
What does 99% accuracy mean in practice?
Accuracy alone can hide which kinds of errors a model makes. For any model used over time, performance should be tracked on new data, with attention to the errors that matter most to the business, not accuracy alone.
How was the ~2.35% sales upside calculated?
It is an estimate of the annual sales gain from acting earlier on underperforming coolers, and we present it as an estimate.