Coca-Cola AI reordering: 83% adoption, check the base

Coca-Cola disclosed usage numbers this week for Perfect Basket, the AI reorder recommender inside its Coke Buddy ordering platform in Malaysia. The headline stat: 83% of participating outlets adopted the recommendations. That is a good number. It is also a number attached to a contest, and the difference between those two facts is most of what a small operator should take from this story.

What actually happened

Coke Buddy is a self-service ordering platform — app, web, and WhatsApp — supporting roughly 39,000 retail outlets in Malaysia, per AI News. Perfect Basket sits inside it and suggests which products a shop should reorder and in what quantity, drawing on previous orders, ordering frequency, seasonality, weather, and purchasing patterns. The retailer reviews the basket and still places the final order.

The 83% figure comes from the "Perfect Basket, Perfect Ride" campaign that ran January to April 2026: more than 4,500 entries from over 4,000 retailers. Coca-Cola says outlets that followed the recommendations recorded higher sales revenue growth than comparable outlets — but, as the reporting notes plainly, the company did not disclose the size of that difference, and published no figures for forecast accuracy, stock availability, inventory levels, or logistics cost.

So: 4,000 self-selected contest entrants out of a 39,000-outlet network, with an unquantified lift. Roughly 10%. Similar suggested-order features have reportedly reached more than 3 million outlets in Latin America, which tells you the pattern works at scale even if this particular stat doesn't prove it.

Why AI reordering matters for your business

The mechanism is boring and that's the point. Order history plus seasonality plus weather is not a frontier model. It is a demand forecast with a good interface. If you run a shop, a distributor, or any business where somebody guesses at reorder quantities on a Friday afternoon, this is the highest-ROI AI you can build, and you can build it against data you already own.

Keep the human as the approver, not the bystander. Coca-Cola's design is the right one: the system proposes a basket, the retailer edits it, the sales rep still shows up. That is what makes adoption possible — nobody has to trust the model to be right, only to be a better starting point than a blank order form. Design your own automations the same way.

Learn to read the denominator. "83% adoption" during an incentivized campaign measures the campaign as much as the model. When a vendor pitches you an AI feature on a percentage, ask three questions: percent of what population, over what period, compared to what baseline. If they can't answer all three, the number is marketing. This is the same discipline we applied to Meta's coding-agent throughput claims.

Your product and order data is the asset here, not the model. Perfect Basket works because Coca-Cola owns the transaction history for 39,000 outlets. If your ordering flow lives entirely inside somebody else's marketplace, you are training their recommender, not yours. We have made that argument about product data before, and reorder history is the same argument with a shorter feedback loop.

Sources: AI News.