Coca-Cola's 83% adoption figure covers a tenth of its network

Coca-Cola now tells Malaysian shopkeepers what to order and how much, through a feature called Perfect Basket inside its Coke Buddy ordering app. AI News reported the rollout. The engine reads past orders, ordering frequency, seasonality, weather and what similar shops nearby are buying. Coca-Cola has not published the architecture, though systems of this kind usually rank options with a Neural Network trained on past behaviour rather than following fixed rules.

The number attached to the story is 83% adoption. It describes a much smaller group than the headline suggests.

“The available Malaysian campaign data does not provide figures for forecast accuracy, stock availability, inventory levels, or logistics costs.”

— AI News, AI News, 8 September 2026

Quote source: AI News, 8 September 2026

Coca-Cola AI ordering: 83% of a tenth of the network

Coke Buddy serves about 39,000 outlets in Malaysia. The 83% figure comes from a campaign that ran from January to April, drew more than 4,500 entries from over 4,000 retailers, and rewarded taking part.

So 4,000 shops out of 39,000 is 10.3% of the network, and they chose to enter a contest. Of those, 83% followed the recommendations, which works out at roughly 3,320 outlets, or 8.5% of the network. The confirmed figure is 8.5%, not 83%.

Coca-Cola also said participating shops grew sales revenue faster than comparable ones, without saying by how much. A comparison with no size attached is a direction, not a result.

The four numbers nobody published

The publication did something unusual and listed what is missing rather than leaving it out. Four things would show whether the system forecasts well:

• Forecast accuracy of the recommendations.
• Stock availability in the shops that followed them.
• Inventory levels held by those shops.
• Logistics costs across the network.

None of them were published. What was published is that shops bought more, which is the outcome the seller cares about. Whether the shopkeeper ended up with the right stock at the right moment, and whether the trucks ran cheaper, remains unmeasured in public.

This matters because the same tool is either a forecasting system or a sales system depending on which numbers you publish, and only one of those needs to be accurate to earn its keep.

What the company has shown elsewhere

Elsewhere the company has more history. It told investors in early 2024 that it and its bottlers had connected nearly eight million customers to business ordering platforms, with AI suggested orders reaching more than three million outlets in Latin America.

Two performance figures exist from other projects. Retailers given AI recommendations were more than 30% more likely to buy the suggested items in early pilots, and a separate three-country trial that added weather and location data recorded sales 7% to 8% above shops without the algorithm. Neither concerns Malaysia, and the article says so.

The pattern is familiar from this week. We wrote about an AI transcription tool sold on time saving whose own evidence pointed the other way, and about a trade that AI removed entirely while nobody counted the people. Adoption numbers are easy to publish. Accuracy numbers are the ones worth waiting for.

Watch for a second release with inventory and stockout data across the full 39,000. Until then the honest description is that a tenth of a network tried a recommendation engine during a promotion, and most of them took its advice.

None of this should be read as personalized investment advice.