Pick n Pay's AI came after two years of groundwork. Your delivery app in South Africa needs its data in place first.

Pick n Pay is rolling out an AI store operations platform from OrderGrid to more than 720 stores, after a pilot cut the time staff took to pick online orders by about 30%, as TechCentral reported on Pick n Pay's AI roll-out. A 30% gain is worth having. TechCentral also says the roll-out is the final stage of a two-year overhaul of the retailer's technology platform, and that is the part a business owner should take from it.

Why Pick n Pay's AI arrived at the end of a two-year overhaul

AI only gets good at running a store when it sits on clean stock and order data. Pick n Pay already used computer vision to spot empty shelves and machine learning to plan picking routes before this deal. The new platform covers real-time inventory, demand forecasting, order management and in-store tasks, and each of those depends on the numbers underneath being right. A picker sent to a shelf the system thinks is full is slower, whatever the software is called.

South African shoppers are moving online fast. World Wide Worx found that South African online retail grew 35% in 2024, reaching R96 billion and 8% of total retail, while physical retail grew 2.5%. In my view, growth like that makes it tempting for a smaller grocer to buy AI inventory management before the basics underneath it are in place.

How we proposed a delivery platform with no AI at launch

In the platform we proposed for an on-demand grocery delivery start-up, AI stays out of the entry tier on purpose. The proposal covers a customer app, a driver app, an admin dashboard and one backend. That entry tier runs on manual driver assignment and one open database for stock and orders. We designed that database so AI features could be added later without a rewrite or a data migration.

The upgrade tier adds a dispatch engine that would rank available drivers by distance from the store and avoid loading one driver with too many orders at peak. It would only make suggestions. The person running the store would keep full manual control and could override every one of them. We recommended that demand forecasting come later still, six to twelve months after launch, once the business has its own order history to learn from.

A forecast trained on a few weeks of data is a guess with decimals. The start-up already has one reliable monthly peak, because student allowances are paid in the last week of each month. You do not need a model for that. Promotions can be timed to that week and stock sized to it, and range and prices can be adjusted weekly from live order data in the admin dashboard. This is what I mean by software that is ready for AI before it uses any.

What to ask before you pay for AI

If you run a shop, a restaurant or a distribution business and you are planning a delivery app, start with a proper stock management system before anything clever. One record of what you have, what was ordered and what was delivered. Then run the operation by hand on top of it for a few months.

When a vendor or developer pitches AI, ask two questions. Where does my stock and order data live, and can I take it with me if I leave? How much of my own history does your AI need before its forecast beats my own judgement? A vague answer to the first one is a reason to walk away, because owning the data and rules your AI runs on is the part you cannot buy back later. It is also how we approach custom software development: the data model is built for AI from the start, and the AI itself comes when the data can carry it.

An SME that buys an AI inventory tool before it has a clean stock record gets a forecast of its own mess.