McGrocer's catalogue grew faster than people could sort it.

A UK grocer with a lot of back-office work.

McGrocer is a UK online grocer. Its products come from supermarkets, wholesalers and specialty retailers, which means a constant flow of new items and updates from many different suppliers.

McGrocer came to us with a broad goal: use AI to make its back-office processes faster and less manual. The biggest piece of that work became category mapping.

Every product needed a home, and finding it was slow.

Each supplier sends products in its own format, with its own names, its own fields and its own categories. Those categories look helpful, but they come from someone else's shop. A supplier's "Dairy" is not McGrocer's "Dairy".

So products often landed in the wrong place. Shoppers could not find things the site definitely stocked, and navigation filled up with noise.

The review team had become the bottleneck.

The fix had always been people. New items were checked and moved by hand. That works for a few thousand products. It does not work for 200,000.

The team that was meant to protect quality ended up setting the speed limit. Supplier updates that should have reached the site in minutes were taking days.

On top of that, McGrocer had redesigned its taxonomy to eight levels. That was good work. But mapping into it still depended on individual judgement, which varies from person to person and from day to day.

We built an AI engine that reads each product two ways.

First, every incoming product is cleaned up into the same shape: title, description, ingredients, brand and attributes.

Then two independent signals look at it. A language model reads the product and proposes where it belongs in the taxonomy. A vector similarity search finds the closest products McGrocer has already classified and checks the proposal against them.

Each signal covers the other's weak spot. The model understands meaning but can suggest a path that does not exist. Similarity search is grounded in the real catalogue but struggles with products unlike anything stocked before.

When the two disagree, a person decides.

A scoring step compares the two signals and gives every product a confidence score. Simple rule checks catch broken paths, such as a category that does not exist under its proposed parent, before they go anywhere.

If the signals agree and the rules pass, the product is published automatically. If they disagree, or confidence is low, it goes to a human review queue.

That way the team's time goes on the genuinely ambiguous products, not on the thousands of obvious ones. Their corrections feed back into the engine, so it keeps improving as the catalogue grows.

Why not a traditional classifier?

A supervised classifier is cheaper and faster when you have plenty of labelled examples. McGrocer did not have them yet: the taxonomy had just been redesigned, and a grocery catalogue has thousands of categories that each hold only a few products.

The taxonomy also keeps changing. Our approach can re-read a new structure and re-check earlier classifications in bulk, instead of starting a labelling cycle again.

What changed.

The engine maps products with 95%+ accuracy across all taxonomy levels, over a catalogue of 200,000+ SKUs.

New products are processed in batches as soon as they arrive from a supplier feed, so they reach the site within minutes rather than days. Large supplier batches that used to mean months of catalogue work now take hours.

People are still in the loop, but only where they add something: the hard cases.

Where McGrocer is now.

The engine learns from corrections, and corrections come from the products people look at most. So the long tail of rarely viewed products needs deliberate checking, or it improves more slowly than the rest.

The taxonomy itself still needs an owner. Deciding when a category deserves its own branch, or when two have come to mean the same thing, is a judgement for McGrocer's team. The engine is built to re-check everything when they make that call.