Mapping, mapping, mapping… it has to be manual in the beginning

Why did we classify our first drink data by hand, and how did that work make AI possible?

Our first batches of ticket lines have arrived, and the data is messy. Very messy.

Messy, very messy

We have received our first batches of ticket lines from bars, cafés and restaurants. Each venue records its products in its own way, so the same drink can appear under many different names, abbreviations and spellings.

To keep things manageable, we decided to focus first on the drinks on each ticket. We will deal with the food later.

Classifying by hand

The data is so messy that we decided to classify every drink manually. To do this well, we developed our own classification interface. It helps our team work faster and, just as important, classify in a consistent way.

How messy is it? Coca-Cola alone is written in 136 different ways in our database. Now imagine classifying beers from micro-breweries. In Belgium there are quite a few of those: about 1,600 beer brands in total. How do you classify all of them?

Coca 33clcokeCCL=Coca-Cola

“Coca 33cl”, “coke”, “CCL”: three ways to write the same drink, and all of them must count as one product.

From hand work to AI

01
Manual classification
Our team maps every drink line by hand, using a dedicated interface to stay consistent.
Result: clean, reliable product data.
02
Training data
Every classified line becomes an example. Over time, these examples add up to a solid training set.
Result: enough data to teach a model.
03
AI-assisted mapping
AI models learn from our manual work and classify new ticket lines at scale.
Result: an accurate mapping system.

Giving AI a try

Once we had classified enough drinks, and so produced enough training data, we decided to give AI a try.

We are fortunate to have access to some of the best AI specialists in the world, in Europe and in a team in San Francisco. They worked with us at a previous company, which was sold to a large US-listed company. Since then, they have moved to the US and built cutting-edge technology in this field.

No shortcut around the manual work

The results have been impressive. But we only managed to build an accurate mapping system thanks to the manual work we did first.

AI is only as good as the data it learns from. By classifying lines by hand first, we gave our models the right examples to learn from, and that is what makes our product data reliable for drink brands today.

Key takeaways

We are focusing on drinks first. Food will come later.
We are classifying every drink by hand, using an interface we built to keep the work fast and consistent.
Once we had enough classified drinks to use as training data, we brought in AI, and the results are impressive.
The lesson: an accurate mapping system starts with careful manual work.

Want to see clean on-trade data for your brand?

Let’s talk data →