If you want to sell across borders, you have to consider a few things. In particular, if there is a customs border between the place of storage of the goods and the delivery address of the customer. For import and export, each item requires a customs tariff number based on its properties. Because without them nothing works at customs! However, identification is a manual, time-consuming process in many companies.
As a specialist in cross-border logistics, MS Direct clears the way for many millions of items to enter the Swiss market. To handle tariffing for all these goods, we rely not only on the knowledge of our customs experts, but also artificial intelligence (AI) using machine learning – our TariffTranslator. The automated determination of the customs tariff number from master data significantly simplifies customs processing so that goods can cross the border smoothly and within the shortest possible time.
How exactly does this work, you ask? Eva Tyssen, our Head of Business Development and Customer Success, conducted an interview for you with Prof. Dr. Siegfried Handschuh from the University of St. Gallen (HSG), who helped us with his team in the development of AI.
You have to realise as a company that you are sitting on treasure with your process data.
Prof. Dr. Siegfried Handschuh, Universität St. Gallen (HSG)
Eva: Siegfried, we approached you almost two years ago with the desire to automate our tariff number determination. How quickly did you realize this was a perfect use case for machine learning?
Siegfried: In the beginning, we suspected it, but we didn’t know it until we worked together on a proof of concept and were able to use concrete sample data. It quickly became clear that we could achieve very good results with machine learning. MS Direct’s good data makes this use case ideal.
Eva: How did we train the machine? What did we need to do this and how long did the process take?
Siegfried: We did a preliminary study and started the actual project with this knowledge. The project took about three months, of which about one month was spent on data handling, one month on working with the training algorithms, and one month on the user interface. Such projects always require a comprehensive exploration and preparation of the data.
The preparation of the data – we had about 57 million records – includes a quality check, an examination of patterns, linking of data sources, and detection of anomalies. With the data thus processed, we developed a total of 75 machine learning classifiers for the task. These models are trained, validated and fine-tuned. In addition, we have created an infrastructure that enables subsequent post-training, making the solution open to the future.
Eva: What do you generally recommend to companies that want to optimize processes with AI?
Siegfried: I see two things directly. First, it must be determined whether the processes can be addressed with AI at all, i.e., what does the current solution look like in terms of costs, error rates, and process inefficiencies, and what is expected from the AI solution? Secondly, modern AI is in many cases data-driven, which means that as a company you have to recognize that you are sitting on a treasure trove of process data that needs to be mined. Or, if you don’t have the data you need, consider strategic data sourcing. The data can then be used to automate business activities and processes.
The most important resource for AI and the quality of AI models is data. Therefore, companies need to develop a data and AI strategy.
Eva: Thank you very much, Siegfried, for the interview!