In Vattenfall’s purchasing department, most invoices are automatically categorized into a category. However, about 25% is not categorized properly or not at all into the correct purchasing category. 

Manually assigning thousands of individual invoices per week is too labour-intensive. The lack of the right insight leads to greater purchasing inefficiencies. How can we efficiently classify 25% of invoices?


Categorize thousands of maverick invoices

Together with the customer we solved the problem by developing an extraction & classification pipeline. Which can classify fully automated invoices into one of the 500 supplier categories. In the end, the invoice model managed to categorize the correct category in 85% of the cases. With full category visibility, this organization was able to implement better procurement strategies and save 5% on total business expenses.


100.000 maverick invoices


Tesseract OCR, HuggingFace Transformers, DataBricks, PowerBI, ClickView


Prototype, Test and Implementation


Realized in 9 months

Level 4

AI optimization & implementation

AI is never a standalone solution. It requires a lot of fine-tuning and creating the right conditions to successfully land a Machine Learning model in an organization. Hemisphere knows what it takes to take these models to the next level. 

There are dozens of reasons why a model does not make it to implementation. Quality of (input) data, availability & costs of sufficient computing power and management are all well-known examples.

Hemisphere works with you to see what is needed to get more out of your AI solution. You can contact us from cleaning data and setting up HPC instances to offering customized Machine Learning SLAs.

Level 3

New AI model design

A costly hick-up in your process and do you have the feeling that only an AI solution can answer this? 

Hemisphere closely follows global AI development. This gives us a good idea of what is possible today. We have mastered various AI techniques on a broad level.

As a result, in most cases up to 80% of the final solution can be achieved in a design sprint of 2 weeks.

Level 2

Data engineering

The shift to cloud services offers organizations the opportunity to simplify their data processing. Companies that currently rely on batch processing can now implement continuous processing methods without disrupting their current processes. Rather than expensive rip-and-replace, the implementation can be incremental and evolutionary. 

By following the Extract-Transform-Load (ETL) methodology, data is processed and passed on in a qualitative and safe manner. This reduces the chance of errors in the data and provides the management team with the right data to base their decisions on.

Level 1

Data research

Data contains rich and often real-time information for a wealth of applications. By analyzing the large influx of customer contacts (anonymously), you are able to gain almost real-time insight into customer needs. Based on these insights, it is also possible to hyper personalize the offer. 

Hemisphere helps you make analyzes to understand the different trends and opportunities at the micro level. By responding to this, you exceed customer expectations and gain trust.