Discover how Unifiedpost Group used AI technologies to implement a scalable solution that provides 250% more accuracy at a lower cost.

About the customer

Unifiedpost Group is a fintech company and one of the biggest players in document digitization in Europe. Unifiedpost Group believes that administrative and financial processes should be simple and smart.

Founded in 2001, the company has won Belgian Scaleup Company of the Year  from Data News, is present in 30 countries, and has over 1300 employees. Unifiedpost Group supports companies in processing documents and payments. They help their customers digitize their business processes. An example of one of those platforms is Banqup, where they manage everything related to invoicing for SMEs.

The problem / challenge

Document digitization is at the core of Unifiedpost Group’ business. The Group handles procurement documents for over 400,000 small and medium-sized businesses and more than 250 enterprise players. The Belgian fintech company processes nearly 350 million invoices and other procure-to-pay docs per year across Europe. It’s of utmost importance that the information automatically extracted is accurate and reliable. Before switching to Document AI, Unifiedpost Group used a template-based solution that needed a template for every document processed. 

There were a couple of pain points in this solution that Unifiedpost Group was looking to alleviate:

The goal

Unifiedpost Group’ goal was to modernise their template-based solution using AI technologies. This would remove the need for creating new templates and make it easier to support SMEs in their digital invoicing. Unifiedpost Group was looking for an AI solution that could make this document digitisation process quicker and more accurate. As they were ramping up their business in Europe very quickly, scalability and cost were 2 key factors in the selection of the AI solution. 

Overall Unifiedpost Group wanted to use AI-driven processes and asked partner Devoteam G Cloud to help them.

The solution

Unifiedpost Group was already working together with Devoteam G Cloud to roll out Google Workspace, so it was a logical partnership to continue when they started to look at Google’s Document AI solution. Devoteam G Cloud helped Unifiedpost Group quickly set up a Proof of Concept on GCP to help them evaluate the performance of Google’s Invoice Parser together with Google’s product team.

After seeing the results, Unifiedpost Group decided to deploy Document AI’s Invoice Parser for their BanqUp platform. In order to ensure scaling to processing possible billions of documents a year, a fully serverless architecture was developed using tools like Cloud Functions, App Engine and Datastore.

In order to lower the impact of this change to the rest of the Unifiedpost Group, the integration went through a proxy using Cloud Endpoints. In combination with the fact that Unifiedpost Groups application was already properly modularized, the change in tech remained transparent to the process of the full flow. 

As accuracy remains one of the main requirements for the application, Devoteam G Cloud also assisted in setting up Google’s Human In The loop tooling. This allowed Unifiedpost Groups’ labelers to verify any labels extracted by the Invoice Parser. In order to match the different pricing tiers that Unifiedpost Group offers, a custom routing application was implemented that combines the confidence of the machine learning model with the subscription type to decide whether or not a result should be reviewed by a human.

The Methodology

Together with the Devoteam G Cloud’s teams and Google’s teams we worked for 9 months on implementing Document AI from A to Z at Unifiedpost Group. 

Devoteam G Cloud adopted the scrum methodology. The project was divided into several epics, which represented the major building blocks of the solution. These epics were in turn divided into smaller parts, also called (user) stories. Each story consists of different tasks that must be completed within a predefined period. This way we spread the project over several sprints, each of about 3 weeks, to finish some of the stories and get one step closer to the completion of an epic and thus the end result. This iterative way of working gave us a lot of room for feedback and testing so Devoteam G Cloud could make the adjustments in the project. The deliverables of the project were very clear for all the stakeholders which is key in such a project. These stakeholders are involved at fixed points according to the Scrum methodology.

The team organised daily stand-ups to follow up very briefly and get rid of blocking factors. This way everyone was kept up to date with the evolution of the project.

The Result

The solution provided Unifiedpost Group with great results:

• Label extraction quality
  Unifiedpost Group noticed an increase in accuracy of 250% when comparing the old templating based solution with the Invoice Parser in Document AI. This has a huge impact on the user experience as they now get accurate information even before any review by a human.
• Lower cost
  Because of the efficiency of the new solution, UPG recorded a 60% lower TCO after deploying the solution. This comes down to two big factors, the first being the cost efficiency of the solution on GCP, and the second being the reduced effort required by human labelers as a direct result of the increased accuracy of the extraction. Apart from the obvious advantage of saving money, it also ensures that employees are not disrupted and can be more effective in their work.
• Scalability
  One of the main goals of this project was to enable Unifiedpost Group to scale their activities across Europe. Again the benefit here is twofold. The first benefit comes from the switch to a fully serverless cloud based environment that is able to scale up to billions of documents. The second benefit comes from the switch to Document AI, which supports 150+ different languages and will help UPG expand to new countries without having to perform a lot of work getting their templating engine to support these languages.