What drove the decision to develop an AI-based use case?
Among other things, we develop solutions for our customers in Microsoft 365, especially for SharePoint, which is mainly used for document management in small, medium and large companies. One customer complained that his employees had to deal with a wide variety of business documents on a daily basis (e.g. orders, applications, letters from authorities, contracts, etc.) and had to manually categorize them in a time-consuming manner to make them easier to find and to trigger automatic workflows.
We then implemented an app ("Easy Tagging") for Microsoft 365 that uses a few sample documents to learn how users tag or categorize documents. After a learning phase, the app adds appropriate metadata to existing or new documents. The advantage for users: They find documents more efficiently and are relieved of routine tasks with documents because automatic workflows with documents can now be implemented.

What were your expectations or requirements for the AI application?
For our customers, it is important that the Easy Tagging App classifies unknown documents with a very high hit rate and specificity and processes large document libraries quickly - and with as few sample documents as possible for training. In addition, the app is expected to be certified by Microsoft, ensuring seamless integration with the familiar SharePoint user interface and software quality.
It is important for us that users all over the world can install the Easy Tagging App directly from the Microsoft Store ("AppSource") into their SharePoint and start using it without any training in AI technologies.

What frameworks/methods did you rely on for development?
Microsoft 365 and Azure as the platform for the application and various Python libraries for text analysis (including NLTK) and machine learning (Tensorflow).

The use of AI solutions must also be accompanied by a change in competences. How have you dealt with this challenge in respect of your employees?
We have solid, field-proven expertise in symbolic artificial intelligence ("semantic technologies") that use explicit rules and ontologies to represent complex models. Since this approach is not optimal for Easy Tagging requirements, we use "machine learning" methods. One of our employees acquired experience with neural networks in his diploma thesis, which put us in a good starting position for the development of the app. Nevertheless, we have deepened our expertise in the context of personal IBM certifications and updated it, especially for newer approaches like Deep Learning.

In what timeframe did your company implement the AI solution?
Since our product development runs alongside our client projects, it took us about a year to implement, which is much longer than originally planned. We spent about half of that time on the actual AI components. Microsoft's certification of the app also took a few weeks.
However, our customers can install the Easy Tagging App from AppSource, get it up and running, and process their document assets with just a few simple clicks.

A conclusion to your AI solution:
The implementation from the product idea to the certified app presented us with some challenges. In particular, the maturity of some central Python libraries left a lot to be desired and was therefore a major effort factor.
On the other hand, the available machine learning libraries allow the implementation of innovative features for text analysis, e.g. automatic summaries or the extraction of facts ("who, what, where") and contexts. These features directly benefit the SharePoint user, who finds and evaluates documents faster and, thanks to automated workflows, is relieved of routine tasks.

One final piece of advice/tip to other entrepreneurs looking to apply AI:
The use of AI methods should not be a goal in itself, but it should fulfill concrete requirements of the users.

More info at:
https://www.diqa-pm.de
https://www.linkedin.com/company/diqa-projektmanagement-gmbh