When Artificial Intelligence sorts the mailbox and assigns tasks

Artificial intelligence (AI) is one of the hyped topics these days. Last November, the German government adopted a future-program for this technology in the amount of 3 billion EUR over the next seven years. The Chinese government even plans to build a $ 150 billion AI industry. 

AI is no science fiction anymore: It is already in the here and now. At ERGO, we use this technology to solve existing problems in a better and faster way than we could do before.

Today, we have launched another AI use case.

In the future all e-mails at ERGO will be read by a trained AI algorithm, which can assign meanings to words and word combinations. Based on that skill, the e-mails are subsequently forwarded to the most likely target address.

At the Moment, already more than 300,000 e-mails are sent to ERGO's 120 different mailboxes. Currently, only 2/3 of these e-mails are addressed to the correct business units. The remaining e-mails either reach wrong business units or are sent to one of the central info addresses. All these e-mails – about 500 per day – must be read by our colleagues from ERGO Direkt and forwarded to the correct recipient individually. Hence, they have to deal with extra work to process e-mails.

Here, the launched AI application comes into play. It automatically assigns all incoming e-mails to the correct target mailbox. However, how can AI manage this task?

To determine the correct recipient, the computer must understand the content of the e-mail. In contrast to a human, a computer cannot easily comprehend the meaning of words. Therefore, to understand the content of a text, e.g. an e-mail body, the words must be translated in a machine-readable form.

For AI applications, which use modern text recognition models, every single word is translated into a number or a multidimensional vector, respectively. Words with similar meaning or words, which are often used in a similar context, are represented as vectors in proximity to each other. This method, the so-called word embedding (Word2Vec), was applied to all e-mails which had reached ERGO in the last two years. By this, a set of data was generated in order to train an artificial neuronal network. This neuronal network is now able to recognize syntax and semantics, as well as text patterns. It forms the centerpiece of our so-called AI classifier.

The development and implementation of the AI classifier has been carried out as part of a joint project between ERGO Direkt, ERGO Digital IT, ITERGO, and ERGO Innovation Management. Thanks to this solution incoming e-mails are now forwarded directly to the correct destination addresses with an accuracy of over 90%. This makes us one of the pioneers in the market – which I am very proud of.

However, regarding the usage of AI, we are still at the beginning, but we will quickly move forward. In the insurance business we speak our own language, we have our own data sets and our own fields of application. In this industry-specific environment, we can learn faster than some technology companies can.

I am looking forward to your comments, questions, or suggestions.
Yours Mark Klein