Artificial intelligence

Every e-mail in the right box

Large companies receive hundreds of e-mail enquiries every day. But who answers which e-mail? Assigning them isn’t as simple as it seems. Swisscom has developed algorithms to take over the chore of sorting e-mails. Artificial intelligence already works more precisely than manual assignment by humans.

Text: Ladina Camemisch, images: © Strandperle, 13 september 2017

Anna works at the help desk for an international company. Every day, hundreds of enquiries land in her inbox. Anna can answer some of them directly, but she has to forward many others: to customer advisors, technical experts or to the invoice centre. Sometimes Anna has no idea who to send the enquiry to. She starts researching and forwards the e-mail to whoever she thinks is best. Sometimes an e-mail like this embarks on an odyssey that lasts days before it ends up with the right person to answer it. The customer has to wait days for an answer. But time is valuable, both for customers and for Anna.

So how can the company speed up the process? The answer is artificial intelligence.

In a feasibility study, Swisscom and its joint venture Swisscom Digital Technology (SDT) developed a solution for automatic e-mail sorting. The artificial intelligence records the content and sends the e-mails directly to the right recipient, without any need for manual e-mail assignment. But the solution is capable of much more. It anonymises confidential data, recognises the language and identifies certain keywords, hot topics or VIP customers.

The feasibility study was a success – experience from the pilot project is now being used to refine the system further. David Rüfenacht, Product Manager at Swisscom Digital Enterprise Solutions, says: “The success or failure of an AI solution is usually due mainly to the people in involved, rather than the technical implementation. That’s why at the start of every project I make sure everyone at the table is talking about the same thing.” For many of those taking part, AI is something completely new. Some customers’ expectations are unrealistic or their data quality isn’t sufficient – but artificial intelligence is absolutely dependent on the available data pool.

How e-mail sorting works

To train the AI system using what is known as the supervised learning method, it was first provided with a set of structured and unstructured e-mails. “Using e-mails that had been processed, we set up an initial algorithm and trained the system with the basic functionalities”, explains Rüfenacht. This phase took around a month. For further improvements, Rüfenacht and his team then concentrated on the processing and classification of the e-mails in real time. To extract context information from the e-mails on the content and sender, the experts used advanced AI techniques such as named entity recognition, boosting, keyword extraction and pattern recognition. Finally, company staff spent four weeks monitoring the results and using a click to rate whether the e-mail was forwarded to the right department. The feedback from these employees was used to refine the algorithm, so that the AI system continued learning all the time. After only eight weeks, the feasibility study was already completed.

A better hit rate than humans

At the beginning, the success rate for automatic e-mail sorting was 50 to 60 percent. After the feedback from the staff, the system demonstrated a steep learning curve and reached a score of around 90 percent. “This means allocation of e-mails by AI is more accurate than by humans”, says Rüfenacht. The results are therefore very promising. With the help of artificial intelligence, the company has dramatically reduced its response time to tickets, reduced its operating costs and downtimes, and at the same time increased productivity. The additional context information is also very valuable to the company, because it makes it easier to process the e-mails.

Building on the experience from the feasibility study, Swisscom has continued to improve the e-mail assignment solution. The client is confident in the system and will now use AI to assist its help desks. For Rüfenacht, the case is clear: “Every large company is faced with an almost overwhelming flood of e-mails – whatever sector it operates in. Artificial intelligence can help stem the tide.”

Focussing on the essentials

And what about Anna? She still gets lots of e-mails. But now she no longer needs to worry about forwarding them, and can get on with customer support straight away. Because she knows that she only gets the e-mails that are actually intended for her. And that’s half the job done already.