Faster processing of service requests

Companies serving millions of customers require efficient services and communication channels. The Deutsche Telekom service team is responsible for the service business and offers Deutsche Telekom’s customers a wide range of different contact options: these include AI-based apps such as MeinMagenta or Frag Magenta, that provide a high self-service solution rate, as well as social media channels, chat, web forms, and the conventional email and telephone options. 

“In recent years, we have been able to take service to a new level – thanks in part to the use of artificial intelligence,” explains Marco Einacker, Lead Automation & DevOps Excellence at Deutsche Telekom Service. However, the team wants to continuously develop its processes to optimize service even further – for customers and the employees in the service team, who can then concentrate on more complex tasks. 

To achieve this, Deutsche Telekom Service relies on extensive automation. For example, customers who fill out a web form with their requests directly trigger an automated workflow that ensures fast processing. Smooth automatic processes are guaranteed by the Oreo process orchestration platform, which runs constantly in the background. “Automation using structured data from web forms is relatively easy to implement. But we ask ourselves: can customer issues be automated without structured data, for example from emails?”

The Automation & DevOps Excellence unit at Deutsche Telekom Service decided to take on this challenge and continuously automate the tricky parts of the service processes. The objective was to ease the workload of the service staff by reducing their manual tasks and increasing the efficiency of the service staff. To achieve this objective, the team looked at one of the main service channels for “unstructured” requests: incoming emails to addresses such as those handling fixed-line network orders. Telekom Service receives around 5,000 emails every week via this address. The approach was for artificial intelligence to automatically record the issue contained in the email, categorize the process, extract relevant data, and transfer it to Oreo to trigger an automated process flow. The managers approached the AI experts from T-Systems for assistance in designing the AI solution.

Large Language Models (LLMs) excel at processing and analyzing language, making them ideal for service automation. "We decided on a Python workflow based on Langgraph," explains Sebastian Wagner, AI Engineer at T-Systems. Using GPT deployed via Azure, the AI identifies email content, detects attachments (e.g., photos or PDFs), and extracts relevant data. However, understanding words alone isn’t enough—the challenge lies in interpreting customer intent. With 20 primary and 40 secondary categories for service requests, technical experts collaborated to define precise descriptions and map customer issues to internal processes, ensuring accurate categorization.

The resulting AI performs tasks sequentially: analyzing emails, extracting data, categorizing requests, and determining if they contain multiple issues. Based on its analysis, requests are either integrated into Oreo’s automation paths or flagged for manual processing. During testing, service employees annotated emails to create a dataset, resulting in the AI achieving 80% accuracy in identifying customer issues. To maintain quality, random checks and a dashboard monitor performance, while MS Teams alerts flag any deviations in key metrics.

With the new AI service, Deutsche Telekom Service further automated its processes. 80 percent of all incoming emails are automatically categorized and imported into existing automation. This significantly reduces employee workload and allows them to focus on other tasks. In addition, this approach achieves annual cost savings of over two million euros. “In the next few years, we expect even more potential once the service has been optimized.” For customers, this means faster processing – and ultimately, a significantly higher quality of service.