AI Support for Service Desk Agents

For many growing IT companies, managing an ever-increasing flood of customer requests can put a heavy burden on the all teams involved, but especially on the service desk agents that are at the forefront of the process. In this project, we helped rku.it GmbH, a leading service provider for the utility and transport sector, to identify and realize automation potential with AI by summarizing incoming requests and making ticket assignment suggestions based on historical cases.

The goal was to help the service desk department “in new and innovative ways, but getting real impact as soon as possible”. As this is exactly how we like to work, we got right to it. In this article, you can learn how we worked in this proof-of-concept phase together with our client, what we achieved so far, what some of the learning were and how the project will continue in the near future.

Client Industry

IT Service, Energy, Transport

Year

2024

Technologies

LLMs and GenAI, Llama3
Pandas, XGBoost
Langfuse, Langchain
DistilBERT
ITSM Application

The Challenge
Agents need to correctly handle complex requests under pressure.

The client for this project provides solutions for the utility and transport sector. rku.it manages a service desk that receives thousands of emails, calls, and tickets monthly, which in itself is a challenge. The service desk is tasked with reading each ticket, which can be quite technical, and understanding its content. The agent then assigns the ticket to the appropriate team for resolution. Understanding and correctly handling the tickets can take significant time because the client manages a complex product landscape. Incorrect assignments can lead to further delays and wasted resources.

Our Approach
Identifying Quick Wins, But Keeping the Big Picture in Mind

Before diving into technical solutions, we worked with rku.it to identify the major problems and corresponding solutions from the operational to the strategic level. It became clear that the employees in the service desk would like support in understanding long and complex request threads as well as assigning tickets to the right team. We decided to work on a proof-of-concept to show how well AI-based solutions can work for summarizing and classifying email-based requests and try both LLM and conventional machine learning approaches. This is well-aligned with Dockside Data’s core philosophy of providing verifiable results during the innovation process and always putting the client’s needs before any technology selection.

The Results
Automatically Summarize Requests and Predict Responsibilities in Existing Tools

Staying in close contact with the users, we decided that any service can only be useful if it fits into the existing tooling landscape and therefore focused on developing integrated prototypes. Dockside Data implemented a service using Large Language Models (LLMs) to automatically summarize each email and attach the summary to the ticket within the ticket system. Additionally, a classification system using an XGBoost classifier made suggestions for the correct responsible team, thus significantly reducing mental overhead for assignments. Both components were integrated into the client’s main ITSM tool to show the agents the information exactly where it is needed.

After three weeks, we were ready to roll out the prototype to allow some service desk agents to test it in limited operation and generate some usage data for us to see the effect of the improvements. While the information extraction and summary showed time savings of up to 75% for longer requests, the automatic assignments had the most potential for helping the agent on the average case, as it also allowed colleagues with less experience to better assign tickets to responsible teams. The users praised the integrated way of presenting information and the ease of use of the new features that integrated seamlessly into their existing toolset. But they also mentioned room for improvement. For example, the prototype did not include information that was not provided in text form, such as screenshots. Additionally, the prediction accuracy needs to be improved for some use cases.

It is not surprising that after only a month, not all the boxes were already checked, but the proof-of-concept showed the potential for AI-based solutions. We are happy to say that rku.it reached out to us shortly after finishing up, and we are currently in talks for continuing the collaboration and achieving production readiness.