Our dual students have long become integral components of COViS's project operations. Along their journey, practical examples have formed the content of their thesis works. This series of articles provides insights into the daily activities of our emerging talents at COViS. As part of the COViS Salesforce team, Jonas Werner, a dual student, explored the opportunities presented by data science and machine learning in his master's thesis for current client projects.
Jonas Werner joined COViS in 2021 as a dual student in Salesforce consulting. Working on various projects with well-known clients, he primarily focused on areas of Salesforce consulting and development. In his master's thesis in 2022, he analyzed the potential of data science and machine learning for addressing issues and managing entire customer service processes. His documented approach and the insights gained, which demonstrate the benefits of AI in process optimization, serve as a blueprint for AI implementation in subsequent projects.
Customer Service with Machine Learning
Almost all IT products regularly require support from knowledgeable personnel. Depending on the issue, this support must sometimes be very fast or performed by specially trained individuals. Simultaneously, many support requests that can be handled with lower priority still pour into service centers. Accordingly, organizing and defining requests by priority and purpose is crucial for efficient processes within a service system.
The deployed IT systems must be reliable enough to correctly recognize the criticality of all service requests. To prevent errors and simultaneously keep costs low, there are various approaches to performing request classification with machine learning. "Process Mining" is particularly suitable for an AI-based analysis of the overall process. This technique uses event logs to visualize and analyze the performance of business processes, identifying inefficiencies, bottlenecks, and areas in need of improvement, ultimately helping companies optimize their processes and increase efficiency.
Systematic Analysis: Classification and Prioritization of Essential Optimization Factors
The critical aspects of setting up a machine learning process are understanding the business logic behind the processes to be optimized and comprehending the data used, its origin, trajectory, and characteristics. Only then can optimization potentials be effectively identified.
In customer service, throughput time emerged as a key factor for improving issue processes. Throughput time describes the period between the logging of an issue and its resolution. Ideally, this should always be defined to match the individual classification and prioritization of an issue. Too rapid processing, for instance, can lead to returns, creating longer processing times per issue, which diminishes customer satisfaction and increases overall effort.
The system studied was based on the Salesforce Service Cloud CRM platform. For the realistic implementation of process mining in an enterprise-scale customer service process, the test instance of a real system operated by COViS, which offers all functions and prerequisites of the production instance, was used. This ensured a realistic environment without jeopardizing operational processes.
In a preliminary analysis of the real data from the production environment of the system under consideration, it was recognized that the assignment of issues represents one of the most significant bottlenecks in the process flow. For assignment, an initial incident classification between service requests and disruptions must be made. Additionally, the content of the issue is specified, and a prioritization determined. The specification and priority then decide which department of Service Management is responsible for the issue. All this consumes many capacities and significantly extends the duration of the issue processing.
Weakness Analysis through Process Mining
To perform an evaluation, a data pool was created with records exported from Salesforce's Case-Activity object. Using this dataset, the Process-AI analysis was conducted with throughput time as the component to be examined.
During the process recording, the use of process mining successfully captured an image of the process and the respective throughput times of the issues. After data standardization, it became apparent: Particularly the customer assignment after issue creation and the creation of a work order after reviewing the liability issue at creation were the process steps that triggered delays.
Thus, the first steps towards process optimization were already clear: Automated classification and prioritization of issues based on content, a summary, and the initiator of the issue would not only be fundamentally possible but would also significantly accelerate the timing of service requests.
The approach to determining and evaluating throughput times was an exemplary and generically applicable analysis of the issue process present in the test system across all incoming channels. The approach was processed generically and recorded in documentation so that the resulting insights could be fully applied to further evaluations. This shortens the analysis phase in future projects where AI-supported process optimization is to be conducted.
Project Execution – Modelling
For the modeling, data mining, or machine learning was applied. Initially, the method and the algorithm to be applied were determined. It is common to test several evaluations, as the initial situation does not necessarily indicate the most accurate algorithm.
As a result of the modeling phase, a model was created per applied algorithm, which could be applied to the existing problem using supervised learning.
After testing various approaches in data modeling, Jonas found that the highest accuracy for the present problemAfter testing various approaches in data modeling, Jonas discovered that the highest accuracy for the current problem was achieved by using the k-Nearest Neighbor algorithm and neural networks.
Through this approach, the throughput times of the test system were successfully determined.
The Opportunities of AI in CRM-based Service Structures
Jonas's work demonstrated that both capturing and improving issue processes through Data Science and Machine Learning are possible. Although the work was conducted on test data from a client's system and the internal issues of Dr. Glinz COViS GmbH, a high focus was placed on documenting the approach generically for subsequent projects in the client environment. For a consulting firm like COViS, which aims to offer expertise in various areas of computing, this is an absolutely essential aspect.
Furthermore, this work proved that classification and prioritization of issues based on content, a summary, and the initiator of the issue are fundamentally possible. It is important to note that the results are strongly dependent on the quality and quantity of the available training data.
“Through my background in mechanical engineering, I have been dealing with the possibilities of Artificial Intelligence for several years. Already in my bachelor's studies, the sub-areas of Predictive Maintenance and classification problems were topics that particularly interested me.
I am very grateful to COViS for allowing me to deepen this knowledge during my two-year dual master's degree alongside operational project business and to expand it in many areas.
The speed at which Artificial Intelligence is developing is very impressive and widely known today. However, for the success of an AI project, in my opinion, it is essential to analyze the business case and the data situation in advance even more precisely.”
Jonas Werner, Developer at COViS
You can find more information about what COViS offers students and vacancies in our careers section.
