Smart Email Triage for More Efficient Customer Service

Every day, enersuisse receives more than 500 customer emails covering a wide variety of issues. The previous solution for categorizing and routing these emails had reached its limits, meaning that actually responding to them always required a significant amount of manual work.

The newly implemented email triage system fundamentally changes this. It steps in to provide support long before the message reaches the correct inbox. The system automatically handles intent recognition and customer identification, clearly summarizes the email content, and directly provides employees with a well-founded suggested response.

At the heart of the technical implementation was the goal of creating a future-proof, modular AI solution that integrates seamlessly into the existing system landscape. Databricks was chosen as the central platform for all AI-based developments. This decision offered key advantages:

• Existing expertise: The knowledge required for operation was already available internally at enersuisse.
• Versatile connectivity: Databricks enables a wide range of connections to various data sources and offers a flexible runtime for custom Python applications.
• Integrated observability: With MLflow, Databricks already provides a comprehensive AI & ML observability platform.
• Seamless monitoring: Existing and new dashboards can easily and directly access logs and observability data.

For the AI components of email triage, we built specialized endpoints within a service-oriented architecture (SOA) for categorization, optical character recognition (OCR), and structured information extraction. The large language models (LLMs) used are securely deployed via the existing Azure infrastructure with data located in Switzerland.

To optimally monitor the AI-based services, all traces are logged in MLflow, where they can be analyzed in detail in the event of an error. To ensure that the quality of the AI services remains consistently high—even with future enhancements or the introduction of new AI models—extensive evaluation datasets with ground-truth data have been established. Corresponding evaluation runs can be initiated at any time via dedicated endpoints, and the results are fully and transparently viewable in MLflow.

The implemented architecture enables enersuisse to operate the solution independently using its own internal expertise. It also provides the flexibility needed to implement any additional AI services and agents in the future, integrate them into the surrounding SAP systems, and utilize the AI functionalities from a wide variety of source systems.

Together with enersuisse, we evolved the existing automated email triage system into an intelligent assistance system.

The solution takes into account the subject line, the entire email thread, and content from attachments and documents. This information is captured using optical character recognition (OCR) and placed within the relevant business context.

Based on this data and supported by a Large Language Model (LLM), the system identifies the customer’s intent and intelligently forwards the email to the appropriate employees or the correct inbox. In addition, the system automatically generates a precise summary of the inquiry, which is displayed directly to service representatives as an “agent summary.”

This allows them to grasp the core issue at a glance without having to read the entire email. At the same time, the system generates more technically sound and better-structured response suggestions, which are derived directly from the company’s internal knowledge base—curated in Confluence—using Retrieval Augmented Generation (RAG), ensuring that responses are always based on validated sources and avoiding “hallucinations.”

Defined communication guidelines—such as tone, forms of address, or wording standards—can also be centrally integrated and consistently applied.

Each suggested response is additionally assigned a confidence score, which shows service representatives at a glance how certain the system is about a particular response. This forms the basis for a controlled, phased expansion of automation.

In addition, relevant customer data—such as business partner and contract account numbers—is automatically extracted from the email, then reconciled with and validated against the backend system. Where possible, the customer is uniquely identified and authenticated, and the corresponding information is retrieved in advance in SAP Customer Engagement (SAP CE). Combined with the suggested email responses, this significantly reduces processing time and enables service representatives to handle inquiries more efficiently and with greater focus.

For employees, this means fewer manual tasks, less effort spent sorting and forwarding emails, and a better foundation for technical review. The technology provides support in the background, while responsibility remains with the employees.

Over time, this further development will also be clearly noticeable to customers. Inquiries are reliably routed to the right department, and responses are of higher quality and follow a clear, consistent structure. The language, tone, and structure are consistently aligned with the defined communication guidelines. This creates a professional and recognizable service experience via the email channel that combines technical precision with a consistent brand image.

Building on the email triage system, additional processes will be automated in stages. Initially, employees will remain in the lead, reviewing, supplementing, or sending the suggested responses. As quality and experience improve, simple, clearly defined inquiries will be processed automatically in stages. In this way, the solution will evolve in a controlled manner from an assistance system to a scalable, automated email processing system.