Initial Situation
A company in the financial sector recognized the strategic opportunity to evolve the processing of incoming customer inquiries using modern natural language processing techniques. Previously, messages had to be manually reviewed, categorized by topic, and forwarded to the relevant specialists. This process tied up valuable resources and made consistent processing difficult as volumes increased.
The goal was to find a solution that could reliably identify various types of requests while meeting the specific data protection and traceability requirements of the financial sector. Beyond simple topic classification, the system needed to extract relevant information from messages, protect sensitive data, and make additional features such as sentiment or attachments available for further processing.
Our Solution
AMAI developed a system for automated email classification and information extraction based on modern natural language processing methods. The solution combines specialized transformer models with a structured data and deployment logic. This ensures that technical requirements, data protection, and reliable operation in the financial sector are all addressed.
- Multi-stage topic classification: Pre-trained and fine-tuned transformer models first assign incoming customer emails to broad topics and then to more detailed categories. This hierarchical classification reflects the technical structure of the inquiries and enables more precise routing than a simple assignment to a few general categories. As a result, staff receive messages with the appropriate context for further processing.
- Information extraction and anonymization: Named entity recognition techniques identify sensitive information within messages, making it manageable for downstream processing steps. Personal or other sensitive data can be automatically anonymized. This ensures data protection is integrated into the processing workflow without losing the content relevant for classification and routing.
- Structured training data: To further develop the models, suitable training data was compiled and annotated using a revised labeling concept. This structured labeling creates a more reliable foundation for text classification and information extraction. Using Label Studio , annotation processes were organized transparently, and new technical categories were systematically incorporated into the model development.
- Advanced features for processing: In addition to the main topic of a message, the solution considers extra information relevant for prioritization and routing. This includes sentiment analysis and the automatic classification of attachments. These supplementary features help capture customer requests more comprehensively and provide more targeted support for subsequent steps in financial services.
- Production-ready model deployment: The trained models are deployed via the NVIDIA Triton Inference Server provided. A central inference service enables uniform model usage and supports the efficient processing of incoming messages. At the same time, the technical structure creates a foundation for updating models in a controlled manner and running various classification and extraction tasks within a single, shared solution.
A particular challenge lay in balancing data quality, data protection, and technical accuracy. Training data had to be carefully vetted and, where necessary, rebuilt. Simultaneously, different approaches for classic transformer models and modern generative language models were examined to evaluate their suitability for hierarchical multi-label classification under realistic conditions. The solution was designed to ensure that technical advancement and regulatory requirements remain aligned.
Results & Business Impact
The developed solution was tested in practical, real-world financial industry environments and refined for broad deployment. Incoming messages can be automatically classified, enriched with relevant information, and forwarded to the appropriate departments. The combination of topic classification, information extraction, and anonymization creates an end-to-end process for the intelligent processing of customer communication.
For the participating financial organizations, this creates a better foundation for consistent and more efficient handling of customer inquiries. Manual sorting and forwarding steps are reduced, while sensitive content can be processed with greater control. The integrated analysis of sentiment and attachments enhances the depth of the classification and supports more targeted responses to various requests.
Technically, the project also establishes a robust basis for the further expansion of AI applications in regulated environments. The experience gained in data preparation, model comparison, anonymization, and production-ready deployment can be applied to other document and message processing use cases.
