Initial situation
A leading IT service provider in the financial sector identified a strategic opportunity to make the processing of financing documents significantly more efficient and flexible. Previously used classification and dedicated extraction models only supported a fixed set of document classes in real estate financing. Every functional expansion required a complete and resource-intensive machine learning cycle: from extensive data collection and manual labeling to complex model training and subsequent evaluation.
Since the specific domain knowledge regarding the diverse document types primarily resided within the specialized processes of the affiliated financial institutions, there was an urgent need for a solution that would eliminate this technical dependency. The goal was to develop a highly flexible service that would allow internal development teams and the financial institutions' business units to integrate new document types independently. This was intended to significantly optimize responsiveness to regulatory or procedural changes in the highly regulated financial sector.
Our solution
To meet this strategic requirement, we designed and developed an API-based AI service for flexible zero-shot classification and information extraction. The system enables the processing of any document type while ensuring the highest compliance and data protection standards. The technical implementation includes the following core components:
- Zero-shot classification of financial documents: The developed service uses advanced large language models to categorize incoming documents immediately without prior specific model training. The caller simply passes a list of possible document classes to the interface, and the system then accurately assigns the document. This innovative approach eliminates the previous bottleneck of time-consuming model training and allows for immediate adaptation to new document types in the financial industry.
- Dynamic information extraction with structured output: Business teams can define complex extraction schemas themselves with complete flexibility via the API. The system converts these requirements into strict data schemas at runtime and transmits them to cloud-based large language models. This technology ensures that the extracted information is returned in a schema-compliant manner in the exact desired data format, such as standardized text, validated numbers, or predefined selection lists. This ensures seamless technical downstream processing.
- Automated document preparation and post-processing: Since financing documents are often provided as unstructured scans or image streams in practice, the solution includes upstream document preparation. Automated orientation correction aligns skewed scanned pages and improves the basis for subsequent text recognition. During post-processing, extracted date formats and other business-relevant entities are normalized. For multi-page documents, an aggregation logic consolidates the page results and adopts the most frequently recognized value for each entity.
- Production-ready service architecture for enterprise use: The high-performance service was implemented according to cloud-native principles and runs in a self-managed data center environment that meets the strict regulatory requirements of the financial industry. Modern dashboards and centralized log management guarantee complete traceability and stability of the operational system. An integrated AI gateway efficiently manages and optimizes requests to the language models, while secure secrets management guarantees the continuous confidentiality of highly sensitive customer data.
- Scalable deployment and evaluation approach: To enable broad adoption, the API service was deployed as a containerized application on modern platforms. Since the zero-shot method cannot inherently provide specific guarantees for new, unknown documents, a methodical evaluation approach was implemented. This allows users to test their individual document types quickly and securely based on their own reference data before they are integrated into productive business processes.
Results & business impact
The developed AI service is sustainably transforming the way affiliated financial institutions expand and adapt their document processes. By successfully decoupling from rigid training cycles, the number of technically supported document classes is now effectively unlimited. Business teams are now able to define new extraction requirements completely independently and use them in production after a short evaluation, which significantly shortens the time-to-market for adapted or entirely new business processes.
The strategic shift of specific domain knowledge directly into the business departments enables highly scalable and maximally flexible document processing in real estate financing. The powerful service is available for immediate integration by internal teams and provides a robust technological foundation to sustainably reduce manual effort in administrative tasks while ensuring the highest quality standards in automated data extraction.
