How does Helvetia Group speed up document processing with GenAI?
- Data & AI
The business problem
The cost of manual document handling
As in other industries, the daily work of many employees in an insurance company involves extracting information from documents and to manually enter this data into a system, thereby making it digitally accessible. This integration enables the company to incorporate the information into its data landscape, allowing machine processing and storage for various purposes. While many document types can nowadays be digitally processed, scans and photos often still require manual processing – particularly when they vary in structure and layout. This takes time:
Depending on the length of the document, this step can take anywhere from a few seconds to several minutes. Finding information on insurance related documents is quickly done, when they are well structured, but becomes harder when 120-page long reports must be seen through.
The time required to type the extracted data into some program depends on the amount of information and the user-friendliness of the system. Typically, this step takes seconds to minutes.
The input, particularly for complex or numerous data fields, is cross-checked with the original document. Sometimes, there is also an automatic cross-check with internal systems, e.g., to match the customer name from the document with a customer ID from an internal database. Depending on complexity, this again takes seconds to minutes.
Overall, the total time for manual document processing ranges from a few minutes to several hours (think, e.g., about the complexity of documents that need to be reviewed for underwriting corporate customers in specialized business areas).
At Helvetia Group (open in a new tab), too, extracting information from documents involves significant manual effort. As a pioneer in adopting cutting-edge technologies, Helvetia Group realized as early as 2023 that generative artificial intelligence (GenAI), particularly large language models (LLMs), could offer a solution to increase efficiency in this area. As part of Helvetia Group’s “Fit4AI” program, the idea for an in-house development of a smart application to extract document information quickly emerged. Helvetia Group brought Wavestone on board as a development partner, having already collaborated successfully and innovatively on several AI projects.
Our solution and approach
Although the implementation project was based on a specific use case, it was clear to us that various departments within Helvetia Group perform the same manual steps. While the content and document types may vary widely, the general approach remains the same. Therefore, instead of creating a tool for just one use case, we designed the application to be easily scalable to new use cases, ensuring the investment in development would pay off multiple times.
The solution is built in a modular way. Its general core includes an OCR (Optical Character Recognition) component and an LLM module. Each use case is represented by a single configuration file, which defines the desired extraction content. OCR is an AI method using a vision model to recognize text in images and convert it into text form. This text, along with the use-case-specific requirements, is sent to an LLM, which extracts the desired information. The result is output in a structured format, making it ready for direct machine processing if needed.
A scalable, flexible, and secure AI solution
In addition to easy scalability across diverse use cases, another crucial design principle was flexibility regarding the AI models and service providers used. Generative AI is evolving at an incredible pace, with new and better models being released almost weekly by various providers. Therefore, it was clear that a sustainable solution must allow the integration of new models. This flexibility is even available at the request level, since, depending on the use case, one model may produce better results than another. Direct queries to multimodal models are also possible. Multimodal models can directly process images, such as scans and photos, to extract information from them.
The flexibility also extends to the provision of AI models. Models can be accessed via different hyperscaler services (e.g., Azure, AWS, Google, etc.). Currently, Azure services are used, but the solution is designed to allow easy migration to other providers.
The in-house development brings not only flexibility but also security. The application operates within Helvetia Group’s infrastructure, and neither query contents nor processed documents leave the “Helvetia world” thanks to Azure’s cloud offerings.
InfoXtractor enables us to quickly interpret millions of digitized documents and will significantly simplify many processes.
Results
Reducing processing time from minutes to seconds, hours to minutes
The new GenAI solution replaces manual steps, delivering results within seconds or minutes, depending on the complexity of the use case and the size of the documents. Thanks to the structured output format, the extracted data can be directly processed by machines or presented to an employee for review in an application interface.
As expected and already while the solution was still under development, more and more teams from various Helvetia Group departments approached the responsible project team with potential use cases. Soon, over 20 use cases were in the pipeline, with different teams evaluating the feasibility of the solution, now known internally at Helvetia as InfoXtractor. The solution has since gone live showing promising results, with high and still growing demand.