Bajaj Allianz General Insurance Company, one of the leading private insurers in India, has automated insurance processes using a machine learning (ML) service from AWS that automatically extracts text, handwriting and data from scanned documents.
The solution, Amazon Textract, helps Bajaj Allianz to reduce turnaround time and improve customer experience.
Bajaj Allianz General Insurance’s president of technology, Avinash Naik told the recent AWS Summit India: “We could significantly reduce waiting time for customers in issuing policies and claim settlements. We processed document extraction quickly with necessary accuracies needed for our business.”
Many processes like insurance claims processing, digitising the sales process, and finance invoice automation are highly dependent on extracting data from scanned documents in various formats.
“Textract can read and process any type of document accurately without any manual effort,” Naik said.
He added that Textract provided data in the most structured format for their teams to further perform analysis. “It has helped the firm reduce errors in manual tasks and also achieve 100 percent compliance,” he said.
“We were able to build processes for automatic approval and rejection of invoices based on document extraction.”
Challenges
Naik said their sales, claims processing and finance teams had faced several challenges in digitising data in various attributes.
“We handle many formats including handwritten annotations, dot-matrix printouts, medical reports, documents with watermarks, logos, stamps or scanned copies which may not be of desired quality. We wanted a solution that could extract data from unstructured documents with high accuracy and also be easy to train upon,” he said.
With a focus to be agile and error-proof in assessing and paying the claims, Bajaj Allianz has built an end-to-end intelligent document processing pipeline using Textract.
“The documents are first ingested either in real-time for single pages or into Amazon S3 for multi-page scan documents. Textract APIs are then invoked to extract all information from the documents and key-value pairs along with field-level confidence scores,” Naik said.
He added that the confidence score could then be used to determine if documents can be processed straight-through or after manual review.
Finally, the extracted data can be ingested in a data lake for deriving further insights or in a transaction processing system for process automation, Naik said.
Improved outcomes
“We probably have the best claim ratio and experience rating in the country,” he added.
Naik said that the firm had not only significantly reduced data entry time but also in processing health, and motor insurance invoices that come with different nomenclature.
As manual data entry is prone to errors, the company had to address more requests for policy changes.
“With the process now being digitised, the requests have started coming down. This was a win-win for the organisation,” he said.
After implementation, Naik warns organisations not to look for perfect accuracy.
“Don't expect 90 or 95 percent plus accuracy on day one. Models must be trained and accuracy improves only over time.”
He also added that his team recognised the need for partnership to help in pre and post-processing of information to achieve accuracy.
