About the Company
The client is a commercial property and casualty insurer that underwrites coverage for businesses across a range of industries. Its underwriters work from a large and frequently updated body of guidance, and the speed and consistency of their decisions depend on getting to the right rule at the right moment.
The Challenge
Underwriters were spending too much of their day searching. The guidance they rely on, underwriting manuals, appetite guides, endorsement language, and the bulletins that revise all of it, was spread across hundreds of documents, and the rules changed often enough that even experienced staff were not always sure they were working from the current version. New hires took a long time to come up to speed, and different people sometimes read the same guidance differently. The company wanted to make all of this answerable in plain language, but with one firm condition: an answer had to come from the company's own approved documents and had to be traceable back to them. An underwriting decision could not rest on something a model had invented.
The Solution
NileForge built the assistant as a retrieval-augmented system on Amazon Bedrock, so that answers are drawn from the insurer's own material rather than the model's general training. The underwriting documents sit in Amazon S3, and Knowledge Bases for Amazon Bedrock handles the retrieval side of the work: it breaks each document into passages, turns them into vector embeddings using Amazon Titan Text Embeddings, and stores them in a vector index on Amazon OpenSearch Serverless. When an underwriter asks a question, the service finds the passages that match and passes them to the model, which writes its answer from that material and returns the source sections alongside it, so the underwriter can open the original wording and check it. Rather than build and run a retrieval stack from scratch, NileForge used the managed service here to keep the moving parts, and the long-term maintenance, to a minimum.
Keeping answers honest mattered as much as producing them. NileForge applied Guardrails for Amazon Bedrock, including contextual grounding checks that catch responses the retrieved passages do not actually support, so when the guidance does not cover a situation the assistant says so instead of guessing. Because the underwriting rules are revised regularly, an automated ingestion job refreshes the knowledge base whenever documents in Amazon S3 change, keeping it in step with the latest approved versions rather than drifting out of date.
The assistant reaches underwriters through a web application with an AWS Lambda and Amazon API Gateway backend, with sign-in handled by Amazon Cognito so only authorized staff have access. Data is encrypted with AWS Key Management Service, calls to Amazon Bedrock stay on the company's network over an interface VPC endpoint, and permissions follow least-privilege through AWS Identity and Access Management. Amazon CloudWatch and AWS CloudTrail provide monitoring and a record of how the tool is used, and the whole environment is defined as code in Terraform.
The Results
- Underwriters get answers in plain language, drawn from the company's own approved guidance
- Each answer points back to the source sections, so a decision can be checked rather than taken on faith
- Grounding checks stop the assistant from answering beyond what the guidance actually says
- New underwriters reach productivity sooner, working from one consistent body of guidance
- The knowledge base updates itself as manuals and bulletins change, so it stays current
Put your own knowledge to work.
NileForge builds generative AI on AWS that answers in plain language, grounded in your documents and cited so your team can rely on it. If that is the problem in front of you, we should talk. Talk to our team
