Query claims in natural language with Amazon Bedrock Knowledge Bases
Amazon Web Services Shreya Pawaskar
AWS showed how Bedrock can answer claim questions from messy files with citations. It’s aimed at regulated work, where the source matters as much as the answer.
Based on reporting by Amazon Web Services, Shreya Pawaskar — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
AWS is pitching a claims assistant that can answer questions in plain English even when the facts are spread across adjuster notes, repair estimates, police reports, payment ledgers, and scanned attachments. The point is simple: in insurance, the answer usually isn’t in one neat database field. It’s buried across a pile of documents, and the person asking may need it fast.
The demo uses Amazon Bedrock Knowledge Bases, AWS’s managed retrieval-augmented generation setup for documents. Bedrock handles the ugly parts — parsing, chunking, embeddings, and vector storage — so the app can retrieve evidence and return cited answers from claim files. AWS says this is a technical walkthrough with synthetic claim records, not a production customer deployment.
The workflow starts with claim documents in Amazon S3, paired with metadata sidecars. The example metadata includes fields like claim ID, claim type, status, date filed, amount, region, adjuster, policyholder, and a few flags such as has_subrogation and has_litigation. AWS uses that metadata to narrow searches, so a user can ask for something like open auto claims over $10,000 filed last month, or a specific claim by ID. Dates are stored as YYYYMMDD integers so number-based filters can handle ranges.
Once the documents are ingested, the app calls AgenticRetrieveStream through the bedrock-agent-runtime client. That API plans the answer, breaks multi-part questions into sub-queries, can run multiple retrieval passes, and streams back trace events, answer text, and citations. AWS says the service can also apply a contextual grounding check so unsupported answers get blocked before they go out. In regulated work, that’s the whole ball game.
The example code also shows multi-turn follow-up questions, metadata filters, and citation rendering that maps answer spans back to source documents. The design is built for claims teams: policyholders asking for status, contact center agents trying not to put someone on hold forever, and adjusters comparing conflicting versions of the same record. It’s not flashy. It’s just the kind of plumbing that makes an AI answer usable instead of merely eloquent.
My take — AI-written commentary, not fact-checked reporting
This is the right kind of AI demo: boring, audited, and chained to source documents. Claims work does not need a poetic chatbot; it needs something that can stop hallucinating the minute a revised estimate shows up. The broader pattern is clear: the winners in enterprise AI will be the systems that admit they need receipts.
Read more about this at: Amazon Web Services