Simpego, a digital-first Swiss property and casualty insurer, automated the extraction of data from claim documents, mostly invoices, which its claims specialists previously transcribed by hand, line item by line item, including complex Swiss VAT details, across thousands of documents a month. Template-based OCR could not cope: Swiss invoices mix German, French and Italian, carry several VAT rates (8.1%, 2.6% and 0%) on one page, and arrive as anything from clean PDFs to blurry phone photos and handwritten notes.
With AWS partner copebit, Simpego built a fully serverless, event-driven pipeline that keeps ingestion separate from model inference so the claims platform never blocks. Its core insurance platform (SIP) calls AWS AppSync over GraphQL to obtain a presigned Amazon S3 URL, and the document is uploaded directly to S3. An S3 event triggers a Lambda function that converts the document to a model-ready format and assembles a prompt containing the document, a strict JSON output schema, field definitions and multilingual hints for Swiss invoice conventions, then calls Anthropic's Claude Opus on Amazon Bedrock. The response is parsed and validated, then pushed back to SIP through an AppSync mutation, and specialists see pre-filled fields appear in real time via WebSocket subscriptions.
Validation has three layers: automated schema validation so downstream systems never receive malformed data; arithmetic checks that line items and VAT amounts sum to the invoice total; and a final human check in which the claims specialist corrects any errors, which also feeds back into prompt improvements. Bedrock was chosen for its Zurich region (eu-central-2) data residency, its policy of not training on customer data, and paying only per inference call for an irregular workload. The pipeline sits within Simpego's multi-account AWS landing zone, alongside Amazon EKS, Aurora Serverless MySQL, CloudWatch, Glue and Athena.
"Automation is only valuable if the output is trustworthy. By pre-filling fields for our specialists, we create a continuous feedback loop that improves our prompts while ensuring our team remains the final decision-makers," said Michael Ammann, CTO of Simpego.
The project went from kickoff to production in two months. The pipeline now processes thousands of extractions a month in roughly 30 seconds per document, and 95% of cases need no manual intervention.
Details
- City
- Zurich
- Organization