Global insurer: 80,000 claims/month, −70% processing time.

80,000 claims per month, 14-day cycle, operators still retyping policy numbers by hand. Document AI is no longer a PoC, it's debt if you don't do it.

The context

Global insurer, 28,000 employees, Paris and Tokyo hubs. Multi-country P&C business, 80,000 monthly claims (police reports, invoices, photos, medical certificates). Aging Tesseract OCR, high error rate, overloaded handling team.

The problem

• 80,000 claims/month processed mostly by hand
• Average cycle: 14 days, 6 of which are documentary waiting
• OCR precision: 71%, systematic human validation
• Estimated overcost: 14M EUR/year in manual operations
• Claims NPS: 28 (vs 52 internal target)

The approach

Three-stage document AI pipeline: layout-aware extraction (LayoutLMv3) for 76% of standard documents, multimodal LLM (Claude Sonnet) for complex or degraded cases, human validation on low-confidence exceptions. Case management workflow rebuilt to push the right cases to the right handlers.

The technical choices

• LayoutLMv3 fine-tuned on 18,000 annotated documents (reports, invoices, certificates)
• Claude Sonnet as fallback for degraded cases (photos, handwriting, multilingual)
• Routing via confidence scoring + document type
• Targeted human validation on 8% of cases (exceptions)
• Camunda workflow redesigned with field handlers

The stack

• LayoutLMv3 fine-tuned, SageMaker hosting
• Claude Sonnet via AWS Bedrock (multimodal)
• AWS Textract in pre-processing
• Camunda 8 for workflow orchestration
• FastAPI for internal APIs, PostgreSQL + S3

The results

1. Average claims cycle: 14 days to 4.2 days (-70%)
2. Extraction precision: 71% to 96.4%
3. Operational savings: 11M EUR/year validated at 12 months
4. Claims NPS: 28 to 47
5. Processed volume: +30% at constant headcount

What we learned

LayoutLMv3 fine-tuned on 18k documents beats Claude Sonnet on cost/perf for standard documents (76% of volume). But Claude Sonnet is unbeatable on Japanese handwriting and blurry photos. Mistake: we annotated documents before defining the final extraction schema, 2,000 docs to re-annotate. To redo: schema co-design with handlers from month 1, and pilot on 5,000 real claims before any rollout.