Agentic & Applied AIAllianz Partners · Home Digital Solutions · 2025–present

5 AI modules live in production claims processing

The case in 30 seconds

Problem: Home insurance claims arrive as messy multi-document PDFs — invoices, quotes, photos, attachments mixed together. Every document meant manual reading, checking, and pricing before a claim could move, and fraud patterns hid in the volume.
What I did: Led the department running an AI document-processing portfolio into production across France and Germany. Modules split and classify documents, extract invoice lines, flag anomalies, screen fraud, benchmark costs, and classify photo damage. Delivered the German modules; now leading the rebuild into a next-generation multi-tenant platform.
Impact: 5 AI modules live in production across 2 markets, screening every incoming claim document. Further modules are in proof of concept and scoping for 2027.

How it happened

2025 — Where it started
Claims documents were the bottleneck: heterogeneous PDFs, hand-read invoices, and no systematic fraud screen. The question was which steps of the document journey AI could reliably take over in production, not in a demo.
The architecture decision
A pipeline of specialised modules rather than one model. Splitting, extraction, classification, red-flag analysis, forensics, and pricing run as separate components with JSON contracts between them. Every automated decision carries its reasoning, so an adjuster can audit it: confidence scores, flag details, and an explanation.
The pipeline in production
1. Multi-document PDFs are decomposed and classified automatically — invoices, quotes, photos, and attachments separated for the right downstream module.
2. Invoice line items are extracted and mapped to trades. 10 layered red-flag checks plus a machine learning (ML) approval score end in an explained pass or fail.
3. Fraud screening combines duplicate detection within and across claims, document-manipulation forensics, and image-similarity analysis.
4. Cost reasonableness is benchmarked against internal pricing references, giving adjusters a fair-price view per claim.
5. Computer vision classifies wall and ceiling damage from customer photos, each detection with a confidence score.
How adoption was driven
Market by market, module by module. Each capability proved itself in production before the next was added; the German modules were delivered under my leadership. Operations act as the human reviewers the outputs serve. Production AI on claims data ships through governance, data protection and security included — not around it.
Today
The portfolio's production learnings now define the requirements for the next-generation multi-tenant rebuild.
5
AI modules live in production
Claims-document processing, France and Germany
2
markets running in production
France and Germany
10
layered red-flag checks per invoice
Plus a machine-learning approval score

What I took from it

AI in production is a portfolio, not a model. The value is in the pipeline — and in every decision being explainable to the human who signs off.