All case studies
Agentic & Applied AIAllianz Partners · Home Digital Solutions · 2025–present
5 AI modules live in production claims processing
- 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.
- Action
- 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.
5AI modules live in production2markets running in production10layered red-flag checks per invoice
How the engagement ran
- 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 a demo.
- Who I worked with
- The product and engineering teams in my department, and operations as the human reviewers the outputs serve. Business partners in both markets, plus data protection and security functions. Production AI on claims data ships through governance, not around it.
- 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.
- 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. The portfolio's production learnings now define the requirements for the next-generation multi-tenant rebuild.
How it worked
- Multi-document PDFs are decomposed and classified automatically — invoices, quotes, photos, and attachments separated for the right downstream module.
- 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.
- Fraud screening combines duplicate detection within and across claims, document-manipulation forensics, and image-similarity analysis.
- Cost reasonableness is benchmarked against internal pricing references, giving adjusters a fair-price view per claim.
- Computer vision classifies wall and ceiling damage from customer photos, each detection with a confidence 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.