Case-Study: Zurich Versicherung

Structured capture of printed legacy contracts

At many companies, documents pile up that conventional OCR technologies can do little with, making manual review and reconciliation processes necessary. AI is more than up to this challenge.

Customer
Industry
Insurance
Service
Software Development
Time
1 Monat
Team
1 Developer
Initial Situation

Thousands of legacy paper contracts need to be captured in a structured format

Zurich Group Germany is developing a new application for rating and quote generation, in which certain contract conditions are to be automatically checked for plausibility. For this, the underlying contract data must be available in a structured, machine-readable form and fit into a complex, domain-specified target schema.

This was not the case for 2,500 legacy contracts: they existed only in paper form. The obvious solution — having clerks capture these contracts manually — would have been costly, error-prone and time-consuming, and would have tied up significant capacity in the business unit for months.

Thousands of legacy paper contracts need to be captured in a structured format
Tech Stack
Claude
AWS Bedrock
Vercel AI Toolkit
TypeScript
Our Aproach

KI-Workflow als Extraktions-Pipeline

We developed an AI-powered pipeline that automatically extracts data from the scanned paper contracts and transforms it into the predefined target schema. The multimodal model used — Claude 4, integrated via AWS Bedrock — understands not just the raw text, but also the layout and domain context. That is precisely what's required to reliably map highly heterogeneous legacy contracts onto a uniform schema.

The structured results are then fed into a human-in-the-loop workflow, where clerks review and approve them in a user interface.

In this project, AI is at the core of the solution. The AI model isn't an add-on but a central component of the workflow: it performs the actual translation from paper contract to the structured target schema, and is what makes the lean human-in-the-loop review possible in the first place — rather than full manual data entry.

This combines the speed of the model with the domain-level reliability that is non-negotiable in an insurance context.

The Result

Dusty legacy contracts become a digital data source

The 2,500 legacy contracts are now available in the target schema in structured form — the application can access the complete dataset, and the business unit saved months of data-entry work.

Manual data entry eliminated:
Case handlers no longer transcribe contracts by hand; they only confirm the AI's structured output. This cuts effort and simultaneously reduces the error rate.
Time savings:
What would have become a multi-month data-entry project with substantial staffing was delivered in one month by a single software developer.
Resources freed up for customer service:
Instead of going into monotonous data entry, the time saved flows into complex cases where professional judgment really matters.

New paths are meant to be explored.

Tell us about your project so we can take the next step together.

Similar Projects