Agentic AI in Action: End-to-End Automated Claims Processing
Together with Provinzial and andsafe, codecentric demonstrated in a production-ready Proof of Concept (PoC) how artificial intelligence can execute complex insurance processes autonomously, transparently, and thoroughly from end to end.
andsafe AG is a digital trailblazer in the German insurance market. Spun out of the Provinzial Group's new business division, the company operates as a purely online insurer without a physical branch network. Serving around 300,000 customers, andsafe delivers a streamlined digital product portfolio spanning private and commercial liability to bicycle insurance.
The project at a glance
- Innovative Approach: Development of an autonomous, agentic AI workflow for end-to-end processing of bicycle insurance claims.
- Efficient Structure: Division of complex regulatory requirements into seven logical processing stages for the AI agents.
- Outstanding Quality: An impressive 98% success rate for correct, fully automated straight-through processing.
- Enormous Leverage: Proven savings potential of 75% in claims settlement costs.
Background
andsafe AG ranks among the top bicycle insurers in the German market. When claims occur, fair, efficient, and swift settlement benefits customers, distribution partners, and insurers alike. For smaller claims, manual processing rarely makes business sense, as handling costs quickly rival the claim amount itself. Historically, these micro-claims were often settled automatically, while larger claims were handled either semi-automatically or manually.
The core guideline for claims processing is a nearly 100-page internal playbook packed with guidelines, General Insurance Conditions (AVB), and requirements from the German Insurance Contract Act (VVG)—all essential for compliant evaluation. For a traditional Large Language Model (LLM), ingesting such a massive context block all at once introduces significant risk of error.
To tackle this, andsafe teamed up with codecentric on a visionary project. The goal: validate through a real-world PoC whether modern agentic AI systems can handle these complex processes entirely autonomously, thoroughly, and compliantly through straight-through processing—relieving claims handlers and slashing costs.
Solution
Hybrid architecture
Specialized AI agents
Deterministic tools
Explainable AI
Hybrid architecture
codecentric's project team opted for a hybrid architecture to bypass the typical pitfalls of standalone AI models. Rather than letting the AI navigate the entire pipeline unguided, the overall flow was mapped deterministically. The 100-page rulebook was split across the seven logical phases of the actual claims workflow: from claim setup and formal/material coverage checks to the final settlement decision and claim closure.
Specialized AI agents
A specialized AI agent operates within each phase. Each agent receives only the subset of rules relevant to its specific step, analyzes the data using structured reasoning, taps into necessary core systems (e.g., for master data), and passes the output downstream via Structured Output.
Deterministic tools
Because language models are notoriously weak at mathematical calculations, codecentric integrated deterministic tools. For example, when verifying whether a bicycle exceeds the maximum age limit of 5 years, the AI steps aside: a precise code tool runs in the background to calculate the exact number of months between the purchase and claim dates.
Explainable AI
To guarantee maximum transparency and auditability, the AI was engineered to provide detailed reasoning for every single decision. Each output is automatically linked to the exact references within the AVB and VVG guidelines, allowing human auditors to track every step effortlessly.
Result
The PoC results spoke for themselves: in a stress test involving 60 claims (based on re-synthesized real data and complex edge cases), the AI system reached the correct decision in 59 cases—a 98% success rate. andsafe's internal claims handlers confirmed that the AI's autonomous decisions and justifications aligned fully with strict insurance guidelines.
An intriguing side effect also emerged: in practice, the AI was occasionally even more precise than human processors. Discrepancies between actual historical claims handling and the AI-settled cases were sometimes traced back to minor human oversight. In these instances, the AI agent proved more consistent and rule-compliant, actively sharpening overall process quality and policy adherence. Furthermore, the system proved its capability to interact seamlessly with external applications via mock interfaces (e.g., policy administration services).
The PoC was a resounding success and has already been presented internally as a flagship AI Success Story. Because codecentric designed the architecture to be completely modular and tech-agnostic, this approach can easily be scaled via modern workflow engines like Camunda 8 or n8n to other high-volume domains—such as substantially more complex auto claims processing.
Dr. Christian Brandt
Executive Board, andsafe AG
The PoC with codecentric impressively showcased the immense potential of agentic AI workflows. Our goal was to realistically prove whether AI can autonomously and reliably process complex, highly regulated workflows end-to-end. The 98% success rate across test cases far exceeded our expectations. codecentric delivered a massive leap forward for us, both technologically and strategically.
Dr. Christian Brandt
Executive Board, andsafe AG
Any questions about the project?
Looking for ways to integrate AI models into your existing system architecture seamlessly, transparently, and with full auditability? Feel free to reach out to me!
Further reference projects
Find out about other successful projects that we have completed with our clients. Perhaps you will find ideas here for a use case in your own organization.
Peter Westergerling