Medical Expert AI
End-to-end redesign of a claims processing system serving 90% of the German insurance market — introducing LLM-powered automation for 90+ customers and surfacing AI recommendations in a way that was intuitive, governable, and trusted by medical and legal experts.
Overview
A full system redesign powered by AI
The integration of LLMs and Intelligent Document Processing (IDP) presented a rare opportunity — not just to improve the existing product, but to redesign and migrate the entire system from scratch into a new use case-based automation pipeline.
This required designing end-to-end user flows grounded in deep domain knowledge: understanding how insurance claims were originally processed by legal experts, medical professionals, and claims assessors — and unifying these into a single coherent, scalable system.
Impact
Measurable gains across the workflow
Beyond the numbers, the system enabled quicker and more confident decision-making during claims assessments — reducing cognitive load for experts by surfacing the right information at the right moment, rather than requiring them to hunt through unstructured documents.
The Opportunity
Automation where it matters most
Manual claims processing was time-intensive by nature — tasks that took hours were bottlenecked by repetitive document review, data extraction, and structured report writing. The emergence of LLM and IDP technology created a direct opportunity to automate the most labour-intensive parts of this workflow.
The challenge wasn't just building automation — it was designing it to be understandable, trustworthy, and adoptable by experts who had spent years doing this work manually.
Understanding where automation opportunities existed required granular knowledge of how different expert types — legal, medical, administrative — approached claims at each stage. This domain research informed every design decision.
Research
90% of the German insurance market. One unified system.
Verisk's platform serves approximately 90% of the German insurance market — meaning this redesign had to work across the full breadth of the industry. Working directly with over 90 insurance customers, I mapped the workflows, edge cases, and expert mental models that the new system needed to accommodate. This wasn't a one-size-fits-all problem — different claim types, user roles, and legal contexts required flexibility at the component and flow level.
The research phase focused on identifying where manual work could be meaningfully reduced without eroding the expert's sense of control or oversight — a critical factor for AI adoption in high-stakes environments like medical injury claims.
- End-to-end user flow mapping across legal, medical, and administrative expert types
- Identification of high-friction, high-volume manual tasks suitable for automation
- AI governance requirements: when and how AI recommendations should be surfaced
- Component-level flexibility requirements across 90+ customer use cases
User interviews were combined with business flow diagrams to bridge the gap between real user behaviour and underlying business logic. This dual lens — qualitative research alongside process mapping — allowed me to build an initial proof-of-concept grounded in both user needs and operational constraints.
I championed the project internally, presenting the proof-of-concept to stakeholders and driving approval to proceed — translating complex UX rationale into business value to secure buy-in across the organisation.
Design
Surfacing AI in a way experts could trust
The core design challenge was making AI recommendations legible and actionable — not just technically functional. Insurance claims involve legal liability and medical judgment, so the system needed to present AI outputs with appropriate confidence framing, traceability, and override capability.
I led the design and development of the AI recommendation layer, working through questions of AI governance: when should the system act autonomously, when should it recommend, and when should it stay silent?
- Scalable generic component system adaptable across claim types and customer configurations
- New front-facing UI that simplified manual workflows for faster adoption
Components were designed to be scalable and generic — adaptable to many use cases through industry knowledge and design expertise, making the system flexible yet targeted for quicker claims processing.
I designed and scaled the design system from the ground up — establishing standards for the broader agentic AI product family at Verisk. The work was featured by Verisk's CTO at the Messe Conference.
Design Tradeoffs
As a Principal UX Designer, I had to balance trade-offs across user experience, business goals, technical constraints, and delivery timelines — especially for the MVP scope.
This meant prioritizing core user flows, clarity, and speed-to-market over edge cases, advanced customization, or long-term scalability in the first release. I also worked closely with stakeholders to align product vision with engineering feasibility and business priorities while managing scope and maintaining momentum.
NDA prevents naming the specifics. Here's how I categorize the judgment calls behind this work:
Systems that scale
- Minimized components by mapping needs across business units. Designed primitives that adapt across contexts instead of building per-business variants. Reduced library entropy at the source.
- Modeled the insurance E2E to predict downstream scale. Used claim-lifecycle domain knowledge to anticipate which components would absorb new states, integrations, and business rules over the next 2 to 3 years.
- Reconciled overlapping business asks into a multi-year-adaptable IA. Structured the layout so new lines of business or AI surfaces could land without a re-platforming pass.
AI with accountability
- Used PM/PO roadmaps as inputs to AI integration mapping. Identified high-leverage AI points from the established business flows, not retrofitted into screens after the fact.
Shipping with judgment
- Pushed back on data-modeling choices optimized for engineering shortcuts. Argued for product-first data foundations even when the easier path was tempting. Wrong shortcuts compound over years.
- Knew what would ship vs. balloon. HTML/CSS/JS and front-end library competency informed component decisions — including page-load and render-performance implications. The "should we build this elaborately?" question came up early, not at code review.
- Probed before designing. Asked: what's the goal of v1? What can we leverage from what already exists? What's the simplest, most intuitive path for this user group? We don't reinvent the wheel — we see what users actually need to do, then make the way.
Retrospective
Next time, I'd cut irrelevant functionalities sooner and focus solely on championing those that would clearly make an impact. Simplification of old systems can spark opportunity — UX heuristics apply, and good ones can win over old habits of working.