Systems | Development | Analytics | API | Testing

We thought the lemon had no more juice. AI proved us wrong.

If there's one thing that stuck with me from Yannick Misteli's session at the Agentic Analytics Playbook event in London, it's this: most AI pilots don't stall because of technology or budget. They stall because nobody answered the "day after" questions. I had the opportunity to sit down with Yannick Mistelli, Head of Engineering at Roche, the global pharma company with 100,000+ employees and heavy regulation across 25+ countries.

AI claims automation is becoming a major focus for insurers as manual claims processing continues to create cost, speed, and customer-experience challenges. Traditional systems often route documents through multiple human reviewers, creating bottlenecks that consume time and operational resources.

Two years ago, GenAI in production usually meant a single LLM serving a single endpoint. In 2026, it usually means much more. The applications shipping in front of users today are compound AI systems: orchestrated pipelines of retrievers, embedders, dialogue models, classifiers, code interpreters, SQL executors, and tools, with a single user request fanning out to several model calls across the stack.

Imagine telling an AI assistant: “Find me waterproof trail shoes, size 10, under €140, delivered before Friday. Use my usual card if the total stays within budget.”

Regulated QA teams carry a pressure most testing platforms weren’t built to solve for at the same time. Every release still needs traceability from requirement to test case to defect that holds up under audit. Approvals and evidence still need to be airtight. At the same time, agile releases, DevOps pipelines, and AI-assisted development keep moving, whether or not the testing platform underneath has kept pace.