The starting point
Insurance brokers work in a highly complex environment with strict regulatory requirements. In Switzerland, data protection, security and compliance aren't optional; they're mandatory. Managing insurance contracts for business clients and public clients such as municipalities takes a lot of work, and so do comparing offers and handling claims.
VERSIFY wanted to digitize these processes completely. Together, we built a platform where clients manage their contracts online, request new offers and compare quotes. AI is used in several places: in a chat assistant that guides clients through filing a claim, in the structured analysis of insurance offers, and in correspondence with insurers.
format_quoteAdding an AI feature is quick. Running AI at business level is a different job.
Stefan Hess
Three requirements at once
Adding an AI feature to software is quick these days. At VERSIFY, though, three requirements had to be met at the same time, and it's their interplay that makes the task demanding.
First, the language models have to be interchangeable at any time, whether they come from OpenAI, Anthropic or Google. The market moves fast, prices and capabilities change constantly, and a platform tied to a single provider loses that freedom.
Second, the processes have to deliver precise results, reliably. An answer that sounds plausible isn't enough when coverage amounts and policy terms are at stake. The result has to be well-founded, and just as well-founded the next time.
Third, the data has to be processed under Swiss data protection law. That limits the choice of models, and with it the first requirement: only models that meet this condition are interchangeable.
What this means for the architecture
These requirements can't be added after the fact. They have to be built into the architecture from the start. That's why data protection and compliance were part of VERSIFY's architecture from day one, not an item on a checklist before go-live.
In concrete terms, interchangeability means the business logic doesn't talk directly to a specific provider. Which model handles a task is a matter of configuration, not a rebuild. Reliable precision means the AI delivers structured results that can be checked, rather than free text you have to take on trust. And data protection means it's clear for every model where and under what conditions it processes the data.
What we take away
The MVP was built in a few months, and the first pilot customers are online. They especially value the analysis features and the noticeably lower effort in case management. Since then, we've been developing the platform further based on real user experience.
The most important lesson isn't a technical one. The difference between “quickly adding an AI feature” and “running AI at business level” lies in the questions you ask before writing the first line of code: What happens if the provider changes its prices? How do we know a result is wrong? Where does the data live? If you answer these questions early, you're building AI-native. If you have to answer them later, you're rebuilding.
