The starting point
SWICA is a multilingual organisation by necessity. As one of Switzerland's largest health insurers, every piece of communication (marketing, legal, product documentation) needs to work in German, French, Italian, and English. A small but highly skilled Language Service team of professional translators handles this across all content types.
For roughly two decades, the team's primary tool was Multitrans. It had served its purpose, but by 2024 it was showing its age in every dimension. The technology was outdated and barely maintained. Workflows were manual and slow, with no meaningful integration into other systems. Most critically, there were no AI or machine translation capabilities at all. In an era where AI-assisted translation has become industry standard, the team was working without any of it.
What we set out to do
The goal was not to replace translators with AI. That framing misses the point entirely, especially in a regulated industry where precision, tone, and legal accuracy matter. The goal was to give a small team the best tools available so they could work faster, handle growing volume, and maintain quality across four languages without scaling headcount.
That meant two things: a modern Translation Business Management System (TBMS) to manage workflows, terminology, and translation memory, and an AI translation engine that could deliver high-quality first drafts for human review and post-editing.
Evaluation and decisions
I led the project from requirements gathering with the Language Service team through vendor evaluation, architecture decisions, and implementation.
For the TBMS, we selected RWS Trados. It is the industry standard for translation management, with strong support for translation memory, terminology management, and project workflows. It replaced Multitrans as the central system the team works in every day.
For AI-assisted translation, we chose Supertext (formerly Textshuttle) — a Swiss company that spun out of the University of Zurich's Department of Computational Linguistics. Two things made them the right fit: first, Swiss hosting and data processing, which matters in a regulated healthcare environment. Second, the ability to custom-train their AI models on SWICA's own terminology and style, so translation quality improves continuously over time rather than staying generic.
How it works today
The system is fully live and in daily use. Translators work in Trados, where they manage projects, leverage translation memory for recurring content, and maintain consistent terminology across all four languages. Supertext's AI engine is integrated into the workflow, providing machine-translated first drafts that the translators then review, refine, and approve.
The human translators remain at the centre. They make the judgment calls on tone, regulatory language, and context that AI cannot reliably handle, particularly in healthcare and insurance content. But they no longer start from a blank page. The AI handles the heavy lifting on first drafts, and the translators focus their expertise where it matters most.
What changed
The impact has been tangible. Translators work significantly faster because they are post-editing AI suggestions rather than translating from scratch. The team handles more volume without growing headcount. And because Supertext's models are custom-trained on SWICA's content, the quality of AI suggestions has improved steadily since go-live.
Beyond the immediate productivity gains, the new stack opens up real integration potential. Instead of manually creating translation tasks for every piece of content, we can now connect other applications (website, mobile apps) directly into the translation workflow. That is where the next wave of efficiency will come from.
What I took away
The biggest lesson was how much this project was about people, not technology. The Language Service team had worked with the same tool for twenty years. Introducing AI-assisted translation required trust — trust that the goal was to empower them, not to automate them away. Spending time on evaluation together, involving the translators in vendor demos and pilot phases, and letting them shape the workflows made the difference between a tool that gets adopted and one that gets resisted.
If you are working on a similar challenge: modernising translation infrastructure, evaluating AI translation providers, or navigating the build-vs-buy question for language services — feel free to reach out. Happy to share what worked and what I would do differently.
