How a Swiss mobile-subscription comparator handled 4,000+ German-language calls a month with one voice agent
A German-language voice agent for a Swiss mobile-subscription comparison service. About 1,000 calls a week, most of them either tech routing questions or purchase intent. The agent identified intent in the first 30 seconds, then either pointed the caller at the right provider, sent an SMS link to buy directly, or handed the complex cases to a human specialist with context already captured.
- Inbound calls per week
- ~1,000
- Monthly call volume
- ~4,000+
- AI success rate (completed calls)
- 95.9%
- Callback data capture
- 91.1%
- Real engagement rate
- ~52%
- Tracking window
- 6 weeks
The setup
A Swiss mobile-phone subscription comparator was getting a constant stream of inbound calls in German. Most weren't sales-ready. The two big buckets:
- People asking which provider's website to use for tech support on a phone or plan they already had
- People who'd already decided what they wanted and needed a link to actually buy
Neither bucket needed a human specialist. The specialist team was burning their day triaging both anyway.
The build
A German-speaking voice agent fronted the inbound line. First 30 seconds: identify intent. Then:
- Tech question about a specific provider → answer is "you want the support page at [provider]". Done. The comparator stopped doing free first-line support for someone else's product.
- Purchase intent → SMS link to the specific plan or SIM the caller had asked about. Buy from the phone after the call.
- Complex case → clean handoff to a human specialist with the captured context attached.
The hard part wasn't the German speech. Modern voice models handle regional German cleanly enough. The hard part was the routing tree: getting the agent to not punt to a provider when it could have answered, and to not try to answer when the right move was to capture details and route.
Six weeks of live data
About 1,000 calls a week. Roughly 4,000+ a month.
- 95.9% AI success rate on completed calls. "Completed" means the caller and the agent had an actual conversation; the agent helped, captured what was needed, or routed correctly.
- 91.1% data capture rate when a callback was needed. Name and number, clean enough that no one had to chase.
- ~52% real engagement on inbound. The rest were hang-ups and silences. That ratio is normal for a multilingual high-volume line. Most teams don't measure it, and so their "success rate" looks like 95% of nothing.
That last point is the one we kept emphasizing internally. The numbers above are only credible because the analytics layer separates real conversations from silences from day one.
What we'd tighten next
The provider-routing tree. There were edge cases where the agent could have answered directly but routed to a provider anyway, which added a step the caller didn't need. A couple of days of prompt work on the routing branch would have removed most of them. We sent the client the diff before handoff. Whether they shipped it, we don't know.
Why this case
Multilingual voice is no longer the hard part. Speech recognition is fine, voice synthesis is fine. What separates a working deployment from a vanity one is the routing accuracy, the capture quality, and the instrumentation that lets you tell the difference between 95% success and 95% silence.