Voice agents have become genuinely usable in German over the last two years. Latency is down to under a second, interruptions work, and a well-built agent can take an appointment without sounding like a phone menu from 2011. That does not mean every company should have one.
The three cases where it clearly works
- High volume of repetitive inbound calls: appointment booking, opening hours, order status, delivery windows.
- Calls outside office hours that currently go to voicemail and are never returned.
- Peak overflow, where the agent takes calls 5 to 12 in the queue instead of letting them drop.
The common factor is that the conversation has a narrow, well-defined goal and a clear data source behind it. If your calendar, your order system or your CRM can answer the question through an API, the agent can too.
Where a human wins
Complaints, negotiations, medical intake and anything emotionally loaded should reach a person quickly. So should technical support that requires diagnosis rather than lookup. A voice agent that tries to handle these creates the worst outcome: the caller repeats themselves twice, then reaches a human anyway, already annoyed.
The actual numbers
Setup for a production-grade German inbound agent starts around 9,000 EUR, including telephony connection, calendar or CRM integration, prompt work and an evaluation set. Running costs are 0.06 to 0.14 EUR per minute of conversation plus the phone number. At 600 calls a month averaging three minutes, that is roughly 150 to 250 EUR monthly. Compare that with the alternative. A part-time reception role costs about 2,200 EUR per month all-in. If the agent handles 60 percent of calls end to end, it pays for itself within seven to ten months and then keeps working at night. Below roughly 200 calls a month the maths rarely holds, and a good answering service is the better choice.
Quality is measured, not assumed
Before go-live we record one hundred real calls from the existing line and run them through the pipeline. We measure three things: word error rate of the transcription, task completion rate, and the share of calls where the agent correctly decided to hand over. Anything below 85 percent task completion on the defined use case is not ready for customers.
An agent that knows when to stop talking is worth more than one that knows one more answer.
Dialect and background noise deserve their own test. Standard German and light regional colouring are handled reliably. Strong Swiss German, Bavarian or Saxon speech reduces accuracy noticeably, and construction-site noise reduces it further. A domain vocabulary of place names, product names and spellings recovers most of that gap. For the remainder, the handover rule matters more than the model, and a caller who reaches a person after fifteen seconds will forgive the detour.
Data protection is a design decision
- Announce the AI assistant and the recording at the start of the call.
- Use a model provider with EU processing and a data processing agreement.
- Keep transcripts only as long as the business purpose requires, typically 30 to 90 days.
- Never store payment data in a transcript; route those steps to a secure form or a human.
- Document the flow in the record of processing activities before launch, not after.
How to start without a big risk
Begin with one call type and one outcome, for example appointment booking for a single location. Route only calls outside office hours to the agent for the first four weeks. Listen to every conversation in week one and to a sample thereafter. If the numbers hold, extend the hours. If they do not, you have spent a feasibility budget and learned something concrete about your call mix.
