Multilingual voice agent unlocked four-times the inbound call capacity.
Inbound calls in four languages — Dutch, English, Arabic, French — answered, qualified, and either resolved or routed in under 90 seconds. Without hiring.
The client runs an EU↔GCC freight brokerage. Inbound call volume had tripled in 18 months but the team hadn't grown — every new hire took six months to ramp on the four languages required, and turnover was 30%/year on the support desk.
Missed-call rate sat at 19%. Each missed call was worth €400 average revenue. The cost of a person to fix this was higher than the missed calls.
The system
A voice agent layer that sits between Twilio's inbound trunk and the human team:
- Pickup within 2 rings. Whisper transcribes in real-time, detects language, switches the agent's voice (ElevenLabs) accordingly.
- Triage agent (Claude). Three buckets: handle directly · qualify and book · escalate to human. Decision in ~3 seconds.
- Resolve directly: order status, ETA queries, document requests — fully automated, <60s call duration.
- Qualify and book: for new business inquiries, captures details and books a follow-up call directly into the team's calendar.
- Escalate: keeps the caller on the line, hands the conversation transcript to a human. Human picks up with full context.
What was hard
The hard part wasn't the AI. It was identifying which 60% of calls could be handled directly without losing the relationship feel. That came out of the audit phase — three weeks of listening to recorded calls before writing a line of code.
The second hard part was the language switching. ElevenLabs voices needed per-language tuning so the customer didn't notice mid-call swaps when the agent escalated and a human took over.
We didn't hire. We didn't lose customers. The voice agent answers in Arabic at 03:00 and our Tuesday-morning post-it about 'who's covering the hotline this week' is gone.