3 wks
Voice assistant
Operators spent their time on repetitive calls: order status, opening hours, simple questions. We launched a voice assistant that answers these itself and transfers only complex conversations to a person.
- Timeline
- 3 weeks from brief to production
- Solution
- Voice bot on the phone line
- Team
- Orchestrator and 6 AI agents
- Stack
- STT · TTS · SIP · realtime
The challenge
Most inbound calls repeated day after day, yet each one needed a free operator. At peak hours callers waited on the line and some calls were dropped.
The assistant had to understand natural speech, reply without noticeable pauses and hand the call to a person in time when the question called for it.
- repetitive calls keep operators busy
- waiting on the line at peak hours
- complex conversations cannot be left entirely to a bot
The solution
- 01
Telephony
The assistant connects to the line over SIP and takes several calls at once, with no queue.
- 02
Real-time speech
Streaming speech recognition (STT) and synthesis (TTS) reply in 1.2 seconds on average, like a normal conversation.
- 03
Scenarios and LLM
Frequent requests follow tested scenarios, while free-form questions are handled by an LLM grounded in the company knowledge base.
- 04
Handover to an operator
When a question is complex or the caller asks for a person, the call is transferred with a short summary of the conversation.
Results
The assistant resolves 62% of calls without an operator and answers with no noticeable pauses. Operators only get complex conversations and see their gist straight away.
Metrics are a reference: we check the result on your data with a prototype in 1–2 days.
Common questions
Does the assistant sound like a robot?
We use modern speech synthesis with natural intonation and pick the voice and manner of speech to suit your company.
Can you connect our phone system?
Yes, if it supports SIP, as most PBXs and cloud telephony providers do. We clarify the details in the brief.
How does the assistant know a person is needed?
From an explicit request by the caller, the topic of the question and recognition confidence. We set up the handover rules together with you.
More cases
Knowledge-base assistant
Search and answers across 12,000 internal documents with source links and hallucination control.
Ticket processing pipeline
Classification, enrichment and routing of inbound requests by a fleet of agents with no operator in the loop.