Call Triage Agent

A voice assistant that answers inbound calls and routes them.

← AI lab

Concept. Not built; this page describes a proposed design.

What it would do

• Answer inbound calls and ask why the caller is calling.
• Handle routine requests itself: opening hours, directions, booking or moving an appointment.
• Transfer the call to a person when the request is urgent, sensitive, or unclear.
• After the call, write a short summary and any follow-up tasks into the CRM.

How a call would flow

1. A SIP gateway receives the call. The agent states that the call is automated and recorded, then greets the caller.
2. Streaming speech-to-text turns the caller’s words into text as they speak.
3. An intent classifier picks the most likely reason for the call and a confidence score.
4. A language model drafts the next reply. A rule layer checks it for blocked topics and required disclaimers before text-to-speech reads it out.
5. If confidence is low, or the caller asks for a person, the call is transferred.
6. When the call ends, the transcript, summary, and actions are saved to the CRM.

Parts

Part	Role	Candidate tools
Telephony	Receive calls, transfer calls	SIP trunk, Twilio or similar
Speech-to-text	Live transcript	Whisper, Deepgram, or similar
Intent classifier	Pick the reason for the call	Small fine-tuned model or LLM prompt
Reply generator	Draft the next sentence	LLM with a fixed system prompt
Rule layer	Block topics, add disclaimers, force transfer	Plain code, not a model
Text-to-speech	Speak the reply	Any low-latency TTS
Integrations	Calendar and CRM writes	Google Workspace, Microsoft 365, Salesforce APIs

Data it would need

• A list of call reasons, with example phrases for each.
• Answers to routine questions, written and approved by the business.
• Rules for when to transfer, and to whom.
• Recorded or transcribed sample calls for testing (with consent).

Targets (assumptions, not measurements)

• Reply delay under 1 second after the caller stops speaking.
• Every call the agent cannot resolve is transferred, not dropped.
• Transcripts deleted after a fixed period (for example, 30 days).

Risks and open questions

• Callers may not want to talk to a bot. How easy is it to reach a person?
• Speech-to-text errors on names, addresses, and accents can lead to wrong actions.
• A language model can say things the business never approved. The rule layer reduces this but does not remove it.
• Call recording consent rules differ by state and country.
• Transcripts contain personal data. Storage, access, and deletion need a clear policy.
• How to test it: replay sample calls and check the chosen intent, the reply, and the transfer decision.