How a call flows

Wetel’s telephony capability lets an AI agent talk to people over a real phone call. This page explains the architecture for integrators.

Current status: Telephony is a test-environment capability available on request. It is not in production.

Status at a glance

Capability	Status
Calls to a demo phone number (test environment)	✅ Wetel-run demo, arranged on request
Outbound calls (Wetel-run internal tests)	✅ Internal test only
Customer outbound calling	🟡 Planned — not available
Production calls	❌ Not available
Live transfer to a person	❌ Not available (verbal offer only)
Call recording	❌ Off by default

Diagrams

Inbound call

sequenceDiagram
    participant Caller
    participant PSTN as Phone Network
    participant SIP as SIP Service (Cloud)
    participant Dispatch as Dispatch Rule
    participant Room as Voice Room
    participant Agent as Voice Agent (Worker)
    participant STT as Speech-to-Text
    participant Brain as Wetel Agent Workflow
    participant TTS as Text-to-Speech

    Caller->>PSTN: Dials the test phone number
    PSTN->>SIP: SIP INVITE
    SIP->>Dispatch: Route
    Dispatch->>Room: Create room, dispatch agent
    Room->>Agent: Join
    Agent->>TTS: Greeting (AI disclosure, offer a person)
    TTS-->>Room: Audio
    Room-->>SIP: Audio
    SIP-->>PSTN: Audio
    PSTN-->>Caller: Hears greeting

    loop Each conversational turn (barge-in supported)
        Caller->>PSTN: Speech
        PSTN->>SIP: Audio
        SIP->>Room: Audio
        Room->>Agent: Audio
        Agent->>STT: Utterance
        STT-->>Agent: Text
        Agent->>Brain: Message
        Brain-->>Agent: Reply
        Agent->>TTS: Text
        TTS-->>Agent: Audio
        Agent->>Room: Audio
        Room->>SIP: Audio
        SIP->>PSTN: Audio
        PSTN->>Caller: Hears reply
    end

    Caller->>PSTN: Hang up
    PSTN->>SIP: BYE
    SIP->>Room: End session

Outbound test lane (Wetel-run test)

This shows how Wetel engineers verify the system. Customer outbound calling is planned, not yet available.

sequenceDiagram
    participant Eng as Tester
    participant Script as Test Script
    participant SIPProvider as A SIP Provider
    participant Phone as Tester's Phone
    participant SIPService as SIP Service (Cloud)
    participant Trunk as Inbound Trunk
    participant Dispatch as Dispatch Rule
    participant Room as Voice Room
    participant Agent as Voice Agent (Worker)

    Eng->>Script: Start (with consent, daytime only)
    Script->>SIPProvider: Place call to tester's own verified number
    SIPProvider->>Phone: Ring
    Phone->>SIPProvider: Answer
    SIPProvider->>Phone: Provider notice (trial)
    Phone->>SIPProvider: Acknowledge
    SIPProvider->>SIPService: SIP INVITE (via trunk credentials)
    SIPService->>Trunk: Match
    Trunk->>Dispatch: Route
    Dispatch->>Room: Create room, dispatch agent
    Room->>Agent: Join
    Agent->>Room: AI disclosure + conversation

    Note over Phone,Agent: Same per-turn loop as inbound.<br/>Hard stop 300 seconds.

System components

flowchart TB
    subgraph Phone["Phone Network"]
        CallerPhone["Caller's phone"]
    end

    subgraph Cloud["Cloud SIP + Rooms"]
        SIPService["SIP Service"]
        DispatchRule["Dispatch Rule<br/>(one room per caller)"]
        VoiceRoom["Voice Room"]
    end

    subgraph Worker["Voice Agent Worker"]
        AgentProcess["Agent process<br/>(one call at a time)"]
    end

    subgraph Services["AI Services"]
        STT["Speech-to-Text<br/>(multi-language)"]
        TTS["Text-to-Speech"]
        Brain["Wetel Agent Workflow<br/>(staging)"]
    end

    subgraph SIPProv["A SIP Provider<br/>(outbound test only)"]
        OutboundAPI["Outbound call API"]
    end

    CallerPhone <-->|"PSTN"| SIPService
    SIPService --> DispatchRule --> VoiceRoom
    VoiceRoom <--> AgentProcess
    AgentProcess --> STT
    STT --> AgentProcess
    AgentProcess --> Brain
    Brain --> AgentProcess
    AgentProcess --> TTS
    TTS --> AgentProcess

    OutboundAPI -->|"SIP INVITE"| SIPService

What each part does

Component	Role
SIP Service	Bridges the phone network to the cloud voice room.
Dispatch Rule	Routes an incoming SIP call to a new voice room and assigns an agent.
Voice Room	Real-time audio session between the caller and the agent.
Voice Agent Worker	Orchestrates each turn: listens, transcribes, thinks, speaks. Currently handles one call at a time.
Speech-to-Text	Converts each utterance to text (supports English and Malay).
Wetel Agent Workflow	The “brain” — runs the agent’s logic and returns a reply.
Text-to-Speech	Converts the agent’s reply to audio.
A SIP provider	Used only in the outbound test lane to originate a call to a verified tester phone.

What you can try today

A demo call. On request, Wetel can arrange a test call to a demo phone number. Dial it and speak to the AI agent. The agent discloses it is AI and offers to connect a person (verbal offer only — live transfer is not yet built). Barge-in (interrupting the agent mid-sentence) is supported.
Observe a Wetel-run outbound test. A Wetel engineer can demonstrate the outbound path by calling their own verified number with their consent, during daytime hours only.

Every call creates a session with a transcript, similar to chat.

What is not available

Feature	Notes
Production calls	The agent workflow runs on the staging environment only.
Customer outbound calling	Designed; not built. Requires a SIP provider with local numbers, legal review of consent and recording requirements, and opt-in controls (off by default).
Live transfer to a human	The agent verbally offers it; the transfer itself is not implemented.
Local (Malaysian) phone numbers	Requires a SIP provider with Malaysian numbers; not available yet.
WhatsApp / Meta voice calling	Not part of this system.
Call recording	Off by default.

Latency

Metric	Current	Target
Per-turn reply time	~3.5–5 seconds (workflow or prompt agent); ~1.4–1.8 seconds (speech-to-speech option, no workflow)	1.5 seconds
Speech-to-text per utterance	0.65–1.8 seconds	—

Phone audio is 8 kHz, which can reduce transcription accuracy for some accents. Latency improvements (streaming transcription, shorter replies, faster models, closer hosting) are on the roadmap.

The agent discloses it is AI at the start of every call.
Outbound calls require explicit consent from the recipient.
No recording by default.
Per-key daily call caps are enforced in the staging environment.
Customer-facing outbound calling, when built, will be off by default and require owner approval.