2 ready-to-build workflows

AI agent workflows for Call centers & BPO

Answer, understand and speak back on live calls — in the caller's own language.

Each recipe below is expressed only in VegaDūta's real workflow building blocks — Trigger, Agent, Knowledge, Condition, Approval, Output, Tool (MCP), Voice, Device, Loop, Code and Parallel — so it maps 1:1 to something you can assemble in the Workflow designer. Consequential actions always pass a human Approval step.

1. Live voice agent that listens and replies out loud

Your queue overflows at peak hours. Simple calls (order status, balance, store hours) tie up agents who should be handling the hard ones.

Trigger: Call (inbound voice)

How the workflow runs

  1. Trigger

    Inbound call is answered. A Call trigger picks up on the Voice channel; the caller's speech is transcribed live.

  2. Agent

    Understand intent. An Agent node interprets what the caller wants and decides the next step.

  3. Knowledge

    Ground the answer. A Knowledge node pulls the real answer from your policies/FAQ/order system so the agent doesn't guess.

  4. Voice

    Speak back naturally. The reply is synthesized with ElevenLabs (or Sarvam for Indian languages) and played on the call — in the language the caller used.

  5. Condition

    Know when to hand off. A Condition node detects frustration, high-value intent, or anything out of scope.

  6. Approval

    Warm-transfer to a human. An Approval/Output node routes to a live agent with the transcript and context attached, so the caller never repeats themselves.

Channels & connectors

  • Voice / Call
  • ElevenLabs
  • Sarvam (22+ Indian languages)
  • Knowledge base

Outcome

Routine calls are fully handled by voice; complex ones reach a human already briefed with the full context.

Why it helps

Deflects the repetitive volume that burns out agents and inflates wait times, while keeping a human in the loop for anything that matters.

Build spec

1 agent — 1 voice agent on the live call; the Knowledge, Condition and hand-off are workflow steps.

System prompt (paste-ready)

You are a phone support agent for {{company}}. Speak in the caller's language, one question at a time, in short spoken sentences. Only answer from the knowledge base; if you don't find it, say so and offer a human. Never quote a price, policy exception, or account change you can't cite. If the caller sounds angry, asks for a human, or the request is out of scope, call transfer_to_human immediately.

MCP connectors

  • Order/CRM lookup — via a Tool (MCP) node
  • Telephony transfer — via the Voice channel

Built-in tools

  • knowledge_search (policies/FAQ/order status)
  • transfer_to_human

Guardrails

  • Grounded-only: answers must come from Knowledge with a citation, or it defers to a human
  • Auto-escalate on frustration / explicit human request / out-of-scope intent (Condition node)
  • No account changes or payments over voice — those route to a human or a separate approved flow

Output & delivery

The agent's text reply is synthesized by ElevenLabs (or Sarvam for Indian languages) and played on the call; on escalation, a warm-transfer carries the live transcript + detected intent so the human never asks the caller to repeat.

2. Zero-effort post-call wrap-up and CRM update

Agents spend minutes after every call writing notes and updating the CRM — time that isn't spent on the next customer.

Trigger: Webhook (call-ended event)

How the workflow runs

  1. Trigger

    Call ends. A Webhook trigger fires with the recording/transcript reference.

  2. Agent

    Summarize + tag. An Agent node writes a crisp disposition, next steps, and sentiment.

  3. Tool (MCP)

    Write to the CRM. A Tool (MCP) node updates the customer record — no first-party CRM connector needed; MCP calls yours.

  4. Output

    Follow-up on WhatsApp. An Output node sends the promised summary/next step to the customer on WhatsApp.

Channels & connectors

  • Webhook
  • Tool (MCP → CRM)
  • WhatsApp

Outcome

Every call is documented and followed up automatically, with the agent freed the instant they hang up.

Why it helps

Recovers after-call work time and removes the note-taking that agents skip when they're busy — which is exactly when it matters.

Build spec

1 agent — 1 wrap-up agent; the CRM write and WhatsApp follow-up are workflow nodes.

System prompt (paste-ready)

You are a call wrap-up agent. Given a call transcript, produce strict JSON {disposition, summary (max 3 lines), sentiment: positive|neutral|negative, next_steps[], follow_up_promised: boolean, follow_up_text?}. Summarize only what was actually said — never infer a commitment the agent didn't make. If a follow-up was promised on the call, draft follow_up_text in the customer's language; otherwise leave it empty.

MCP connectors

  • CRM — via a Tool (MCP) node (no first-party connector)
  • WhatsApp — send channel
  • Call-ended event — inbound via Webhook trigger

Built-in tools

  • http_request (fetch the transcript/recording by reference)

Guardrails

  • Summarize only what was said — no invented promises, figures, or next steps
  • CRM write is keyed to the call id — idempotent, never double-posts the same call
  • A customer follow-up is only sent when one was actually promised on the call

Output & delivery

The agent's JSON drives a Tool node that updates the customer's CRM record via MCP, and — when a follow-up was promised — an Output node sends the recap/next step to the customer on WhatsApp.

Related industries

Build one of these in minutes

The sandbox gives you a live agent workspace — no account, no card. Or head back to the full catalogue and compare patterns across every industry.

Try this in the sandbox — no signupBrowse all industries
See how VegaDūta compares to n8n, Dify and BotpressWhy VegaDūta's architectureEstimate your WhatsApp API costs