Outbound lead qualification

You will build a campaign that calls a list of leads, says who is calling and why, asks four qualifying questions, writes the answers to your CRM, and leaves a call record you can filter into a shortlist. The result is one published agent, one published tool, an outbound-capable number, and a batch that dials inside a calling window and retries unanswered phones.

A call that reaches voicemail is billed like any other connected call. With voicemail detection on, the agent leaves a message, hangs up, and the call ends as voicemail. A list with many machines costs money while reaching nobody.Every call in a batch runs the agent’s current published version, read before each dial. Republishing mid-campaign changes every later call. To hold a campaign steady, pin a version.

What you need

Provider keys for Deepgram and OpenAI (the models below).
A provisioned carrier account and an imported phone number.
A CRM endpoint that accepts a qualification write, with its token stored as the workspace secret CRM_TOKEN.
Credit on the workspace, and the EDITOR role.

Create the CRM write-back tool

The model fills the arguments while the person is on the line; the tree posts them with a token the model never sees.

record_qualification.json

{
  "name": "record_qualification",
  "description": "Record how this lead qualified. Call it exactly once, after they answer the budget and timing questions or as soon as they say they are not interested. Send 'unknown' for anything they did not tell you.",
  "json_schema": {
    "type": "object",
    "properties": {
      "interested": { "type": "boolean", "description": "True if they are open to a follow-up. False if they are not interested or asked not to be contacted." },
      "current_tool": { "type": "string", "description": "What they use to plan capacity today, in their words, or 'unknown'." },
      "budget_band": {
        "type": "string",
        "enum": ["none", "under_10l", "10l_to_50l", "over_50l", "unknown"],
        "description": "The annual budget band in rupees. 'none' means no budget this year."
      },
      "timeline": {
        "type": "string",
        "enum": ["this_quarter", "next_quarter", "this_year", "no_plan", "unknown"],
        "description": "When they want something in place."
      },
      "other_stakeholders": { "type": "string", "description": "Who else decides, by role or name, or 'unknown'." }
    },
    "required": ["interested", "current_tool", "budget_band", "timeline", "other_stakeholders"]
  },
  "operations": [
    {
      "kind": "http",
      "silent": true,
      "config": {
        "method": "POST",
        "url": "https://crm.northwind.example/v1/leads/{{userdata.crm_id}}/qualification",
        "headers": { "Authorization": "Bearer {{secrets.CRM_TOKEN}}" },
        "body": {
          "interested": "{{args.interested}}",
          "current_tool": "{{args.current_tool}}",
          "budget_band": "{{args.budget_band}}",
          "timeline": "{{args.timeline}}",
          "other_stakeholders": "{{args.other_stakeholders}}",
          "called_at": "{{system_vars.now}}"
        }
      }
    },
    { "kind": "say", "config": { "text": "Got it, thank you. I have noted that down." } }
  ]
}

A body value that is exactly one token keeps its type, so interested is sent as a boolean.
silent: true hides the CRM response from the model. With nothing returned to the model, the tool is silent: the say line plays and the agent does not add a reply of its own. See speech operations.
{{userdata.crm_id}} comes from each batch row (step 5), so validation warns that it is read before anything publishes it; that warning is expected here. {{system_vars.now}} uses the agent’s timezone.

TOOL_ID=$(curl -sS https://api.talqing.com/v1/tools \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d @record_qualification.json | jq -r .id)

curl -sS -X POST "https://api.talqing.com/v1/tools/$TOOL_ID/publish" \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"changelog": "First version"}'

Create the agent

An outbound prompt has to say who is calling, why, and how the person gets off the phone. The greeting and the voicemail message are personalized from the recipient’s userdata; a missing field resolves to nothing, so place tokens where an empty value still reads ("Hi {{userdata.first_name}}, this is Maya").

agent.json

{
  "config": {
    "name": "Northwind outbound qualification",
    "channel": "voice",
    "language": "en",
    "timezone": "Asia/Kolkata",
    "greeting": {"message": "Hi {{userdata.first_name}}, this is Maya from Northwind Tools, calling about the capacity planning guide you downloaded. Is now a bad moment?"},
    "prompt": "You are Maya, calling for Northwind Tools, which sells warehouse management software to mid-market distributors. You are calling someone who downloaded our capacity planning guide.\n\nWithin your first three sentences, offer to send an email instead. If they take it, thank them and end the call.\n\nIf they are willing, find out in order:\n1. What they use to plan warehouse capacity today.\n2. Whether there is budget for a change this year, and roughly what band.\n3. When they would want something in place.\n4. Who else is involved in deciding.\n\nRules:\n- One question at a time.\n- If they are not interested, thank them, confirm nobody will call again, and end the call. Do not push.\n- Never state a price, discount, delivery date or another customer's name.\n- Before the call ends, call record_qualification exactly once.\n- Short sentences. This is a phone call.",
    "stt": { "provider": "deepgram", "model": "nova-3-general" },
    "llm": { "provider": "openai", "model": "gpt-5.6-luna" },
    "tts": { "provider": "deepgram", "model": "aura-2", "voice": "aura-2-thalia-en" },
    "tools": [{ "tool_id": "3f8b0f4c-9a1e-4a7d-9c2b-1d5e6f70a821" }],
    "voicemail_detection": {
      "enabled": true,
      "message": "Hi {{userdata.first_name}}, this is Maya from Northwind Tools about the capacity planning guide you downloaded. I will follow up by email."
    },
    "analysis": {
      "enabled": true,
      "summary": true,
      "outcome": {
        "prompt": "Success: a real person agreed to talk and answered both budget and timing, even negatively. Failure: voicemail, cut off in the first fifteen seconds, or asked to be left alone."
      },
      "fields": [
        { "name": "fit", "type": "string", "description": "How well the lead matches a mid-market distributor with a warehouse. One of: strong, possible, poor, unclear." },
        { "name": "next_step", "type": "string", "description": "The single next action for a salesperson, as an instruction. 'none' if they asked not to be contacted." }
      ]
    }
  }
}

AGENT_ID=$(curl -sS https://api.talqing.com/v1/agents \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d @agent.json | jq -r .id)

curl -sS -X POST "https://api.talqing.com/v1/agents/$AGENT_ID/validate" -H "Authorization: Bearer $TALQING_API_KEY"
curl -sS -X POST "https://api.talqing.com/v1/agents/$AGENT_ID/publish" -H "Authorization: Bearer $TALQING_API_KEY"

Publishing pins record_qualification at its current published version; republish the agent to pick up a later tool version.Tool arguments versus analysis fields. budget_band on the CRM write is what the agent was told on the call — act on it. fit and next_step are what a model infers from the finished transcript afterwards, on the agent’s own LLM; they are never written into userdata. If a value has to be right, make it a tool argument. See call analysis.If the same agent also answers inbound calls, branch its opening on {{system_vars.direction}} in an on_enter hook.

Enable outbound on the number

A batch dials from from_phone_number_id, which must be active on a ready carrier account with can_outbound set. Otherwise the request is refused with this number is not enabled for outbound.

curl -sS -X PATCH https://api.talqing.com/v1/telephony/phone-numbers/8b41d0e6-2c77-4a19-b3f5-91ad6e40c8b2 \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"can_outbound": true}'

Each recipient’s to is normalized against the region of this number. See outbound calls.

Call yourself once

POST /v1/calls/outbound runs the published agent with the same userdata a batch row carries, exercising the greeting, the tool and the analysis. It rings a real phone and is billed; use your own number and a crm_id you can delete.

curl -sS https://api.talqing.com/v1/calls/outbound \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "from_phone_number_id": "8b41d0e6-2c77-4a19-b3f5-91ad6e40c8b2",
    "to": "+919812345678",
    "agent_id": "9f1c0a3e-4b21-4f7a-9c6d-2a5b8e13d704",
    "userdata": { "first_name": "Priya", "company": "Asha Textiles", "crm_id": "lead_test_1" }
  }'

The response carries session_id. After hanging up, check:

GET /v1/calls/{session_id} — the greeting was substituted, the email offer came within three sentences, and record_qualification ran once.
Your CRM — the row itself. The http operation is silent, so a 200 with the wrong body looks the same as a correct write from the call.
session.analysis on the same call. If a field is vague, rewrite its description and try it against this transcript with POST /v1/calls/{session_id}/analysis/preview (writes nothing, costs one LLM call).

Create the batch

One request carries the policy and up to 10,000 recipients. Each recipient is to plus a flat map of string userdata.

curl -sS https://api.talqing.com/v1/calls/batches \
  -H "Authorization: Bearer $TALQING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Capacity guide, week 36",
    "agent_id": "9f1c0a3e-4b21-4f7a-9c6d-2a5b8e13d704",
    "from_phone_number_id": "8b41d0e6-2c77-4a19-b3f5-91ad6e40c8b2",
    "timezone": "Asia/Kolkata",
    "start_at": "2026-11-02T04:30:00Z",
    "calling_window": { "start": "10:00", "end": "18:00", "days": [1, 2, 3, 4, 5] },
    "max_concurrency": 3,
    "max_attempts": 3,
    "retry_after_minutes": 90,
    "recipients": [
      { "to": "+919876543210", "userdata": { "first_name": "Priya", "company": "Asha Textiles", "crm_id": "lead_88213" } },
      { "to": "+919812345678", "userdata": { "first_name": "Arjun", "company": "Ravi Foods", "crm_id": "lead_88401" } }
    ]
  }'

import { readFileSync } from "node:fs";

type Lead = { phone: string; first_name: string; company: string; crm_id: string };
const leads: Lead[] = JSON.parse(readFileSync("leads.json", "utf8"));

const batch = await talqing.calls.batches.create({
  createCallBatchRequest: {
    name: "Capacity guide, week 36",
    agent_id: "9f1c0a3e-4b21-4f7a-9c6d-2a5b8e13d704",
    from_phone_number_id: "8b41d0e6-2c77-4a19-b3f5-91ad6e40c8b2",
    timezone: "Asia/Kolkata",
    start_at: "2026-11-02T04:30:00Z",
    calling_window: { start: "10:00", end: "18:00", days: [1, 2, 3, 4, 5] },
    max_concurrency: 3,
    max_attempts: 3,
    retry_after_minutes: 90,
    recipients: leads.map(({ phone, ...userdata }) => ({ to: phone, userdata })),
  },
});
console.log(batch.id, batch.status, batch.next_dial_at);

import csv

with open("leads.csv", newline="") as handle:
    recipients = [
        {"to": row.pop("phone"), "userdata": {k: v for k, v in row.items() if v}}
        for row in csv.DictReader(handle)
    ]

batch = talqing.calls.batches.create(
    name="Capacity guide, week 36",
    agent_id="9f1c0a3e-4b21-4f7a-9c6d-2a5b8e13d704",
    from_phone_number_id="8b41d0e6-2c77-4a19-b3f5-91ad6e40c8b2",
    timezone="Asia/Kolkata",
    start_at="2026-11-02T04:30:00Z",
    calling_window={"start": "10:00", "end": "18:00", "days": [1, 2, 3, 4, 5]},
    max_concurrency=3,
    max_attempts=3,
    retry_after_minutes=90,
    recipients=recipients,
)
print(batch["id"], batch["status"], batch["next_dial_at"])

Field	Range	Default	For this campaign
`max_concurrency`	1–10	1	2–3 while you are still reading transcripts
`max_attempts`	1–5	1	2–3 for a cold list
`retry_after_minutes`	5–1440	30	90 or more; a retry in the same hour usually hits the same voicemail
`calling_window`	`start`, `end`, `days` (ISO, 1 = Monday)	any hour, any day	Set it, or a list created at 23:00 dials at 23:00
`timezone`	IANA zone	required	The recipients’ zone; it interprets `calling_window`

Run it first with two of your own numbers and max_concurrency: 1. In the dashboard this is Telephony → Outbound Calling → New batch.

Running the campaign

Pause, resume, cancel, appending recipients (duplicates are skipped and reported), the circuit breaker that stops a batch after ten consecutive setup failures, and the automatic pause when credit runs out are covered in batch calling. next_dial_at is when the batch next dials, with next_dial_reason saying why it is waiting (start, window or daily_cap).

Reading the results

Where	What it gives you
`GET /v1/calls/batches/{batch_id}`	`status`, `counts` (including `voicemail`), `next_dial_at`, `failure_reason`
`GET /v1/calls/batches/{batch_id}/recipients?status=failed`	Per person: `attempts`, `last_close_reason`, `last_close_reason_label`
`GET /v1/calls?batch_id=…&outcome=success`	The calls that met the outcome, each with `summary` and `outcome`
`GET /v1/calls/{session_id}`	The analysis `fields` (`fit`, `next_step`) — the list does not carry them
`session.completed` webhook	`analysis` with summary, outcome and fields, plus `batch_id`, for every call as it ends

A shortlist is the successful calls whose fit is strong or possible; next_step is what a salesperson reads first. To build it in your own system, subscribe to session.completed instead of fetching each call.

Limits

Voicemail detection is the model’s judgement of the greeting it hears. An unusual greeting can pass as a conversation, and each retry that reaches the same mailbox leaves the message again.
One batch is one agent, one number, one timezone. Only userdata varies per recipient; a different agent or caller ID is a second batch.
The batch runs as the user who created it. If they leave the workspace or lose the EDITOR role, dialling fails and the breaker trips.
Calling law is yours. Consent, do-not-call registers, permitted hours and recording disclosure are not enforced by Talqing; calling_window does not know public holidays.

Batch calling

Statuses, retries, policy edits, daily caps and pinning a version.

Call analysis

Outcome, fields, preview and backfilling past calls.

Webhooks

Push each finished call, with its analysis, to your system.

Get started

Build

Guides

Outbound lead qualification

What you need

Running the campaign

Reading the results

Limits

Next

Batch calling

Call analysis

Webhooks

​What you need

​Running the campaign

​Reading the results

​Limits

​Next

Batch calling

Call analysis

Webhooks

What you need

Running the campaign

Reading the results

Limits

Next