Building an AI Voice Outreach Pipeline with n8n, Apify, and ElevenLabs
How to architect an end-to-end outbound sales pipeline that discovers leads, enriches them, calls them with an AI voice agent, qualifies the conversation, and writes a clean CRM record — reliably and without babysitting.
Outbound sales is mostly logistics. You need a steady supply of qualified leads, a way to reach them, a way to capture what was said, and a record of it all. Every one of those stages can be automated today. I have built this pipeline end to end, and the surprising lesson is that the AI voice is the easy part — the durability of the workflow is what makes or breaks it.
The five stages of the pipeline
- Discovery — find potential business leads that match an ideal customer profile.
- Enrichment — add the context a caller needs: industry, size, likely pain, a decision-maker name.
- Outreach — an AI voice agent places the call, opens the conversation, and qualifies interest.
- Qualification — capture structured answers, not a rambling transcript.
- Storage — write a clean, deduplicated CRM record and trigger the next action.
Discovery is a data-quality problem
I use Apify actors to extract leads from directory-style sources. The temptation is to grab everything and process later. Resist it. A pipeline that ingests ten thousand noisy records will drown your callers and burn your phone numbers. Apply the ideal-customer filter at ingestion time, not at the end.
- Normalise phone numbers to E.164 the moment they arrive. A number in the wrong format fails silently three stages later.
- Reject records without a verifiable business identity. Voicemail to a wrong number is a wasted call and a reputational risk.
- Keep the raw source URL on every record so a human can audit where a lead came from.
Enrichment turns a phone number into a conversation
An AI agent that opens with a generic script gets hung up on. An agent that says 'I saw you run a twelve-bay auto shop in Mesa and wanted to ask about your appointment bookings' gets a reply. Enrichment is what buys you those first ten seconds. In n8n this is a sequence of HTTP and LLM steps that fill a fixed schema before the call is ever placed.
{
"company": "string",
"industry": "string",
"size_band": "1-10 | 11-50 | 51-200 | 200+",
"pain_hypothesis": "one sentence, specific",
"decision_maker": "name or null",
"source_url": "url",
"dedupe_key": "normalised domain + phone"
}
Schema first, prompts second
Define the enrichment schema before you write a single prompt. When the schema is fixed, you can swap models, add validators, and route bad records to a human queue without rewriting the workflow.
The voice agent: constrain it
A voice agent that improvises is a liability. With ElevenLabs and a structured dialogue flow, the agent should have a narrow objective — confirm the right contact, ask two or three qualifying questions, and either book a follow-up or politely close. Constrain the branches. Every extra branch is a failure mode in production.
- Give the agent a hard time budget. Long calls cost money and annoy people.
- Always have a graceful exit and a compliance line. Do not let the agent argue.
- Capture answers as discrete fields, not as free text, so downstream automations can act on them.
- Log the recording reference and a consent flag on every call.
Qualification output must be machine-readable
The value of the call is not the audio. It is the structured outcome. Map every call to a small, stable result object: qualified or not, the reason, the best follow-up channel, and a confidence score. Everything after this point is a simple rules engine.
Storage and deduplication
Write to the CRM with an idempotent upsert keyed on your dedupe key. Workflows retry, and a retry must never create a second record for the same lead or dial a recently contacted number again. In Airtable this is a merge on a unique field; in a database it is an ON CONFLICT upsert.
INSERT INTO leads (dedupe_key, status, last_call_at, payload)
VALUES ($1, $2, now(), $3)
ON CONFLICT (dedupe_key)
DO UPDATE SET status = EXCLUDED.status,
last_call_at = EXCLUDED.last_call_at,
payload = EXCLUDED.payload;
Making the seams boring
The stages are simple. What separates a demo from a system that runs for months is what happens when something fails. Build for that first.
- Every external call needs a timeout, a retry with backoff, and a dead-letter path.
- Rate-limit yourself against each provider before they rate-limit you.
- Emit a metric per stage so you can see exactly where throughput drops.
- Add a kill switch. When a provider changes something at 2am, you want to pause the pipeline, not debug it live.
- Include a suppression list that is checked before every dial.
The AI is the flashy part. The boring engineering around retries, idempotency, and observability is what makes it a product instead of a demo.
Built this way, the pipeline runs itself: leads flow in, qualified conversations flow out, and a human only touches the records that actually deserve attention.