The problem
Inbound leads arrived on WhatsApp, which is where the agency's market actually talks. Nobody was watching the inbox out of hours, questions about services and pricing went unanswered until the next morning, and whatever was learned in the conversation stayed in the conversation. Nothing reached the CRM unless somebody retyped it.
The two failure modes were response latency and lost context. Neither is solved by hiring; both are solved by making the first exchange automatic and the handover structured.
What was built
A WhatsApp-first intake flow, orchestrated in n8n:
- Intent detection. An LLM classifies what the prospect actually wants — a price, a capability question, a human — rather than pattern-matching keywords.
- Grounded answers. FAQ responses are answered from the agency's own material, so the bot does not improvise about services it does not sell.
- Qualification. The conversation collects the fields that decide whether a lead is worth a call, in the order a human would ask for them.
- Booking. Qualified prospects are routed to scheduling rather than into a queue.
- CRM write-back. The lead lands in Pipedrive with custom fields populated from the conversation, not as a bare phone number.
- Consent and audit trail. Opt-in is captured and every exchange is auditable — a requirement, not a feature, once you are processing personal data under GDPR.
Why n8n rather than a chatbot SaaS
Because the agency keeps the asset. A hosted bot platform rents you an integration and owns the logic; an n8n workflow is a file you hold, version, and move. When the CRM changes, the workflow changes. When the vendor raises prices, nothing happens to you.
The trade is that somebody has to run it properly: an error workflow, alerting on failed executions, and pruning of execution data before it fills the database. A WhatsApp bot that silently stops answering is worse than no bot, because the prospect assumes they were ignored.
What this does not do
It does not replace a salesperson. It answers the questions that have answers, captures what a human would have had to ask for anyway, and hands over a warm, structured lead. Anything genuinely consultative still goes to a person — the flow is designed to get there faster, not to avoid it.
Stack
- Orchestration: n8n (self-hosted)
- Language model: OpenAI, for intent classification and grounded FAQ answers
- Channel: WhatsApp Business API
- CRM: Pipedrive, with custom fields written from the conversation
Verifying the engineering behind this
Most people advertising n8n work assemble existing nodes. When an integration has no node, or the
node lacks the endpoint you need, that gap becomes the project. Our Apollo.io integration node is
published on the public npm registry as
n8n-nodes-apollo (MIT, maintainer arturl95) —
you can read the source rather than take our word for it. We also hold n8n Verified Creator status.