Built a local AI lead-recovery workflow with n8n, Ollama, Telegram and Google Calendar

Built with n8n
1

Hi everyone,

I’m sharing a self-initiated lead-recovery automation that I built and verified locally with n8n. The goal is to capture incoming leads, qualify them with a local AI model, store them in a structured way, notify the right person, and support appointment booking without calendar conflicts.

Workflow overview

Complete n8n lead-recovery workflow overview
Complete n8n lead-recovery workflow overview1200×671 44.9 KB

Public-safe overview of the complete n8n workflow.

What the workflow does

  • Receives new leads through a webhook
  • Uses Ollama for local AI-assisted qualification
  • Stores structured lead data in an n8n data table
  • Sends Telegram alerts for actionable leads
  • Checks Google Calendar availability before booking
  • Keeps the flow modular so individual steps can be adapted for different businesses

Stack

  • n8n
  • Ollama
  • n8n data tables
  • Telegram
  • Google Calendar
  • Webhooks and REST APIs

The project is intended as a practical demonstration of how missed or slow-followed-up leads can be moved into a consistent qualification and booking process.

The repository includes the project details and implementation notes:

I’d appreciate feedback from the community, especially around production hardening, edge cases, and useful additions. I’m also open to small, clearly scoped, fixed-price n8n automation projects.

Thanks!

4

como manejas los errores y alucionaciones que podria tener ollama a la hora de calificar los leads?

5

Good question. In the current demo I’m keeping the LLM’s responsibility quite narrow: it extracts a fixed set of fields and returns structured JSON rather than making open-ended decisions.

I also validate the output before the workflow continues, and missing information is kept empty instead of being invented. For a production version, I’d add stricter schema validation, retries/fallback handling for invalid outputs, and route uncertain or incomplete leads to manual review rather than automatically acting on them.

I’m still improving this part, so I’d be interested to hear what safeguards you normally use with local models like Ollama.