Outreach Email Classification Automation

Every reply to an outreach campaign had to be opened and read by hand just to tell an interested prospect apart from a bounce, an out-of-office, or an unsubscribe request — and n8n's native Gmail Trigger gave no way to filter or rate-limit what reached AI classification.

Outreach Email Classification Automation was built with n8n, Google Apps Script, Gmail, Google Sheets, Gemini AI, Slack.

This was built as commercial/client work, so the source code isn't publicly available. Get in touch to discuss the implementation.

AI-powered email reply classification uses a large language model to read the content of a cold-outreach reply and automatically categorize its intent — such as Interested, Not Interested, Schedule Later, Wrong Person, Out of Office, or Unsubscribe — removing the need for a human to manually read and triage every incoming reply.

A Google Apps Script bound to a Google Sheet can apply custom, spreadsheet-editable filtering rules — allowed recipients, blocked domains and senders, warm-up and unsubscribe keyword detection — and throttle outbound delivery to a downstream API at a precisely controlled rate, giving finer-grained, more easily tunable control than relying solely on a workflow engine's built-in polling trigger.

Rate-limit avoidance combines proactive throttling — sending requests at a deliberately slow, controlled pace, such as a fixed number per minute with a delay between each — with reactive handling, like automatically retrying a request when the API returns a 429 rate-limited response, so bursts of incoming replies don't overwhelm the AI provider's limits.

Large language models sometimes wrap structured JSON output in markdown code fences or format it inconsistently, so a parsing step needs to strip that formatting and explicitly throw a visible error on failure — rather than silently passing malformed or partial data further down the pipeline, which would otherwise produce a confusing, malformed notification with no clear root cause.

A prompt-driven classification system means the AI model's system prompt — not surrounding code — defines how inputs are categorized, so fixing a misclassification issue means editing the prompt's instructions rather than modifying the workflow's code nodes, making behavioral tuning faster and more accessible to non-developers.

A resilient webhook-based automation has the sending side — here, the Google Apps Script poller — detect failed delivery attempts and re-queue that item for the next scheduled run rather than discarding it, combined with persistent state tracking (a processed-IDs list and a pending queue) so retries don't create duplicate notifications.

Sending AI-classified outreach replies to a dedicated Slack channel gives a sales or outreach team near-real-time, at-a-glance visibility into new reply intents as they're classified, integrating directly into a tool the team already actively monitors rather than requiring them to check a separate dashboard or inbox.

Storing filter criteria — blocked domains, blocked senders, keyword lists — in Google Sheet tabs lets non-developers update outreach-filtering rules by editing a spreadsheet directly, instead of requiring a workflow or code change and redeployment for every new domain, sender, or keyword that needs to be added or removed.