Gmail Support Email Replies with OpenAI & Claude (n8n)

Drafts and quality-checks Gmail support replies with OpenAI & Claude—only sending when it’s confident

This n8n workflow reads your latest knowledge base from Google Drive, understands incoming support emails, and automatically drafts Gmail responses using OpenRouter LLMs (OpenAI & Claude style models). It then runs a second “critic” quality check and sends replies only when citations and critic scores pass—otherwise it routes the ticket to a human reviewer with weekly metrics.

What this workflow does

• Nightly knowledge sync: Lists files in a Google Drive folder and compares versions against a Postgres ingest ledger to ingest only new or changed documents.
• Builds a vector knowledge base: Downloads changed files, deletes existing chunks for that document, splits text, generates OpenAI embeddings, and stores vectors in pgvector with document metadata.
• Ingests support tickets: Triggers on new unread Gmail messages with a support label or receives helpdesk tickets via a header-authenticated webhook, normalizing into one ticket format.
• Policy + sensitive handling: Masks PII in the ticket body, classifies the request with an OpenRouter chat model, logs it to Postgres, and applies routing rules to drop noise, restrict sensitive categories, or allow auto-answering.
• Auto-reply with citations: For eligible tickets, an OpenRouter drafting agent searches the pgvector knowledge base and generates a cited reply with confidence/unsupported-claims metadata.
• Critic-gated delivery: A second OpenRouter model critiques the draft; the workflow gates delivery using critic score/verdict, drafter confidence, citation presence, and zero unsupported claims.

Use cases

• SaaS support teams that want safe, grounded Gmail auto-responses from a controlled knowledge base.
• Automation engineers implementing an LLM + pgvector support assistant with human fallback.
• Teams that need audit-friendly routing and weekly metrics emailed to reviewers.

Technical details

• n8n nodes/tools: if, code, gmail, switch, webhook, postgres
• Knowledge source: Google Drive → Postgres/pgvector embeddings
• LLMs: OpenRouter chat models for classification, drafting, and criticism (OpenAI embeddings used for vectorization)
• Safety checks: PII masking, citation requirement, critic score/verdict gating, and unsupported-claims filtering