Before you start

n8n (Cloud or self-hosted)
Google account: Gmail + Sheets
OpenAI API key
A knowledge base: one row per question and answer

1. Pick where it runs

n8n Cloud

Log in to your n8n Cloud workspace.
Workflows → Import from file.
Pick the file(s) in workflow/.

n8n self-hosted

Open your n8n (Docker or npm, v1.x/2.x).
Workflows → Import from file.
Same files; nothing else changes.

ChatGPT or Claude

Create a Custom GPT or Claude Project.
Paste gpt/INSTRUCTIONS.txt.
Upload your sheet or paste rows in.

2. Setup checklist

Download the files from your welcome emailUnzip. You'll use the workflow/, templates/ and guide/ folders.
Import knowledge-base-template.csv into Google SheetsKnowledge Base tab: id, question, answer (5 example entries, replace with yours)
Import tickets-log-template.csv into Google SheetsTickets Log tab headers
Import ai-customer-support-agent.n8n.json into n8nWorkflows → Import from file (n8n Cloud or self-hosted).
Connect Google SheetsUsed by: Get Knowledge Base, Log Ticket.
Connect OpenAIUsed by: Classify & Draft.
Connect GmailUsed by: Get Support Emails, Create Reply Draft, Label Processed.
Fill in the Config nodecompany_name, tone, signature, categories, escalate_categories, gmail_query (default: unread emails with the plugstaff-support label from the last 3 days), processed_label_id, confidence_threshold (default 0.75), batch_size.
Run it on 3 real rowsCheck the drafts and the status column before you raise the batch size.

3. Connect your tools

Service	Nodes that use it
Google Sheets	Get Knowledge Base, Log Ticket
OpenAI	Classify & Draft
Gmail	Get Support Emails, Create Reply Draft, Label Processed

In each node, choose Create new credential and sign in (Google uses OAuth; OpenAI takes an API key). n8n stores credentials in your instance, not with us.

What the statuses mean

Status	Meaning	Your next step
`drafted`	Reply draft is in the thread	Review, edit, send
`needs_human`	Escalated (see escalation_reason)	Answer yourself; add the answer to the KB so it's covered next time
`error`	See the error column	Fix, then remove the plugstaff-processed label from the email so the next run picks it up

Troubleshooting

Symptom	Fix
Everything is escalated	KB too thin or threshold too high → add Q&A rows, lower `confidence_threshold` to 0.65 and compare
Replies include quoted history	Some clients quote differently → the cleaner strips `>` lines; paste a sample in support and we'll extend it

Guardrails

Drafts only: a human sends every reply. Keep it that way until you trust the answers on your own data.
Customer emails are sent to your LLM provider. Mention AI-assisted processing in your privacy policy, and don't put secrets in the KB.
If you serve EU users and later let it send automatically, tell customers they're talking to an AI (EU AI Act transparency duties for AI that interacts with people). Not legal advice.

Set up the AI Customer Support Agent