AI email replies from Gmail with feedback saved to Sheets

AI email replies from Gmail with feedback saved to Sheets

An n8n workflow that answers new Gmail messages from the senders you list with a reply written by OpenAI. It saves each message and its reply to a Google Sheet. Every reply carries a Yes and a No link, and a click writes the answer into the sheet, so you collect rated examples for fine-tuning later.

What is inside

• workflow.json: the workflow as n8n exports it, fifty nodes: forty-one that do the work and nine sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The credential links are gone. Select your own for Gmail, Google Sheets and OpenAI.
• Generate reply was n8n's older OpenAI text node. The file names no model, so the node used its default, gpt-3.5-turbo-instruct. OpenAI's deprecations page, read on 12 November 2026, lists that model as shut down on 28 September 2026 with gpt-5.6-terra as the replacement. The model page for gpt-5.6-terra shows the legacy Completions endpoint as not supported.
• Generate reply is now an HTTP Request to Chat Completions. It names gpt-6-luna, sets reasoning_effort to none and takes max_completion_tokens from replyTokenSize. OpenAI's reference counts reasoning tokens in that limit, so the default of 300 could run out before any text appears with reasoning switched on.
• A new Reply Fields node follows it. It turns the answer into text and finish_reason, the two fields the rest of the workflow reads, and Format data reads text from it.
• The first line of the email parser Code node was a comment with its author's link. That line is removed and the parser is as it was.
• recipients held [UPDATE ME]. It holds <YOUR_RECIPIENT_EMAILS> now.
• The notes are rewritten. One of them named Pipedrive, which belongs to another workflow.
• Nothing else.

Before you start

• An n8n instance with the WEBHOOK_URL environment variable set to the address people reach it at. The Yes and No links in each reply are built from it. n8n's security page lists N8N_BLOCK_ENV_ACCESS_IN_NODE as false by default, which leaves that access on.
• A Gmail account, a Google account for Sheets, and an OpenAI API key.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add a Gmail credential and select it in On email received and Send reply to recipient.
3. Add a Google Sheets credential and select it in Create spreadsheet, Create or update rows, Paste data and Send feedback for fine-tuned data.
4. Add an OpenAI credential and select it in Generate reply.
5. In Configure, replace <YOUR_RECIPIENT_EMAILS> with the addresses to answer, separated by commas, or with * to answer every sender.
6. Leave spreadsheetId and worksheetId empty and the workflow creates a spreadsheet named by spreadsheetName. To use your own sheet, fill both in and give it the columns ID, Initial message, Generated reply and Good response?.
7. Save and activate the workflow, then send a test message from a listed address.

How a run goes

1. On email received polls Gmail every minute. Identify trigger #1 marks the item as gmail and Configure adds the settings.
2. Determine which trigger ran sends it on. Only continue for specific emails checks the sender against recipients.
3. Generate UUID gives the exchange an id. Extract message content (advanced) keeps the newest text of the message and drops quoted older mail.
4. Is text within token limit? compares the text length divided by four with maxTokenSize minus replyTokenSize. A longer message is ignored.
5. Generate reply asks OpenAI, and Reply Fields picks out the answer. If reply is complete lets only a reply that ended on stop through to Email template, and Send reply to recipient answers the message in its thread.
6. Format data builds a row with the id, the first message, the reply and an empty Good response?. The sheet nodes create the spreadsheet if it is missing, or add the row to it.
7. On feedback given is a webhook. A click on Yes or No sends the id and the answer. Identify trigger #2 marks it webhook, Record feedback passes it on and Send feedback for fine-tuned data writes the answer into the row with that id. The visitor sees a thank-you page.

Change the look

• The model and the prompt: the body of Generate reply.
• The reply length: replyTokenSize in Configure.
• Who gets an answer: recipients in Configure.
• The look of the reply: the HTML in Email template.
• The names of the spreadsheet and the tab: Configure.

Good to know

• The workflow sends real replies to every matching sender. Test with your own address first.
• A reply that stops at the length limit is not sent. Raise replyTokenSize if that happens.
• maxTokenSize is 4000, a cautious number. The model page for gpt-6-luna lists a context of 1,050,000 tokens, so you can raise it.
• The spreadsheet id is kept in the workflow's static data. Copy the workflow into a new one to refresh it.
• Anyone who has a feedback link can click it, and the link carries the exchange id.
• This copy was not run against Gmail, Google Sheets or OpenAI.
• The text of each email goes to OpenAI, and the sheet keeps messages and replies.
• The JSON holds no credential.