Chat agent that asks a human on Slack when it is stuck

Chat agent that asks a human on Slack when it is stuck

An n8n workflow with a chat. An AI agent answers what it can and remembers the last five exchanges. When it has no answer
or is unsure, it calls a tool that runs a second workflow inside the same file. That one looks for an email address in the
visitor's message. With one, it posts the message to a Slack channel so a person can reply. Without one, it asks the
visitor to send the question again with an address.

What is inside

• workflow.json: the workflow as n8n exports it, ten nodes and seven sticky notes. It holds no credential.
• README.md: these notes.

What we changed from the original copy

• The chat model was called gpt-4, which OpenAI shut down on 23 October 2026. The node is now named OpenAI Chat Model
  and calls gpt-5.6-sol, the replacement on OpenAI's deprecations page, read on 12 November 2026.
• The model node asked for a temperature of 0.2. The page for gpt-5.6-sol lists no temperature setting, so the copy
  leaves it out. Add it back in the node's options if your model takes it.
• The Chat Trigger lost its webhook id, and the Slack and OpenAI nodes lost their credential links. Select your own.
• Nothing else. The file has two small Code nodes, and neither needed repair.

Before you start

• An n8n instance, on n8n Cloud or your own server.
• An OpenAI API key that can use gpt-5.6-sol.
• A Slack workspace, an app that can post messages, and a channel for the questions.

Set it up

1. In n8n, create a workflow, open its menu and choose Import from File. Pick workflow.json.
2. Add an OpenAI credential and select it in OpenAI Chat Model.
3. Save the workflow. The tool Not sure? calls this same workflow by its own id, so it works only once the workflow is saved.
4. Add a Slack credential. In Message Slack for help, choose the channel by name. The file leaves it empty.
5. Open the chat with the Chat button at the bottom of the canvas. Ask the agent a question and tell it to answer as if it
   does not know. It should ask for your email address.
6. Send the question again with an address in the same message. Slack gets the message, and the agent confirms that a human
   was messaged.
7. To let visitors use the chat outside n8n, set Chat Trigger to public and read the notes below first.

How the tool and its sub-workflow connect

Not sure? is a Tool Workflow node. Its workflow is {{ $workflow.id }}, this workflow, and it starts at
Execute Workflow Trigger. It sends query, the text the agent hands to a tool, and one field of its own, chatInput, set
from the Chat Trigger's chatInput: the visitor's latest message. Only chatInput is used downstream. The tool's description
tells the agent to use it when it does not know or is not confident.

The sub-workflow ends in one of two Code nodes. Each returns a response text, and the agent's system message tells it to
answer with that text.

How a run goes

1. Chat Trigger takes the visitor's message.
2. AI Agent (version 1.2) reads it with Window Buffer Memory, whose context window is the node's default of five past
   interactions, and answers with OpenAI Chat Model.
3. If the agent cannot answer, it calls Not sure?. Check if user has provided email tests the message with a regular
   expression for something shaped like an address.
4. With a match, Message Slack for help posts the message and Confirm that we've messaged a human returns the thank-you.
   Without one, Prompt the user to provide an email returns the request to send the question again with an address.

Change the look

• The agent's rules: the system message in AI Agent.
• The tool's trigger: the description in Not sure?.
• The two replies: the texts in the Code nodes Confirm that we've messaged a human and Prompt the user to provide an email.
• The Slack message: the text in Message Slack for help.
• How much it remembers: Context Window Length in Window Buffer Memory.
• The model: OpenAI Chat Model.

Good to know

• AI Agent is version 1.2 with no agent type set, which n8n reads as a Conversational Agent. n8n's code marks that mode
  deprecated and says it goes in n8n v3. If your n8n shows that notice, replace the node with a current AI Agent and reconnect the
  model, the memory and the tool.
• The email check looks only at the latest message, not at the memory, so the visitor must put the address in the message that
  carries the question. It checks the shape of an address and does not confirm it exists.
• Window Buffer Memory keeps messages on the n8n instance. They are not shared between workers in queue mode.
• Each visitor message costs one or more OpenAI calls. Slack posts cost nothing extra.
• A public chat lets anyone with its address spend your OpenAI key, so add limits before you share it.
• The visitor's messages go to OpenAI, and one that asks for a person goes to your Slack channel with its email address.