Inbox triage without losing the human touch.
An LLM-assisted classification step can sort support messages before a person sees them. This demo shows the four branches — confident auto-reply, needs-human, frustrated customer and FAQ deflection.
What triggers this workflow
Email, web form or chat inbox event lands in the support queue.
Source: Email / web form / chat
Workflow steps
- Chatwoot + n8nCaptureMessage lands in Chatwoot with the customer, channel and full body.
- n8n + LLM endpointClassifyLLM step proposes a category, confidence and sentiment.
- n8nBranchConfident / needs-human / frustrated / FAQ deflection — different reply paths.
- ChatwootReplyAuto-reply goes to the customer; human-needed and frustrated paths go to the on-call queue.
Tools used
Human handoff
A person reads every reply before it goes out
Confident auto-replies are still reviewed by a team member for the first 30 days. Frustrated-customer replies are always reviewed by a human first.
- Customer is on the priority account list
- Sentiment is flagged as negative
- Reply would change the agreed policy
Try it in your browser
Edit the sample inputs, choose a branch, and see what the workflow produces. The sandbox runs offline — no accounts touched, no messages sent.
Try it in your browser
Edit the inputs, pick a branch, and see what the workflow produces. Runs entirely in your browser — no messages are sent, no accounts are touched.
1. Inbox intake
Incoming message from Hannah on email. Initial summary: "My order arrived damaged."
2. Classification
- Suggested category
- Warranty
- Confidence
- high
- Sentiment
- neutral
3. Routing branch
Subject: Re: My order arrived damaged. Hi Hannah, Thanks for reaching out. I've logged this as a warranty request and our team will be in touch within one business day with the next step. Your reference is SUP-9481. — Outback Gear Repairs
4. Disclaimer
Conceptual demo — the AI classification step requires a configured LLM endpoint and customer-history access. Production deployment needs a review queue and a human-in-the-loop approval policy.
Sample data only. Numbers, names and messages are placeholders for illustration. Wire the live workflow separately using the setup notes below.
Setup for production
Wiring it for production
Live deployment needs the LLM endpoint, the inbox source and a human-review policy.
- Connect the LLM endpoint (self-hosted or hosted) — credentials in n8n, never in the workflow JSON.
- Set the confidence threshold below which a reply routes to a human (default 0.7).
- Configure the human-review window for confident auto-replies (default 30 days after launch).
- Add the FAQ library so the deflection branch points to real guides.
Production prerequisites
- LLM endpoint configured with a clear data-handling policy
- Human reviewer on shift during business hours
- FAQ library kept up to date
Nothing here runs by default. Configuration has to be in place before any live send, booking or contact write.
Questions
Does the AI ever send a reply on its own?
Confident auto-replies are reviewed by a team member for the first 30 days. After that, a confidence threshold + a complaint trigger keep a human in the loop.
What if the LLM misclassifies the message?
A low-confidence message always goes to a human. We also review the classification weekly and tune the prompt when patterns drift.
More demos you can run in your browser
Want this customer support triage workflow running on your stack?
Pick the demo you're curious about or describe your own. We'll come back with what a real deployment looks like for your business — the workflow, the prerequisites, the people involved.