Support triage: an AI workflow that drafts and checks ticket replies
A customer writes in: they were charged twice for August and your bot closed their ticket. The reply has to be careful, accurate about the refund policy and quick.
This workflow splits the job the way a support team would. A cheap model classifies the ticket, a retrieval step finds the relevant FAQ, a stronger model drafts the reply and a policy checker reviews it for refund risk before a person approves it.
Free in Sim. No account or API key.
Sample input
Customer says: I've been charged twice for August and your bot closed my ticket. Write a careful support reply and flag refund risk.
How the workflow runs, step by step
1. Start
Ticket In
The customer's message starts the run.
Next: Classify
2. Agent
Classify
A small model labels intent and urgency.
Next: FAQ Retrieve (intent)
3. Tool
FAQ Retrieve
Searches your FAQ or help-centre text for the passages that apply.
Next: Draft Reply (kb)
4. Agent
Draft Reply
A stronger model writes the customer-facing reply from the ticket and the FAQ.
Next: Policy Check
5. Agent
Policy Check
The most capable model checks the draft for refund and policy risk. It is usually the most expensive step.
Next: Human Gate
6. Human approval
Human Gate
A person approves the reply before it is sent.
Next: Send Reply
7. Result
Send Reply
The approved reply is the result of the run.
Why build it as a graph
When every ticket goes through one large prompt, you cannot tell which part of the reply costs the most or where it went wrong. As a graph, each step has its own model, cost and output, so you can downgrade the cheap steps and keep the expensive check only where it earns its place.
Things to try
- Run it in Sim and look at the cost per step: most of the spend sits in the draft and the policy check.
- Apply the suggested insight to sample or downgrade the policy check, then re-run and compare cost and quality.
- Paste one of your own tickets as the task to see how the route changes.