A support agent that answers, looks things up, and knows when to stop.
An AI agent with email and lookup tools handles the routine tickets. A human gate, a budget and a full trace keep it inside the lines.
Three workflows teams build here first.
- 01
First response on every ticket with the order, subscription or account already looked up.
- 02
Refunds, cancellations and address changes drafted by the agent and approved by a person before they execute.
- 03
Escalation with context: the full conversation, the agent’s reasoning and the tool calls, attached to the ticket a human picks up.
What runs, step by step.
The agent answers every ticket; a refund waits for a person before it is issued, and the reply goes back on the same ticket.
- 01New ticket
- 02Agent with order lookup
- 03Refund requested?
- 04Approve refund
- 05Issue refund
- 06Reply on the ticket
Every step is a node type in the catalog. Diagram only: no data, no live run.
Before you start
- A Zendesk connection, or a help desk that can call a webhook.
- An order-lookup tool: any published workflow that returns the order for a customer.
- A credential for the model provider, and a workspace AI budget.
- A Stripe connection for refunds, and a person who approves them.
What you get
Routine tickets get a first answer with the order already looked up; refunds are issued only after a person approves them, and every step the agent took is on the trace.
Limitations
- The agent can act only through the tools you give it. It cannot look up what no tool returns.
- When the AI budget is reached, the agent stops and tickets go to people. Keep a human queue staffed for that.
- The Intercom nodes cover contacts and notes. Replying on an Intercom conversation goes through the HTTP node.
- Approval adds the approver’s response time to the refund. The run waits until the approval timeout, then takes the fallback you set.
AI support agent with order lookup
A webhook hands a customer message to a tool-using agent that can look up the order over HTTP, then the answer is emailed back.
See each step's output, and approve the risky ones.
Illustrated with a support workflow; the run view and the approval request look the same for every workflow.
Built for this, not adapted to it.
An agent node with tools
The agent calls tools you define: fetch an order, search a knowledge base, send an email. Any published workflow can be a tool.
Email in and out
Mailbox triggers and send nodes for Gmail, Outlook, SendGrid and SMTP. The agent reads the thread and replies on it.
A human gate on risky actions
Refund above a threshold? The run suspends, a person approves or rejects, and the run continues. Nothing irreversible happens unattended.
Budgets and circuit breakers
Monthly token and dollar budgets per workspace and per provider. When the budget is hit the agent stops and the ticket routes to a human.
A per-iteration trace
Every reasoning step, tool call and token count is recorded. When a customer asks why the agent said something, you can show them.
4 model vendors and local models
Anthropic, OpenAI, Google and DeepSeek, plus any OpenAI-compatible inference server you host (vLLM, Ollama, LM Studio, llama.cpp) for on-prem. Switch vendor without rewriting the workflow.
Start from one of these.
Each one imports into a workspace; swap the connections and run it.
- AI support agent with order lookupA webhook hands a customer message to a tool-using agent that can look up the order over HTTP, then the answer is emailed back.
- AI screen, human approve, then sendClassify an inbound request, ask a reviewer to approve or reject it, and email the requester the outcome.
- Q&A over a fetched documentAnswer a question about any public document: fetch it over HTTP, hand the text plus the question to an LLM with a grounding prompt, and post the answer to a callback URL.
Zendesk
Intercom
Gmail
Anthropic
Slack
SendGrid
StripeRead before the first live run.
- How to test a workflow before it touches anything realStatic validation, a dry run against real config with side effects held back, then a first live run on one record. A practical order for shipping automations safely.
- How to keep an AI agent’s spending close to the number you setA retry loop, a huge tool result or a growing prompt can turn one agent into a runaway bill. Here’s how per-tenant budgets and a circuit breaker limit it.
- How much a dry run actually proves before you trust it liveA dry run executes your workflow with side effects suppressed and shows each step’s output shape. That shows the path resolves, not that the result is right.
See a workflow run on your own use case.
Book a technical demo. Bring an n8n file or pick a template, and we build it with you on the canvas while Workflow Health checks it before it ever runs.