# Customer Support Triage Workflow

> Classify, prioritise, summarise and route every incoming ticket

| | |
|---|---|
| **Category** | Customer Support |
| **Type** | Automation |
| **Difficulty** | Beginner |
| **Automation** | Mostly automated |
| **Runtime** | ~3 min |
| **Est. cost** | $0.09 |
| **Version** | 1.0.0 |

## Purpose

Support queues fail at triage, not at answering. This workflow reads every incoming ticket, works out what it is actually about, how urgent it is and who should own it, then hands the agent a summary and a suggested reply.

The agent still sends. What changes is that they open a ticket already understood.

## Problems solved

- Urgent tickets queued behind trivial ones
- Agents re-reading long threads to find the actual question
- Inconsistent categorisation makes reporting useless
- Non-English tickets wait for a specific agent

## Benefits

- Urgency detected from content, not from the customer ticking a box
- Agents open tickets already summarised
- Consistent categories make reporting trustworthy
- Non-English tickets triaged immediately

## Limitations

- Suggested replies need agent review before sending
- Sarcasm and understatement fool sentiment detection
- The category taxonomy needs maintaining as the product changes
- Does not resolve tickets — it prepares them

## Success metrics

- Median first response time
- Routing accuracy (tickets reassigned)
- Agent handling time per ticket
- CSAT versus the pre-automation baseline

## Prerequisites

- **Api key** — LLM API key
- **Account** — Helpdesk with webhooks
- **Software** — Category taxonomy

## AI tools used

- **ChatGPT**
- **Claude**

## Steps

### 1. Ticket created trigger

Webhook fires with the ticket body, customer record and any thread history.

- _Time:_ 1 min

### 2. Classify and score urgency

Assign a category, detect sentiment, and score urgency from the content rather than the customer's self-selection.

- _Model:_ GPT-4o
- _Time:_ 1 min

```text
Classify this support ticket.

Ticket:
{{ticket_body}}

Available categories: {{categories}}

Return JSON:
{
  "category": "one of the above",
  "urgency": "low|normal|high|critical",
  "sentiment": "positive|neutral|frustrated|angry",
  "language": "ISO code",
  "is_churn_risk": true|false,
  "one_line_summary": "what they actually need"
}

Judge urgency from the impact described, not from tone. A calm message about total data loss is critical.
```

### 3. Summarise the thread

For tickets with history, condense the whole thread to what has been tried and what is still outstanding.

- _Model:_ GPT-4o
- _Time:_ 1 min

### 4. Draft a suggested reply

Produce a first-draft reply grounded in the help centre, in the customer's language.

- _Model:_ Claude Sonnet
- _Time:_ 1 min

> ⚠️ The draft may confidently state a policy that does not exist — agents must verify before sending

### 5. Route and notify

Assign to the owning queue, set priority, and alert the on-call channel for anything critical.

- _Tool:_ Slack
- _Time:_ 1 min

### 6. Agent review and send

The agent reads the summary, checks the draft against policy, edits and sends.

- _Time:_ 3 min

## FAQ

**Can this reply to customers automatically?**

It can, and for a narrow set of well-understood queries that is reasonable. Start with agent review on everything, measure where the drafts are consistently right, and only then automate that slice.

**How do I stop it misjudging urgency?**

Instruct it to score on described impact rather than tone, then sample a week of classifications against what agents actually escalated.

**What about languages we do not support?**

It will triage and summarise them regardless, which at minimum tells you what has arrived and how urgent it is.

---

_Exported from AIToolsay — https://invitationbuddy.com/workflows/customer-support-triage-workflow_