Customer support
Cutting first-response time without auto-sending replies
A workflow classifying tickets, retrieving account details and drafting responses while keeping agent review before every send. This is an example workflow design; the exact approach and results depend on the organisation’s process, data and systems.
- INDUSTRY
- Consumer services
- SERVICES
- AI & Agentic Workflows, Internal Tools
The problem
Agents often spend the first minutes of every ticket working out what it is about and looking up the customer’s order in other tools.
Typical manual workflow
- 01Ticket arrivesmanual
- 02Agent reads and tagsmanual
- 03Order looked upmanual
- 04Account looked upmanual
- 05Reply writtenmanual
- 06Sentmanual
Dashed steps are typically done by hand.
The approach
Tickets are classified on arrival, with the relevant order and account details attached and a draft reply prepared from the support knowledge base. Agents still edit and send.
Example system design
Design decisions
Why nothing is auto-sent
Refunds and account changes carry real cost. The goal is to make agents faster, not remove their judgment.
Failure handling
Example outcomes
| MEASURE | MANUAL PROCESS | AUTOMATED PROCESS |
|---|---|---|
| First response | 3h 10m | 1h 31m |
| Lookups per ticket | 2–3 | 0 |
Technology
- OpenAI
- Zendesk API
- PostgreSQL
- Make