Case study · AI
SupportAI ticket automation
Make + Claude workflow that triages support tickets and drafts replies for human approval.
#4812 · Billing · 94%
Charge posted twice after the March plan change.
Project context
Client labels may be portfolio names. Signed-off numeric outcomes appear only on FleetFlow in this redesign.
Context
Tier-1 volume ate the queue every launch week.
Problem
Tier-1 tickets consumed 60% of support capacity, response times slipped during product launches.
Research
Ticket categorization study, macro analysis, and review of 500 closed conversations.
Challenges
Tone consistency, PII handling, and agent trust in AI drafts.
Approach
AI drafts. Humans approve. Nothing sends itself.
Make workflow with Claude triage, draft replies, and human approval before send.
#4812 · Billing
Charge posted twice after the March plan change.
- Billing94%Route
- Login62%Manual
- #4812BillingSent
- #4808LoginHeld
- #4803ShippingSent
Claude routes; humans stay in control.
Make workflow with Claude triage informed by ticket categorization and macro analysis.
Drafts for approval, never auto-send.
Draft replies land in an internal review console with confidence scores and one-click edit/send.
Low confidence falls back to the manual queue.
Zendesk integration, audit logging, and fallback to the manual queue when confidence is low.
UX / UI decisions
UX and UI decisions that shaped the build.
Interface choices
Internal review console for support leads with confidence scores and one-click edit/send.
Solution
Ticket piles, then routed work with human gates.
Make workflow with Claude triage, draft replies, and human approval before send.
Before
Tier-1 volume dominated capacity; response times slipped during launches.
After
Make + Claude triage with draft replies and human approval before send, plus low-confidence fallback.
Inbox.xls
| Ticket | Status |
|---|---|
| #4812 | Unread |
| #4808 | Unread |
| #4803 | Open |
#4812 · Billing · 94%
Charge posted twice after the March plan change.
- #4808LoginManual
- #4803Shipping88%
Illustrative Before / after pattern · no client data shown
Engineering
Make, Claude, Zendesk · with a manual fallback.
Make scenarios, Zendesk integration, logging for audit, fallback to manual queue on low confidence.
Frontend
- Internal review console
- Confidence scores
Delivery
- Make scenarios
- Audit logging
- Manual-queue fallback
Stack
- Make
- Claude
- Zendesk
Outcome
What shipped for support ops.
Make workflow with Claude triage, draft replies, and human approval before send.
Measured outcome metrics are not published on this case until signed off. The narrative above is from the verified case source.
Where this work lives