Disclaimer: Composite persona for illustration — not a single named client.
数値は illustrative ranges であり、監査済みの財務結果ではありません。
1. Problem
A 12-person ecommerce team handling 800+ inquiries/month across marketplaces and their own store. Responses lived in personal inboxes and ad-hoc ChatGPT threads — no shared templates, no SLA visibility.
2. Before
| Area | State |
|---|---|
| Intake | Email · marketplace messages · spreadsheets |
| Drafting | Copy-paste · individual AI tools |
| Approval | Manager review via chat |
| Metrics | None · anecdotal workload |
3. Why existing tools failed
- Helpdesk SaaS did not match Japanese marketplace message formats.
- Generic AI bots lacked approval gates and audit-friendly metadata.
- MVP stopped at a FAQ page without workflow integration.
4. Solution (illustrative engagement)
- Classify inquiries into 8 template types (shipping · returns · sizing · …)
- Server-side draft generation · human approval before send
- PoC: one marketplace channel · 4 weeks
- Production phase: rate limits · logging · STG/PROD separation per MVPと本番の境界
5. Results (illustrative)
| Metric | Before | After PoC |
|---|---|---|
| First-response prep time | ~18 min avg | ~9 min avg |
| Template reuse rate | ~20% | ~65% |
| Escalations to manager | ad hoc | tracked queue |
6. Pitfalls
- Auto-send without human approval
- Pasting customer PII into external AI
- PoC success treated as full rollout
7. Summary
- EC support DX starts with one channel · template taxonomy · approval
- Link PoC scope to 診断結果からPoC範囲を切る
- Prioritize themes via 無料DX診断