AI customer support for e-commerce
High concurrency, many SKUs, pre- and post-sales tangled — how to stand up an open-source AI support stack that controls cost and lifts conversion.
Recommended stack
ChatwootDifyFastGPTn8nTypebot
Monthly cost
$80 - $300 (incl. LLM tokens, 5k monthly conversations)
Compliance notes
Consumer protection laws, cross-border payment compliance, GDPR / PIPL. Retain chat logs for at least 3 years.
Key challenges
- Inquiry volume spikes 10-50× during promotions
- 60-70% of tickets are order status, returns, shipping
- Multi-platform data fragmentation (Tmall / JD / Shopify / DTC)
- Pre-sales needs SKU + inventory awareness; post-sales needs order-system access
Why open source fits e-commerce#
Three traits make SaaS painful:
- Spiky volume — at 1k orders/day SaaS pricing is fine; at 50k orders/day Intercom / Zendesk pricing explodes
- Deep system integration — when a customer asks “where is my order,” the AI must hit ERP / WMS live, beyond superficial webhooks
- Multi-storefront — one open-source KB can serve many frontends
Recommended architecture#
Deployment shopping list#
- Chatwoot — multi-inbox: WeChat, email, widget, marketplace IM
- Dify — train returns policy, SKU details, shipping rules as KB
- FastGPT (optional) — dedicated product + review KB with its own vector store
- n8n — order lookup, refund initiation, ticket creation
- Typebot — visual triage at the web entry
Key KPIs#
| Metric | Target |
|---|---|
| AI deflection rate | 45-65% |
| First response time | < 3s |
| Pre-sales handoff rate | 8-15% |
| Order-status accuracy | > 95% |
Promo-day playbook#
- Load-test Dify 7 days ahead (~100 QPS on default config)
- Precompute top-SKU FAQ, cache in Redis
- Enable Chatwoot auto-assignment queue
- Add rate limits in n8n to protect downstream ERP