AI Customer ServiceAI Customer Service
Not a wrapper-around-ChatGPT toy. The bot actually reads your product catalog, FAQ and refund policy, handles roughly 70% of routine questions, and hands complex ones off to a human with the full conversation context preserved.
What's included
Every engagement is built from scratch, not templated. Each item is tailored to your actual needs.
- 01LINE OA / Messenger / Web widget — deploy to multiple channels→
- 02Trained on your FAQ, product catalog, and internal SOPs→
- 03Automatic hand-off to a human when needed, context preserved→
- 04Weekly conversation analytics report→
- 05Multilingual — Traditional Chinese / English / Japanese→
Scenarios
Common use cases for this service — to help you size the scope and expected outcome.
Cram school LINE bot
Parents ask about courses, schedules, pricing — the bot replies instantly; complex cases hand off to staff.
E-commerce FAQ assistant
Returns, shipping, size charts — answered from your own product pages and policy docs.
Answering the same question a hundred times — leaving no time for actual customer relationships
80% of questions are the same five
Shipping time, returns, sizing, stock, how to exchange — five questions eat 80% of your support day. The answers are already on your product pages, but customers don't want to hunt.
Off-hours = lost deals
Someone wants to buy at 10pm with a question. By the time you reply tomorrow, they've bought from a competitor. 50% of e-commerce orders happen after 8pm.
Your old bot doesn't understand
"My order" → "How can I help?" → "Order" → "Please wait..." — tree-menu bots make customers angrier and damage the brand.
High support turnover
Every new hire gets re-trained on products, processes, return policies. Two months later they leave. That knowledge belongs in a system, not a person's head.
An AI that actually handles 70% of your support
Not a "Hi, how can I help" template toy. A system that has read your products, FAQs and SOPs, and answers in your brand voice.
Knowledge base setup
- FAQ, product catalog, policy document ingestion
- RAG vector storage (pgvector or Pinecone)
- Answers cite sources — no hallucination
- Weekly auto-refresh of knowledge base
- Brand voice training (you provide 20 sample conversations)
Multi-channel integration
- LINE Official Account (essential in Taiwan)
- Facebook Messenger
- Website chat widget (one-line embed)
- WhatsApp (international)
- Unified conversation history across channels
Human handoff + analytics
- Auto-handoff to human when AI is uncertain, full context preserved
- Flags low-confidence answers so you handle them first
- Weekly analytics (top questions, hardest questions, complaint hotspots)
- Monthly AI accuracy tracking
- Abuse detection and blacklist
Three weeks from zero to live, first analytics report by week four
- Week 1
Audit + training data
Audit your FAQs, products, policies; structure into training data; collect 20 real conversations as samples.
- Week 2
Training + internal testing
RAG setup, prompt tuning, internal testing with 50 sample questions.
- Week 3
Channel integration + launch
Connect LINE / Messenger / website; 50% rollout; live monitoring.
- Week 4
Full rollout + first report
100% traffic + first-week analytics (AI accuracy, handoff rate, CSAT).
Pick the right model = save 80% on cost
The biggest trap in AI support isn't shipping it — it's a monthly API bill that exceeds a human CSR salary. We tier the models: cheap ones for routine questions, smart ones only for complex cases.
Handles 80% of routine questions. US$ 1 per 1,000 — 100× cheaper than a human.
For questions needing judgement (return disputes, special requests). Smarter than Haiku, cheaper than Opus.
Open source, cheap, sufficient. Lives inside Supabase / Neon free tier.
What Taiwanese customers actually use. Free up to 500 messages/month.
Prevents API bill blowups from abuse. Free tier works.
Better UX than reCAPTCHA, free. Blocks 90% of bot abuse.
You're probably wondering
- Q01
Does the AI hallucinate? What if it gives wrong answers?
With RAG, the AI can only quote from your documents — it can't invent facts. Every answer includes a source link the customer can verify. When the system can't find a confident answer, it hands off to a human instead of guessing. - Q02
Will the API bill get expensive?
For typical SMBs with 1,000-5,000 conversations a month, API costs run NT$ 500-2,500. We set daily usage caps — anything over pauses automatically, so bills never blow up. Still far cheaper than one full-time CSR. - Q03
Can it integrate with our existing LINE account?
Yes. LINE Official Account is the standard integration. If you don't have one yet, setup takes about a week (verification + business account approval). - Q04
Will customer conversations leak?
No. Conversations live in your own DB (Neon / Supabase). API calls go through Anthropic's zero-retention route — they don't store your conversations for training. We can also add PII filtering (auto-mask phone numbers, card numbers). - Q05
What if the AI can't answer — will customers get angry?
The AI is designed to say "let me connect you with a person" when uncertain, not to fake it. Most customers are actually satisfied with a clean handoff. What makes them angry is "AI pretends to know and gets it wrong" — exactly what this architecture prevents. - Q06
Will the AI model become obsolete?
Provider-agnostic architecture — swapping models is an environment variable change. Claude 4.7 ages out → swap in Claude 5 or GPT-5 without rewrites. Every 3-6 months we can tune prompts; small updates fall within the retainer. - Q07
Can we start with just the website widget, no LINE?
Yes, and we recommend starting there. The website widget is the simplest, most controllable channel. Add LINE after 1-2 months of stable operation to minimise risk.