How to Build an AI Customer-Service Chatbot: From RAG Knowledge Base to Human Handoff, Cost, and a 90-Day Roadmap

At 11 p.m., a customer asks on your site whether you are open weekends and can they return this. Nobody replies, and by morning they have bought elsewhere. A mid-size e-commerce store gets 1,000+ support emails a month, 60% of them repeats: shipping cost, delivery time, how to return. That is the pain an AI chatbot solves: auto-answer the repeats, save humans for conversations that truly need a human. But just plug in ChatGPT is the most expensive mistake; without your own data, it will confidently make things up.

When It Fits vs When It Doesn''t

Good fit for an AI chatbot:

Large volumes of repetitive Q&A (shipping, hours, specs, returns)
Stretched support staff; no coverage after hours or on weekends
Complex product/service info customers struggle to find
You already have documents, FAQs, product data to feed it

Not a fit (or hold off):

Highly personalized questions needing complex judgment (medical diagnosis, legal advice, complaint negotiation)
Very low message volume (fewer than 5 a day; just use a human)
No existing documentation at all (organize your knowledge first)
Industries where a wrong answer is catastrophic (finance, healthcare) needing strict human review

Alternatives Matrix

Option	Pros	Cons	Cost band
Custom RAG AI support	Answers your own data, human handoff, controllable	Must build/maintain a KB	Build NT$80,000–300,000 + monthly
Rule-based chatbot	Cheap, controlled answers	Only preset questions, no nuance	NT$1,000–5,000/mo
Plug in ChatGPT / generic LLM	Fast, natural conversation	Does not know your data, hallucinates, security risk	API usage-based
SaaS support platform with AI (Intercom Fin, Zendesk AI)	Full features, ticket integration	High monthly fee, data on their platform	US$50–hundreds/mo

Full Process Breakdown (with tools and deliverables)

Week 1 | Knowledge inventory: consolidate FAQ, product data, return policy, chat logs. Tools: Notion / Google Docs. Deliverable: a structured knowledge-source list.
Weeks 2–3 | Build the RAG KB: chunk documents, embed, store in a vector DB (pgvector, Pinecone). Deliverable: a searchable knowledge base.
Weeks 4–5 | Conversation design: design the bot-answers-then-hand-off funnel plus brand voice. Tools: Figma. Deliverable: conversation flow + voice guide.
Weeks 6–7 | Human handoff + admin: connect LINE OA / site widget; build one-click-to-human and a chat-log backend. Tools: Laravel + LINE Messaging API. Deliverable: a launchable support system.
Week 8 | Test + tune: stress-test with real historical questions, fix wrong answers, fill gaps. Deliverable: an acceptance report.

Real Cost Breakdown

Development: mid-size RAG build about NT$80,000–300,000 depending on KB size and integration complexity.
LLM API usage: per-token; typical SMB monthly usage about US$20–200.
Vector DB: self-hosted pgvector nearly free; SaaS like Pinecone from US$0–70/mo.
Maintenance (hidden but essential): KB must be kept current, wrong answers fixed; budget NT$5,000–20,000/mo.
Human agent hours: the bot will not catch everything; do not assume you can cut to zero.

Reality vs Client Imagination

Client thinks plug in AI and it answers everything correctly. Reality: without your data it hallucinates; RAG KB quality decides everything.
Client thinks set it and forget it. Reality: the KB needs ongoing maintenance; update it whenever a product or policy changes.
Client thinks fully replace human support. Reality: catching 50–70% of repeats is already a great result; the rest needs humans, and the handoff experience must be good.

Common Traps vs How to Avoid

Trap: the bot makes up answers (hallucination). Use RAG bound to your own data and design if-not-found say so and hand off; do not let it bluff.
Trap: stale KB. Build an update-the-KB-whenever-product-or-policy-changes process with a named owner.
Trap: no human exit. Every conversation needs a one-click human button so customers are not trapped in a bot loop.
Trap: data leakage. Do not feed customer PII into third-party training; choose an API that lets you disable training and sign a DPA.
Trap: robotic tone. Write a brand voice guide for the model so replies feel human.

Success Metrics + 90-Day Roadmap

30 days: watch auto-resolution rate (share ending without human handoff), aim for 40%+ first, and track the wrong-answer list.
60 days: fill knowledge gaps, push auto-resolution to 50–60%; analyze which question types get handed off.
90 days: add auto-update flows for new common questions; measure labor saved and CSAT.

Decision Checklist

☐ Do I have large volumes of repetitive support questions daily?
☐ Do I have existing FAQ/product data to feed the bot?
☐ Can I assign someone to maintain the KB?
☐ Do I accept the bot solving 50–70%, rest to humans?
☐ Do I have an after-hours/weekend no-reply pain?
☐ Can my industry tolerate occasional wrong answers (with human oversight)?
☐ Am I ready to budget monthly API + maintenance?
☐ Do I want chat logs kept in my own backend?
☐ Do I need LINE OA / site widget integration?
☐ Am I willing to start small (one product line)?

FAQ

Will the AI say something wrong and get me in trouble?

It can, if you plug in a generic model without binding your data. The right approach uses RAG to constrain answers to your knowledge base and designs an if-not-found hand off, minimizing hallucination risk.

How much and how long to build one?

A mid-size build runs about NT$80,000 to 300,000 over 6 to 8 weeks, then about NT$5,000 to 20,000 per month for API plus maintenance, depending on knowledge-base size and number of channels.

Can it fully replace human support?

Not advisable. In practice, offloading 50 to 70% of repeats is a strong result; complex, emotional conversations still need humans, and handoff smoothness directly shapes customer perception.

Call to Action

Want to know whether your support workload suits AI and what it costs? We offer a free support automation health check that calculates your automatable share and real cost. Contact us: