LINE Support Bot

The AI takes the eighty percent, a person takes the rest

An AI support agent for LINE official accounts. A reusable template - each product or client gets its own bot, and only one knowledge file changes before it goes live.

Type: Support automation
Scope: Conversation design
Behaviour protocol
Deployment template
Industry: Customer service for small teams

A customer asks on LINE; the AI answers common questions instantly and hands off to staff when a human is needed (demo conversation) — Fig.1A customer asks on LINE; the AI answers common questions instantly and hands off to staff when a human is neededdemo conversation

The challenge

A small team cannot staff a chat channel around the clock, yet eighty percent of what arrives is the same handful of questions: how do I buy it, what does it cost, will it work for my case. Leaving them unanswered loses customers; answering them consumes the day. The difficulty is the other twenty percent - the ones that genuinely need human judgement. The bot has to recognise those and hand them over, rather than inventing an answer.

024 key points

What we built

Three behaviours, chosen explicitly
Answer directly, point toward a person, or escalate. On escalation the owner is notified and the AI goes quiet for that customer, so the bot never talks over a human mid-reply.
Every failure returns a safe reply
No error ever surfaces to the customer as a technical message. When a support bot breaks, the person on the other end should not be the one who finds out.
One knowledge file per account
The core is shared across every deployment and only the knowledge file differs. Running on edge compute, the marginal cost of one more client approaches zero.
Adversarial review caught three inherited bugs
Generalising from the existing production version went through adversarial review, which surfaced eleven issues - three of them real bugs inherited from the original — the bot replying over a muted conversation when a sticker arrived, a failed push permanently swallowing its own notification, and unsanitised push content that could forge a fake system message.