The assistant is live. Guests still leave. It is usually not the model
Nine p.m., one person closing the shop. LINE: “Weekday haircut price? Seat at noon tomorrow?” They type the price. The guest asks about colour. They type that too. Then silence. The chair is empty the next day — not because nobody asked, because nobody finished the booking.
A newer model will not magically turn that thread into a reservation. The assistant we build with partners — kakinang (askai.tw) — is meant to search approved facts, show a human-readable summary before a lead or booking is stored, and hand off when it should not guess. Their own site says it is not just another chat box.
Four steps shops actually miss
- Clarify: what decision is still missing — people, slot, cut or colour?
- Check: only published prices and hours. Live stock and payments come from the system, not a made-up number.
- Confirm: show a short summary before submit. The same confirm must not create a second ticket.
- Handoff: refunds and complaints go to a named person.
The confirm screen does not need to look clever. The guest should see the slot, headcount, which price list, and who will call next. Only then does “send” write to the back office. A sticky network or a double tap must not open two bookings from the same confirm — that is two angry calls in the morning. Replay the original ticket; do not insert a second one.
Channels can be the site, LINE, Facebook. The failure is each one answering from a different sheet. Do not merge two strangers into one customer because the names look alike.
Write these down first
- Which of the top 20 questions must never auto-answer
- Who updates prices, and whether the assistant sees the new file
- Whether the page still claims “pay online” before payments are commissioned
- Which inbox gets night-time leftovers
- What you review after two weeks: handoffs, misses, duplicate leads
Chatbot vs agent: last week’s note. This one only adds: if people still leave, knowledge and confirm are usually missing. See AI services.
FAQs
Is kakinang a ZhuYun product?
It is an AI assistant service we build with partners. Public site: askai.tw. The same checks apply when we deploy support for a client: knowledge, confirm, handoff.
Must we switch platforms?
No. LINE, the website, or an existing desk can work. What matters is approved copy, a confirm step before writes, and where after-hours messages go.
Can we launch with no knowledge base?
Menus and handoff, yes. Open-ended answers with no files means guessing. One wrong price is worse than no bot.
Scope a support pilot AI agents
Product site: kakinang. Not a performance claim for that platform.