1. Where the industry stands: older cars, fewer hands
Taiwan has more than 8.5 million registered cars (MOTC vehicle registration statistics), and roughly 5.6% of the fleet is over 25 years old. Older cars mean more service visits — structurally good news for independent shops. The constraint is people: entry-level automotive technicians in Taiwan start at around NT$27,000 a month, so shops struggle to hire and to keep staff. In most 5-to-8-bay shops one service advisor answers the phone, writes estimates, replies on LINE and chases retention — all at once.
Overseas shops are already offloading that. The 2025 Ratchet+Wrench industry survey found 61% of shops using an AI estimate generator, 59% using AI digital vehicle inspection and 52% using AI scheduling.
2. The four use cases worth building first
- Missed-call recovery: at peak hours three lines ring at once and calls get dropped. AI catches the missed number and sends a LINE message — "You just called; would you like to book a service?" — with open slots attached.
- Estimate write-ups: the technician dictates "left front wheel bearing noise, brake pads at 3mm"; AI turns it into plain language for the owner, split into "do now" and "can wait".
- Personalised service reminders: next-visit timing derived from the model's maintenance schedule and the last recorded odometer reading, instead of blasting one canned message to everyone on the same day.
- Google review drafts: every review answered within 24 hours, AI drafts, owner sends.
3. Two real cases
A. Hotai Motor's TOYOTA LINE official account (relaunched September 2025): Hotai merged vehicle browsing and owner services into a single entry point, added real-time body-and-paint repair status lookup, and upgraded push messages from plain text to rich cards with direct-action buttons. The transferable idea is collapsing booking, job status and reminders into one channel.
B. Sanyang Motor's SYM LINE AI assistant (launched June 2026): scoped into spec comparison, preliminary analysis of after-sales issues, and maintenance-interval guidance. The lesson is the boundary: SYM states explicitly that the AI offers "preliminary information and direction, not a replacement for a professional technician's actual inspection," and routes any noise, warning light or brake complaint to a service point. Neither company lets AI diagnose; shops that skip this step pay for it in trust the first time the AI is wrong.
4. Recommended tool stack
| Layer | Tool | Why |
|---|---|---|
| Channel | LINE Official Account + Messaging API | Owners are already on LINE; no new app to teach |
| Orchestration | n8n or Make | Chains "missed call → look up owner → send LINE → write back to the job card" without a custom backend |
| Model | OpenAI API or Claude API | Estimate rewriting and review drafts; a small model is enough |
| Speech | Whisper-class models | Technicians' hands are dirty — dictation beats typing |
| Data | PostgreSQL / Supabase | Service history needs a single source of truth |
Do not replace your whole DMS on day one. Run the numbers with n8n first.
5. ROI model (assumptions stated — rerun it with your own numbers)
Baseline: 5 bays, 260 vehicles serviced per month, average ticket NT$3,800, 1,800 owner records, 38% scheduled-maintenance return rate.
| Item | Assumption | Monthly effect |
|---|---|---|
| Missed-call recovery | 120 missed calls/month, 25% with buying intent, AI recovers 40% of those | +NT$45,600 |
| Return rate | Personalised reminders lift 38% → 46% | +NT$45,600 |
| Advisor hours | 70 min/day down to 25 min, 19.5 hours saved per month at NT$250/hour | +NT$4,875 |
| Running cost | LINE OA from NT$800, model NT$600–1,500, hosting and support NT$4,000–7,500 | −NT$8,000 |
Net gain is about NT$88,000 a month, so a NT$150,000 build pays back in roughly two months — that is the optimistic case. Discount the benefits to 30% (about NT$26,400 net per month) and payback is still under six months. If it does not work at 30%, your volume is too small: don't build it yet.
6. Rollout timeline
- Phase 1 (weeks 1–3) — clean the data: complete plate, model, last odometer reading and last service date. The dullest step, and the one that decides the outcome.
- Phase 2 (weeks 4–7) — one use case only: missed call to LINE booking. Measure the recovery rate.
- Phase 3 (weeks 8–12) — estimates and reminders: dictation-to-estimate, plus reminders to owners due within 90 days.
- Phase 4 (weeks 13–20) — reviews and reporting: review-reply drafts and a weekly bookings/return-rate report.
7. Four common failure modes and fixes
- Dirty data: blank odometer fields, the same owner recorded three times. Fix: do not enter Phase 2 until Phase 1 is finished.
- Letting AI make technical calls: owner asks about odd brake feel and the AI says "probably fine". Fix: safety keywords always escalate to a human.
- Over-messaging: three pushes a week drives block rates and message cost up together. Fix: cap marketing pushes at one per owner per month.
- No success metric: three months in, nobody can say what improved. Fix: write down the baselines before launch.
8. What should not be automated
- Fault diagnosis and safety judgement — a technician has to inspect the car.
- Collision estimates and insurance negotiation — liability is at stake.
- High-value complaints — the first sentence should come from the owner of the shop.
- Technician skill development — feel cannot be taught with a prompt.
9. How ScriptWalker can help
- Shop AI front desk (LINE booking, missed-call recovery, after-hours reception): from NT$68,000
- Estimate drafting and reminder engine (dictation-to-estimate, reminder segmentation): from NT$120,000
- Full integration (data cleanup, review replies, weekly reporting): from NT$180,000, support from NT$6,000/month
Every engagement starts with a free missed-call audit. If we cannot show recoverable revenue, we don't take the project.
10. FAQ
Is this worth it for a three-bay shop?
Look at volume first. Under 40 missed calls a month and fewer than 600 owner records, start cheap: a LINE official account, a one-page booking form and manual reminders.
Will customers notice it's a bot and hold it against us?
They will notice, and they should. Say up front that it's an online assistant and that complex questions go to a technician — far better than being caught pretending to be human.
Is it safe to hand plate and mileage data to an AI vendor?
A licence plate is personal data. Keep the database in-house or in a Taiwan data centre, send only required fields per request, and disclose the purpose when customers add you on LINE.
Can monthly AI costs run away?
What usually runs away is LINE messaging, not model spend. With a small model and capped response length, model cost typically lands at NT$600–1,500 a month.
How long until results show?
Missed-call recovery is measurable by weeks 4–6. Return rate is a lagging indicator — allow at least one maintenance cycle, three to six months.
11. Decision checklist and next step
- ☐ Can you state last month's missed-call count?
- ☐ How many owner records, and what share have complete mileage?
- ☐ What is your current return rate, and how is it calculated?
- ☐ Who answers the phone at peak, and what else are they doing?
- ☐ How long does an estimate take from dictation to delivery?
- ☐ How many days did your last Google review wait for a reply?
- ☐ Will you run one use case for a full six weeks before expanding?
- ☐ Have you written the rule that safety questions always go to a human?
- ☐ At 30% of projected benefit, does it still pay back in six months?
If questions 1, 3 and 9 have no answer, measure three weeks of baseline first.
- Email: [email protected]
- Phone: 0916-224-047
- LINE: @ufv9089p