The AI Duty Roster for Presale Sales Centers: Visitor Intake, Lead Tiering, and Reviving Dead Lists

In a year when transaction volume is down, every walk-in at a presale sales center costs more to acquire. This piece skips the usual "AI valuation" and "listing recommendation" talk and focuses on the three things a Taiwanese presale agency or developer should automate first at the sales center itself: visitor intake, lead tiering, and reviving dead lists.

1. Where the Industry Actually Stands

Taiwan's housing market hit a nine-year low in 2025: registration data from the Ministry of the Interior Real Estate Information Platform shows just 261,308 building sale transfers for the full year. Volume shrank, but the way sales centers handle data has barely changed — handwritten visitor cards, Excel trackers, and client preferences living in an individual agent's phone. When that agent leaves, the list is effectively dead.

Overseas data shows the same gap. The National Association of REALTORS® 2025 Technology Survey found only 32% of agents had never used AI — meaning roughly seven in ten had. Yet in the same survey, 46% reported no noticeable impact on their business and only 17% saw a significantly positive one. The problem is not model quality. It is that AI was never wired into the floor process and the CRM, so it stayed a browser tab people open separately.

2. Four AI Scenarios Worth Deploying at a Sales Center

2.1 Structured Visitor Intake

Turn handwritten visitor cards and LINE conversations into fixed fields with an LLM: budget band, size requirement, household size, motivation, mortgage status, decision maker, timeline. The output is JSON written straight into the CRM — not a nicely worded summary paragraph.

2.2 Project Knowledge Base Q&A (RAG)

Chunk the project fact sheet, floor plan notes, common-area ratio, parking layout, payment schedule, neighborhood amenities and escrow arrangements into a vector store, so the LINE official account answers 80% of the repeat questions (location, size, parking, launch date) outside office hours.

2.3 Lead Tiering With a Recommended Next Step

Combine rule-based scoring (budget fit, number of visits, whether the decision maker showed up, urgency) with an LLM-generated next-step suggestion and risk note. The day's walk-ins are auto-ranked A/B/C for the morning meeting instead of being sorted from memory.

2.4 Segmented Outreach to Dead Lists

A sales center's biggest asset is usually that three-thousand-row list of people who never closed. Re-cluster it by original requirement (price-sensitive, wrong unit size, waiting for a discount, bought elsewhere), map each cluster to a different outreach script, push in batches over LINE or SMS, and route anyone who replies straight back to a human agent.

3. Two Real Cases

Case 1: Yungching Realty's "Yungching AI Assistant." Buyers describe what they want in natural language ("near the MRT, suitable for a small family, park nearby") and the system proactively pushes matching listings. Per Commercial Times reporting, buyer showing rates rose 200% and purchase decision time accelerated 160% after launch, across more than 100,000 groups and 2.5 million group interactions. The lesson is not "they have an AI chatbot" — it is that the assistant sits inside LINE groups agents already use daily, and pushes rather than waits.

Case 2: Sinyi Realty's internal AI rollout. According to an iThome interview, Sinyi began deploying Microsoft Copilot across departments in 2023 for knowledge base building, marketing copy, industry analysis reports and CRM work, running internal training first and tool selection second. The takeaway: get back-office and sales-support staff productive first, then push to the customer-facing side. Resistance drops sharply.

The pit both fell into: data has to be machine-readable first. When listing description fields are inconsistent and every sales center defines its own visitor-card columns, AI can only produce vague answers. Both companies spent a large share of their timeline on data standardization, not prompt tuning.

4. Recommended Tool Stack

Self-host n8n rather than going all-SaaS: visitor data contains personal information, and the audit and contract cost of routing it through an offshore automation platform usually exceeds the cost of one VPS.

5. ROI Model (one agency selling three projects concurrently)

Deal-side assumptions are deliberately conservative: if this yields only 3 additional closings a year at NT$450,000 agency fee each, that is NT$1,350,000 in incremental revenue. Adding time savings, first-year net benefit is roughly NT$1,593,000 against NT$488,000 invested (NT$380,000 build + NT$108,000 run), a payback of about 3.7 months.

The better question is the stress test: assume zero incremental closings. On the 117 hours/month alone, payback is about 17 months. If your sales center runs fewer than two support staff, or you are selling only one project, that stretches past two years — in which case do Phase 1 and 2 only, not the whole thing.

6. Rollout Timeline, Phase 1-4

• Phase 1 (weeks 1-2), inventory: export the existing CRM and visitor cards, unify field definitions, pick one live project as pilot, and write the red-line list of things AI must never answer.
• Phase 2 (weeks 3-6), knowledge base MVP: chunk project data into pgvector, connect the LINE official account, configure human handoff and business-hours routing, and open it to internal agents only at first.
• Phase 3 (weeks 7-10), intake and tiering: Whisper plus the OpenAI API for field extraction, n8n writing to the CRM, a daily A/B/C ranking, and two weeks of calibrating weights against the sales manager's own judgment.
• Phase 4 (weeks 11-14), outreach engine and rollout: cluster the dead list, A/B test scripts, build the results dashboard, then replicate to the remaining projects and establish version control for project data.

7. Five Common Failure Modes and How to Avoid Them

• Deploying AI before structuring data: every sales center defines its own visitor-card columns and Excel has merged cells. Fix: build a field dictionary in Phase 1; five fewer fields is fine, inconsistency is not.
• Project details change but the knowledge base does not: prices adjust, unit types sell out, and the AI still quotes old figures — that is false-advertising exposure. Fix: single source of truth with reindex-on-edit, and stamp every answer with an update date.
• Treating the AI tier as a verdict, not a hint: one mis-ranked A lead from a black-box score and agents stop looking. Fix: show the reason fields next to every score and let agents override with feedback.
• No human-handoff threshold: when a buyer asks about negotiating room, loan-to-value or escrow specifics, an AI that answers anyway is a liability. Fix: red-line keywords trigger immediate handoff with a 5-minute SLA in business hours.
• Skipping the privacy process: visitor data sent to an external model without proper notice and consent. Fix: update the intake consent clause under Taiwan's Personal Data Protection Act and de-identify names and phone numbers before the API call.

8. Where AI Does Not Belong (an honest list)

• Deciding negotiating room and discount authority — that is a delegation problem, not a model problem.
• Building trust during a showing and closing the final gap; AI can only supply ammunition.
• Final legal review of the property description document, contract terms and escrow arrangements, which belongs to a licensed land administration agent or lawyer.
• Compliance judgment on building permits, occupancy permits, urban renewal and hazardous-building reconstruction procedures.
• Defect determination and liability allocation at handover inspection.

9. What ScriptWalker Builds for This Industry

We build the back office and the on-site mobile layer with Laravel and Flutter, and the AI layer with the OpenAI and Claude APIs plus pgvector and self-hosted n8n. Your data stays on your server.

• Project AI reception assistant (LINE official account + RAG knowledge base): from NT$180,000
• Structured visitor intake and CRM automation (including Whisper transcription): from NT$120,000
• Lead tiering and dead-list outreach engine: from NT$250,000
• Full Phase 1-4 rollout: from NT$380,000, with monthly operations from NT$9,000

10. FAQ

Q1: We are only selling one project. Is this worth it?

Phase 1 and Phase 2 are (about NT$180,000) — the knowledge base Q&A and structured intake. Hold off on tiering and the outreach engine until you have two or more concurrent projects, or a dead list over 2,000 rows; otherwise payback stretches beyond two years.

Q2: Will the AI quote wrong prices and expose us to false-advertising claims?

It will, if the knowledge base has no version control. Our approach: price, unit type and payment terms are not vector-retrieved at all — they are read live from the backend tables. Any price-related answer carries an "on-site posting prevails" note plus a timestamp, and anything out of scope goes to a human.

Q3: Is it safe to send client data to OpenAI or Claude?

API data-usage policies differ from consumer products, but the safer pattern is to de-identify before the call — tokenize names, mask phone numbers, and let the model see only the requirement description. Raw personal data stays in your own database.

Q4: What if the sales team refuses to cooperate?

Start by removing work, not by adding surveillance. Ship voice-to-text visitor cards first so agents type less, then introduce tiering two weeks later. Sinyi's sequence — train people first, choose tools second — is worth copying.

11. Decision Checklist and Next Step

• ☐ Do all our sales centers use identical visitor-card fields, exportable as structured data?
• ☐ Does our dead list exceed 2,000 rows with original requirements recorded?
• ☐ Do sales assistants spend more than an hour a day on data entry and CRM backfill?
• ☐ How long does an after-hours LINE or form inquiry wait for a reply?
• ☐ Can we write down the red-line list of things AI must never answer?
• ☐ Do price, unit type and payment terms have a single source of truth and an update process?
• ☐ Does our intake consent clause already cover outsourced processing?
• ☐ Is the sales manager willing to spend two weeks calibrating the AI tiering against their own calls?
• ☐ Do we have at least two more projects launching in the next 12 months to amortize the build?

If you ticked five or more, your floor data is clean enough to start at Phase 1. For a rollout assessment and quote scoped to the projects you are selling right now, get in touch:

• Email: [email protected]
• Phone: 0916-224-047
• LINE: @ufv9089p