在 AI 處理前，如何將地產平台客戶查詢電郵解析成結構化紀錄

Many property portals still deliver leads to small agencies as notification emails. Before you add an AI assistant to the follow-up, turn those emails into clean, structured records. Clean input makes everything downstream simpler, from dedupe to follow-up timers.

This intake step is the unglamorous start of real estate lead automation for small agencies, and it is worth getting right.

The shape you want

{"source": "portal_a", "received_at": "2026-10-19T09:12:00+05:30",
 "name": "...", "phone": "...", "email": "...",
 "listing_ref": "...", "message": "...", "raw_id": "<message-id>"}

One parser per sender

Each portal formats its emails differently, so write one small parser per sender address and route by the From header.

import re

def parse_portal_a(body: str) -> dict:
    def grab(pattern):
        m = re.search(pattern, body)
        return m.group(1).strip() if m else None
    return {
        "name": grab(r"Name:\s*(.+)"),
        "phone": grab(r"Phone:\s*([+\d][\d\s-]+)"),
        "email": grab(r"Email:\s*(\S+@\S+)"),
        "listing_ref": grab(r"Ref(?:erence)?:\s*(\S+)"),
        "message": grab(r"Message:\s*([\s\S]+?)\n\s*\n"),
    }

PARSERS = {"leads@portal-a.example": parse_portal_a}

def parse(email):
    parser = PARSERS.get(email.sender)
    if parser is None:
        return route_to_human(email, reason="unknown_sender")
    record = parser(email.text_body)
    record.update(source=email.sender, raw_id=email.message_id,
                  received_at=email.date)
    return record

Validate before you store

REQUIRED = ("listing_ref",)
ONE_OF = ("phone", "email")

def validate(r):
    problems = [f for f in REQUIRED if not r.get(f)]
    if not any(r.get(f) for f in ONE_OF):
        problems.append("no_contact")
    return problems

If validate returns anything, send the email to a person with the reason attached. Do not guess.

Rules that save pain later

• Keep the raw email ID so you can re-parse when a portal changes its format, which happens without notice.
• Normalise phone numbers to one format before matching against existing contacts.
• Fail loudly. A parser that silently returns None for the phone number creates leads nobody can call.
• Use AI only for the free-text message. Pulling budget, move-in date or "wants to view this weekend" from the message is a good AI job. Store that output in separate fields so it never overwrites what you parsed deterministically.
• Write a test per portal using a saved real email with personal details removed.

Dedupe and follow-up timers come after this step. They are much easier when every lead arrives in the same shape.

Disclosure: I run NxFlowAI, an AI automation agency. Mapping intake like this is part of our first audit, which takes 72 hours for one workflow.