01 · Overview
The problem it solved
Paradyse Homes sells fractional property co-ownership in Bali. Every inbound lead arrived with a different level of awareness, intent, budget, and readiness to buy. The sales team handled each one manually: same call, same brochure, same response time regardless of signal quality.
Before
- Manual brochure delivery: often delayed hours
- Sales called every lead with the same pitch
- No qualification data before the first call
- Hot leads and tyre-kickers treated identically
- Follow-ups arbitrary, not intent-based
After
- Brochure delivered instantly on WhatsApp
- AI qualifies budget, intent, timing in conversation
- Sales receives scored, context-rich leads
- Suppressed follow-ups for uninterested contacts
- Human handoff on complex or high-intent signals
02 · Scope
What the bot is actually responsible for
One WhatsApp thread, six jobs. Nala carries a lead from first message to booked consultation, or to a human, without a separate tool for each step.
📄
Info delivery
Sends the brochure and listing details the moment a lead asks, no waiting on a rep.
🎯
Qualification
Extracts budget, intent, and timing through natural conversation, not a form.
🏡
Property recommendation
Filters listings to eligible options based on what's been shared so far, budget included.
💬
Answer & guide
Handles follow-up questions and keeps the lead moving instead of stalling in the thread.
🔁
Nurture
Re-engages based on how engaged the lead has actually been, not a fixed follow-up schedule.
🤝
Human handoff
Escalates to a sales agent when intent, complexity, or frustration calls for a person.
03 · Architecture
Let the model talk. Let the workflow decide.
LLMs handle language: understanding intent, generating natural responses, extracting structure. Workflow logic handles sequencing, routing, delays, and CRM updates. Neither tries to do the other's job.
→
WhatsApp Inbound
Lead sends message · workflow triggers · brochure sent instantly
01
Context Management
Chat history built and passed to all prompt stages every turn
02
Qualification Conversation
"Nala" asks budget, intent, timing, Bali relationship (one question at a time)
⚡
QUALIFICATION_COMPLETE detected
03
Answer Extraction
Reads full conversation → returns structured JSON. Never responds to customer.
04
Normalization
Maps raw answers to controlled enum values for downstream matching
05
Property Recommendation
Budget + intent → max 2 live eligible properties, with matched USPs
⚡
Post-recommendation mode
06
Nurture & Convert
Guide → Clarify → Convert → Handoff. No consecutive booking CTAs.
07
Human InterventionEscalation path
Explicit escalation: frustration, edge cases, direct human requests
08–09
Recovery & Follow-up
23h stall detection · context-aware suppression when inappropriate
04 · Qualification
Conversation in, structured data out.
Every answer is captured through natural conversation, then mapped to a controlled vocabulary, not free text, so the workflow can branch deterministically without a model interpreting prose downstream.
Dimension 01
Reason for buying
"Are you exploring for personal use, investment, or both?"
Determines which property USPs to emphasize. Investment intent unlocks ROI projections; personal use shifts to lifestyle framing.
reason_for_buying
own_use_only
own_use_and_investment
mainly_invest_some_use
investment_only
Dimension 02
Bali relationship
"Have you spent time in Bali or are you planning a first visit?"
Shapes conversational register and context depth. Bali regulars want less explanation; first-timers need more lifestyle framing.
bali_relationship
lives_in_bali
frequent_visitor
visited_before
first_time
Dimension 03
Budget
"What investment range are you working with?"
Filters the property recommendation to eligible options only. Budget mismatches are surfaced gracefully without ending the conversation.
budget_range
under_100k
100k_150k
150k_200k
200k_plus
still_exploring
Dimension 04
Purchase timing
"Are you thinking in the next few months or a longer horizon?"
Informs urgency of follow-up. Short timelines trigger booking CTAs; longer horizons shift to nurture mode with a lighter touch.
purchase_timing
within_3_months
3_to_6_months
6_to_12_months
just_browsing
Data flow: how conversation becomes CRM data
Natural language answers
→
Extraction JSON
→
Normalized enums
→
Property match
→
HubSpot contact
05 · Prompt Architecture
Every prompt does exactly one thing.
Separation of concerns prevents the system from doing conflicting things in the same turn: qualifying while recommending, or nurturing while escalating. Each prompt stage is isolated, focused, and replaceable.
Conversation layer: what Nala says and how
01
Personality
Defines Nala's tone: warm, concise, one question at a time. Prevents invented information and enforces WhatsApp-appropriate brevity.
02
Qualification
Guides conversation through four dimensions while responding naturally to prospect questions first. Never skips a dimension.
DATA
Extraction
Reads full chat history → returns structured JSON of four answers.
Never responds to the customer: reads only.04
Normalization
Maps extracted answers to controlled enum values, e.g.
mainly_invest_some_use, for downstream property matching logic.05
Prop Recommend
Selects max 2 live, budget-eligible properties. Explains the match using USPs. ROI included only for investment-intent prospects.
CLOSE
Post-qual
Switches mode after recommendation: Guide → Clarify → Convert → Handoff. Avoids consecutive booking CTAs.
07
Human handoff
Detects when to emit
HUMAN_INTERVENTION: [reason] for frustration, edge cases, or explicit requests for a person.08
Follow-up
Generates context-aware follow-ups. Suppresses with
NO_FOLLOW_UP when contact is uninterested, human-handed, or already engaged.06 · Guardrails
Human handoff is a feature, not a fallback
The system emits structured machine-readable signals the workflow can route on, not vague fallback responses that silently change behavior.
paradyse-ai · signal output
HUMAN_INTERVENTION: [reason]
Triggers immediate human escalation. Routes to sales via HubSpot task with reason attached. AI stops responding to that conversation.
QUALIFICATION_COMPLETE
Tells the workflow to pause the conversation loop and trigger the extraction → normalization → property recommendation sequence.
NO_FOLLOW_UP
Suppresses scheduled follow-up messages. Applied when contact expressed disinterest, requested a human, or is already mid-conversation.
HUMAN_INTERVENTION triggers
→User expresses frustration, anger, or dissatisfaction
→AI cannot confidently answer a specific question about a property
→User directly requests a human or wants to call someone
→User asks about full-villa pricing or full ownership (not fractional)
→User is in Bali now and wants to physically visit a property
→User is already in a meeting or wants to speak directly now
07 · Interactive Demo
Step through the qualification flow yourself
The exact conversation and system log from the production walkthrough. Click through it turn by turn.
N
Nala · Paradyse Homes
online
Step 0 of 31
Live fit score
–
Not yet scored
System log
08 · Results
Measurement framework
Instrumented around consultation bookings as the primary success metric, with human handoff volume and follow-up response rates as secondary signals.
Funnel stages · confirmed figures only
Headline metric
+37%
Increase in consultations booked vs. pre-launch baseline
20+ hrs
Manual work saved per week
100%
New lead coverage, 24x7