← Portfolio

AbitAI · Chow Tai Fook Jewelry Group · 2018

A 24/7 assistant for high-value luxury purchases

A conversational shopping assistant built 0→1 for Chow Tai Fook's Tmall flagship store. I strategized the end-to-end user journey, designed the NLU / NLG logic, and led a cross-functional team of engineers and analysts to decide what the bot should understand — and when it should step aside for a human.

0→1 productConversational UXNLU / NLG designCross-functional lead

+316%

conversion lift

95%

NLU precision · 203 core intents

82%

of user queries covered

01Overview

Chow Tai Fook's Tmall flagship store faced a critical service gap: 60% of customer inquiries happened outside business hours.

  • 60% of customer inquiries occurred 8PM–8AM, when no staff was available
  • High-value jewelry purchases ($500–$5,000+) require personalized consultation
  • Customer-service costs increasing 40% year-over-year
  • Conversion rate dropping during night hours

The UX challenge

Luxury purchases require empathy and trust — can automation deliver personalized consultation without feeling robotic?

My role

Lead Product Designer & UX for Conversational AI. Strategized the end-to-end user journey, designed NLU / NLG logic, and led a cross-functional team of engineers and analysts.

Team & timeline

  • 2 conversational designers
  • 2 AI engineers & 2 data analysts

Apr 2018 – Aug 2018 (5 months)

02Architectural process

From taxonomy to deployment. Three phases took the jewelry domain from a raw pile of transcripts to a documented, engineer-ready logic spec.

  1. Phase 01

    Taxonomy & mapping

    I structured the complex jewelry domain into 6 core logic pillars to ensure scalable intent mapping.

    Product infoCommercePolicyCuration

  2. Phase 02

    Interaction patterns

    Developed a design system of 5 reusable conversational components to keep the UI consistent across all 203 flows.

    • Static informational (FAQ)
    • Guided branching (sizing)
    • Contextual carousel (recommendations)
    • High-confidence handoff

  3. Phase 03

    Validation & spec

    We used AbitAI's internal authoring portal to build the flows directly — dragging and connecting modules to compose each conversation path — and stress-tested edge cases there before handoff to engineering.

    Outcome — reduced engineering rework by 30% through comprehensive logic documentation.

03Key decisions

1

Phase 1 · Foundation

Intent recognition & brand voice

Solves understanding user needs and responding naturally

The problem

Build a bot that understands different customer intents and responds naturally enough that users don't realize they're talking to AI.

  1. Intent recognition: “What ring size?” vs “Need help choosing size” vs “Size recommendation?” = the same intent.
  2. Response generation: template responses reveal the bot's nature, breaking trust for luxury purchases.

Research approach

Built a training corpus from Chow Tai Fook's customer-service data:

  • Analyzed 5,000+ conversation transcripts
  • Identified 203 unique customer intents
  • Clustered similar queries under each intent
  • Extracted response patterns from top agents

NLU design 1 · Intent classification

Handle real-world query variations

  1. Identified and defined 203 intents across 6 categories
  2. Created a corpus for each intent with variations
  3. Trained a classifier on the corpus to recognize intent despite varied wording

Example — the “return policy” intent includes:

  • “What's your return policy?”
  • “Can I return this?”
  • “How long do I have to return?”
  • “Refund policy?”

NLU design 2 · Response generation

Maintain human-like conversation

  1. 3–5 response templates per intent
  2. A variation-selection algorithm to avoid repetition
  3. Brand-voice modeling from human-agent language

Example — bot responses for “return policy”:

  • “You have 7 days from delivery — plenty of time to make sure it's perfect!”
  • “We offer a 7-day return window. Most customers know right away, but we want you to feel confident!”
  • “No rush — a full week from when you receive it!”

Validation

203

intents

95%

intent-recognition accuracy

82%

queries covered

2

Phase 2 · Flow design

Personalized conversation flow

Solves accurate answers, efficiently

Building on the 203 intents from Phase 1, I focused on designing flows for personalized queries.

Research

From the transcript analysis:

  • 35% of intents require personalized recommendations (“What ring size for my girlfriend?”)
  • Different products require different information:
    • Children's products: age + height (no weight)
    • Adult jewelry: height + weight
    • Some users already know their exact measurements

The problem

Generic answers don't work for personalized questions:

  • “Here's the size chart” → user still confused
  • “Size 6.5 for your 165cm / 55kg girlfriend” → helpful

But personalized answers require information. How do you gather it without wasting the user's time?

The decision

Smart branching logic — only ask what's necessary. Different conversation paths based on the user's specific situation; the bot skips every question it can already answer from context.

Example · size-recommendation flow

triggerObtain age + product link
children's product
skipSkip sizing questions → answer with child-specific products
free-size product
skipSkip questions → answer “free size”
ring
askAsk for finger size
necklace / bracelet
askAsk for height, then weight — only the parts still missing
checkSuitable size from [height] [weight] [category]
answerRecommend [size]
fallbackNo measurements known → provide size guide

Result

85%

completion rate

4 → 2.3

avg. questions asked

3

Phase 3 · Optimization

AI–human collaboration model

Solves the optimal AI–human balance

After testing the Phase 2 flows, an internal pilot with 100% AI automation revealed the limits of pure automation.

Testing insight

Internal pilot, 100% AI automation:

  • Good coverage (82%), but conversion didn't improve for high-value products ($2,000+)
  • Users abandoned complex / emotional conversations

Pure automation can't replace human expertise for high-stakes luxury purchases.

Identifying AI limitations

Analyzed pilot conversation patterns:

  • Confidence <80% → 3× higher abandonment
  • Emotional keywords present → users needed empathy, not just information
  • High-value context (>$2,000) → 60% of users requested a human agent anyway
We want to handle complex cases — that's where we add most value. The bot should filter routine queries.— Chow Tai Fook service team

The decision

Hybrid intelligence — AI handles volume, humans handle value. A confidence score plus context analysis runs on every message right after intent recognition, so routing happens before the user waits on a weak answer.

AI handles

>80% intent confidence · routine informational queries

Escalate to human

  • Confidence ≤80%
  • Emotional keywords
  • High-value (>$2,000)
  • User asks to “talk to a person”

The trade-off

Phase 2 — max cost reduction

100% automation

Risked lower customer trust and loss of brand empathy in high-stakes luxury moments.

Phase 3 — premium trust & quality

82% AI + 18% human

Preserved the personal-concierge experience by escalating complex emotional needs to human experts.

Overall conversion

+316%

High-value products

3× higher completion

04The impact

Multi-round conversations, empathy at scale, and smart escalation added up to 24/7 personalized shopping assistance.

82%

Logic coverage

95%

NLU precision

30% ↓

Operating cost

3×

Night-time ROI

The bot filters simple queries effectively, allowing our team to focus on complex cases. That's where we add most value.— Director of Digital Innovation, Chow Tai Fook Jewelry Group
← Portfolio