Team 1
LogisticsFulfillment, supply chain, delivery, warehouse operations
1 lead + 5 ICs
01 Background
TikTok is the fastest-growing global e-commerce business — 1.58 billion users globally, 170 million in the US, and 15 million sellers powering Live Shopping, Shoppable Video, and Store experiences across the US, UK, ID, MY, PH, SG and TH. After scaling design teams at Amazon, Microsoft and Flexport, I joined TikTok as a Senior Design Manager leading the Business Platform Design department under the Global E-Commerce Design Team.
This case study walks through how I led a 17-person design org across multiple business platforms, and dives deep into one of our most impactful initiatives: the redesign of TikTok's Customer Service Platform (ByteHi), where AI-driven design lifted CSAT, doubled agent capacity, and cut average handling time.
COMPANY
TikTok
TIME
2023–2024
ROLE
Senior Design Manager
SCOPE
17-person design org
TikTok Global E-commerce
The fastest-growing business at TikTok — 16.1 Billion in revenue in 2023, with three core shopping experiences: LIVE Shopping, Shoppable Video, and Store.
02 Organizational scope & my role
The Global E-commerce Design Team at TikTok was organized around multiple design pillars. My department, Business Platform Design, owned the systems that powered the day-to-day operation of TikTok Shop — spanning logistics, finance, customer service, user growth, and internal dev tools.
Department footprint
Business Platform Design
A 17-person design organization across three core platform teams and two embedded focus areas.
17
members
3
core teams
5
platform areas
Team 2
Ops + FinanceOperations, payment, settlement, finance tools
1 lead + 4 ICs
Team 3
Customer ServiceSupport experience, CRM, agent tools, service workflow
1 lead + 3 ICs
Embedded
User GrowthGrowth experience, engagement, retention
1 IC
Embedded
Dev ToolsInternal tools for design, engineering, and platform operations
1 IC
Leadership & impact
My role in the team
Head of Business Platform Design Department
Strategy & ownership
Owned design vision, priorities, staffing, and operating processes for a global business platform design department spanning operations, logistics, finance, customer service, user growth, and dev tools.
Led a 17-person design organization across 5 platforms
Team leadership
Managed three platform teams with team leads, plus embedded designers across user growth and dev tools. Hired, mentored, and established team rituals from weekly critiques to quarterly OKR alignment.
Partnered with senior leadership across the organization
XFN partnership
Worked closely with product, engineering, research, content design, creative design, and senior leaders to drive design direction in a fast-paced global e-commerce environment.
A structure built for clarity, ownership, and cross-functional velocity — enabling designers to solve operationally complex business problems at global scale.
03 Case study — Customer Service Platform
To show how this organization translated into product impact, this case study zooms into ByteHi — TikTok Shop’s internal customer service platform, designed end-to-end by my Customer Service team.
Global E-commerce Design Team
Business Platform Design
Owns systems powering TikTok Shop operations
Logistics
Fulfillment, supply chain, delivery, warehouse
Ops & Finance
Operations, payments, settlement, finance tools
Customer Service
Support experiences, CRM, agent tools
User Growth
Growth experiences, engagement, retention
Dev Tools
Internal tools for design and engineering teams
Product overview
What is ByteHi?
ByteHi is TikTok Shop’s internal customer service platform — a suite of tools that helps service teams support buyers, sellers, and creators across messaging, ticket handling, workforce planning, training, quality assurance, and operational monitoring.
Platform role
A unified workspace for customer service operations
Instead of forcing agents and operations teams to switch across fragmented tools, ByteHi brings the core service workflow into one platform: live conversations, ticket resolution, staffing visibility, training, and quality control.
01
Agent efficiency
Reduce repeated manual work during high-volume service moments.
02
Operational visibility
Help managers monitor workload, staffing, and service quality.
03
Scalable service quality
Support consistent training, QA, and resolution standards across markets.
Core workflow
Instant messaging
A live support workspace where agents manage buyer and seller conversations with order context, service history, and response tools in one place.
Resolution system
Ticket center
Structured ticket handling for complex cases, escalation, follow-up, and cross-team resolution.
Operations planning
Schedule planner
Workforce planning tools that help operations teams forecast demand, assign shifts, and manage service capacity.
Quality & enablement
Training
Training and QA workflows that help teams onboard agents, improve service consistency, and maintain quality at scale.
Service context
Why ByteHi was needed
Customer service is inherently labor-intensive: agents need to respond quickly, calm emotional customers, document every case, and maintain consistent service quality even during high-volume promotional periods. Before redesigning anything, the team and I spent time on-site with agents in BPO centers to understand the day-to-day workflow pressures behind the support experience. Four pain points consistently surfaced:
Overwhelming message volume
During promotional periods, user inquiries often surged beyond agents’ daily reception capacity.
Threatening and insulting behavior
Some customers made threats or unreasonable demands, consuming significant agent time.
Conciliate customers
Agents not only solved problems, but also had to calm highly emotional customers during difficult moments.
Repetitive recording issue
After each chat, agents manually created tickets for future reference, slowing down timely responses.
04 Solution framework
Rather than ship one-off AI features, I framed the roadmap as a four-stage capability ladder — borrowing the mental model from autonomous driving. This gave the team, PMs, and leadership a shared vocabulary for what “AI in customer service” means at each step, and what success should look like.
“
I defined AI capabilities in the Customer Service Platform by comparing them to autonomous driving levels.
Design principle
AI excels at efficiency
Routine generation, summarization, and real-time assistance.
Humans excel at judgment
Risk control, edge cases, and accountable decisions.
L0
No AI
1 agent : 3× customers
L1
Agent Assistance
- Recommend replies
- User sentiment summary
- Automatic ticket filling
- ...
L2
Conditional Automation
- AI hosting
- Automatic ticket creation
- ...
L3
High Automation
- AI copilot
1 agent : 10× customers
The North Star was simple: at L0, one agent serves about three customers in parallel. At L3, with a true AI copilot, that ratio could move toward 1:10.
05 Design iterations & next-generation platform
We shipped AI capabilities progressively from L0 to L3, iterating on key experiences and measuring impact at every step. Each iteration shaped better outcomes and informed the foundation for a more integrated AI Copilot that guides agents end-to-end.
L0
Common phrase replies
Baseline to save typing time on peak hours.
→
L1
Recommend replies,
User sentiment summary
Smarter suggestions and emotion insights.
→
L2
AI Hosting,
Automatic ticket creation
AI-led handling and automated workflows.
→
L3
AI Copilot
in progress
Integrated guidance across the full service process.
L1
Recommend Replies
We tested multiple interaction patterns for our AI suggestion experience.
Design question:
How do you combine AI-composed text with manual input without interrupting the agent’s flow?
Design A
Design B
Design C
Final solution
Final Solution
Why the placeholder won
Showing the AI suggestion inside the composer as ghost text let agents accept it with a single Tab keystroke — or keep typing normally if they wanted to ignore it. This zero-friction “accept or continue typing” model drove adoption from 10.2% to 23.3%.
Adoption rate
Final Solution 23.3% ↑
Final Solution 23.3% ↑
L0
Common Phrase Replies
Simplest baseline: agents can search and insert pre-defined phrases from a Common Phrase Library, saving time on typing during peak hours.
Typing time saved
Baseline
L1
User sentiment summary
AI-powered sentiment analysis automatically detects user emotions and surfaces them at the top of the conversation, helping agents prioritize what matters.
Customer Satisfaction Score
+4.21% ↑
L2
AI Hosting
AI handles well-bounded, low-risk scenarios — simple status check-ins or case reconciliations — and hands back to a human when confidence drops.
Contacts per agent per day
+22.7% ↑
L2
Automatic ticket creation
AI auto-generates a structured ticket summary — tag, order ID, problem description, and recommended action — so agents can review and submit in seconds.
Average Handling Time
−20.31% ↓
L3
AI Copilot in progress
Our L3 AI Copilot integrates L1 and L2 capabilities into a single side-by-side panel. It gives agents full-process guidance — recognize the user, retrieve the order, check user requirements, propose solutions, and draft the reply — all in one place.
- Full-process guidance
- Unified context and actions
- Faster, more accurate responses
06 Team management
Designing the platform was only half the job. The other half was building the team, systems, and operating cadence behind execution — onboarding new hires quickly, creating lightweight intake and prioritization, measuring impact, enabling research under constraints, and keeping morale healthy in a high-pressure environment.
Team enablement
Onboarding & talent development
I rolled out a standardized onboarding guide, paired every new hire with an experienced buddy, and created tailored growth plans that set high standards while supporting each person’s potential.
Onboarding guide template
Standardized the onboarding path and continuously updated it based on new hire feedback.
Buddy system
Paired new hires with experienced team members to support acclimation and day-to-day guidance.
Tailored growth plans
Set high standards, unlocked growth opportunities, and provided timely feedback.
Operating model
Intake process & SLA
Seventeen UX designers supporting 160+ PMs could easily become chaotic without a shared process. I introduced an online intake workflow, automated notifications, weekly prioritization with PM leaders, and a 2-day acknowledgement SLA.
Background
17 UX designers supporting 160+ PMs, with high demand and tight project schedules.
Intake process
PMs submitted requests through an online form, with automated notifications and shared tracking.
SLA
Requests were acknowledged within 2 days and prioritized based on strategic significance.
Measurement
Metrics & measuring success
Every initiative was tied to measurable business or UX outcomes. We defined success metrics before design started, ran usability and A/B testing, and monitored data after launch.
Business metrics
Connected UX initiatives to penetration, delivery rate, aging rate, lost rate, CSAT, and NPS.
UX metrics
Measured task success, learnability, work efficiency, time on task, and usability quality.
Data-driven decisions
Used usability testing, A/B testing, and post-launch monitoring to guide design decisions.
Research practice
User research under privacy constraints
TikTok operated under strict privacy and legal restrictions, so I worked through pre-qualified participant groups with UX research and used enterprise tooling exceptions where needed. Designers also led usability research, design sprints, site visits, and internal dogfooding.
Multiple channels
Pre-qualified participant groups, UX research collaboration, and enterprise tooling exceptions.
Designer-led research
User interviews, usability studies, design sprints, and site visits for logistics projects.
Dogfooding program
Launched product and design dogfooding so the org could experience TikTok Shop like users do.
Team operations
Team routines
Weekly team meetings, monthly group meetings, and quarterly OKR alignment kept everyone pointed in the same direction. A two-level design critique process raised the bar without becoming a bottleneck, while morale rituals kept the team connected.
Team meetings
Weekly team meetings, monthly group meetings, and quarterly OKR alignment.
Design critiques
Two-level design reviews, priority-based critique requirements, and XFN participation.
Team morale
Frequent morale events, anniversary and birthday gifts, and yearly team offsites.
07 Results and learning
Platform outcomes
Results across the Customer Service Platform
CSAT
+4.21%
↑ vs previous quarter
Contacts per agent / day
+22.7%
↑ vs previous quarter
Average Handling Time
−20.31%
↓ vs previous quarter
The L1 and L2 features shipped across all seven TikTok Shop markets
US
UK
ID
MY
PH
SG
TH.
On the team side, intake SLA stabilized at 2 days, design throughput doubled relative to the previous quarter,
and we hired and ramped four new designers without a quality dip.
Key learnings
What the work reinforced
Borrow mental models
Framing AI capabilities as L0–L3 gave PMs, engineers and leadership a shared language. Half the design work in a platform org is alignment.
Friction is the signal
Recommend Replies adoption more than doubled when we moved from “click to accept” to “Tab to accept” in the composer. Tiny interaction details compound at scale.
Process scales people
Onboarding guides, buddies, intake forms and SLAs sound boring — and they’re what let a 17-person team support 160+ PMs without burning out.
© Copyright 2026 - Designed & coded by Frank Wu