Finance · AI Transformation · Conversational Design
Strategy meets AI in financial services.
I turn complex finance workflows into intelligent, self-service experiences — combining competitive strategy, data analysis, and conversational AI design.
About
Strategy foundation, AI-native execution.
I'm a Fashion Marketing & Content Creation student at the University of the Arts London (Class of 2027, GPA 3.9, graduating July 2027), with a strategy-consulting foundation built at Landor (WPP), where I worked on competitive analysis in the financial payments sector (UnionPay vs. Alipay & WeChat Pay).
I use AI tools daily — Claude, ChatGPT, Nation AI — to compress research and reporting cycles, and I'm now going deeper: designing conversational AI agents for finance operations. My strengths are structured problem-solving, cross-functional coordination, and translating messy business needs into clear, deployable plans.
I'm looking to apply this at the intersection of finance and AI transformation.
Featured Project
Supplier Payment Inquiry — AI Agent Design
Self-initiated concept project · End-to-end conversational agent design
The problem
Finance / Accounts Payable teams spend a large share of their time answering the same low-complexity questions: "When will invoice #X be paid?", "Why is my payment on hold?", "What documents are still missing?" These inquiries are high-volume, repetitive, and time-sensitive — an ideal candidate for an AI assistant that lets suppliers and internal requesters self-serve.
My objective
Design an AI agent that resolves routine supplier-payment inquiries without human intervention, targeting a 70%+ self-service resolution rate while routing genuine exceptions to a human cleanly.
1 Stakeholder & needs analysis
I mapped the three groups who generate inquiries and what each needs:
| User | Typical question | Underlying need |
|---|---|---|
| Internal procurement requester | "Has my supplier been paid yet?" | Status visibility, deadline confidence |
| Supplier (external) | "When will I receive payment for invoice #1042?" | Cash-flow certainty |
| AP / Finance staff | "Can the bot stop forwarding me these?" | Deflection of routine load |
2 Knowledge organization
I structured the knowledge base the agent needs to answer accurately:
Payment status definitions
Received → Under review → Approved → Scheduled → Paid → On hold
Payment cycle & SLA
Standard net-30 / net-60 terms and cut-off dates
Common hold reasons
Missing PO, price mismatch, incomplete bank details, pending approval
Required documents
Invoice, PO, delivery confirmation, tax forms
Escalation paths
When and to whom the agent hands off
3 Conversation flow design
The agent follows a branching dialogue:
flowchart TD
A[User opens chat] --> B{Identify user type}
B --> C[Ask for invoice or PO number]
C --> D{Record found?}
D -- No --> E[Apologize + offer to log a ticket / escalate]
D -- Yes --> F{Payment status?}
F -- Paid --> G[Show payment date + amount + reference]
F -- Scheduled --> H[Show expected pay date + SLA note]
F -- On hold --> I[Explain hold reason + required action]
I --> J{User can resolve?}
J -- Yes --> K[Guide self-service fix]
J -- No --> E
G --> L[Anything else? / Close]
H --> L
K --> L
4 Try the agent
A working prototype. Ask about a mock invoice — or request a human to see the escalation path.
Connects to a live Claude-powered agent (Haiku 4.5) when deployed with an API key; falls back to built-in rule-based replies otherwise. Mock data only.
5 KPI framework (success measurement)
| KPI | Definition | Target |
|---|---|---|
| Self-service resolution rate | Inquiries closed without human handoff | ≥ 70% |
| Answer accuracy | Correct status/info vs. system of record | ≥ 95% |
| Escalation rate | Inquiries routed to a human | ≤ 30% |
| Avg. handle time saved | Finance-staff minutes saved per inquiry | benchmark vs. baseline |
| User satisfaction (CSAT) | Post-chat rating | ≥ 4.2 / 5 |
6 User Acceptance Testing (UAT) plan
Participants
2 AP staff, 3 internal requesters, 2 pilot suppliers
Scenarios tested
Paid invoice, scheduled invoice, on-hold with resolvable cause, on-hold needing escalation, invalid invoice number
Pass criteria
≥ 90% of scenarios return correct status; 0 incorrect "paid" confirmations (zero-tolerance); clean human handoff in all escalation cases
Feedback loop
Log misunderstood queries → refine intents and knowledge base → re-test
What this demonstrates: the exact skill set the role asks for — knowledge organization, conversational workflow design, KPI definition, and UAT ownership.
Parallel Module
Internal Expense Reimbursement — AI Agent Design
Self-initiated concept project · End-to-end conversational agent design
The problem
Employees flood finance with repetitive questions: "What's the status of my reimbursement?", "Why was my claim returned?", "What's the meal allowance?", "Which receipts do I need?" High-volume, low-complexity — an ideal candidate for self-service.
My objective
Resolve routine reimbursement inquiries without a human, targeting a 70%+ self-service resolution rate with clean handoff for real exceptions — and raise the first-time-right submission rate by answering policy questions before people submit.
1 Stakeholder & needs analysis
I mapped the three groups who generate inquiries and what each needs:
| User | Typical question | Underlying need |
|---|---|---|
| Employee submitter | "What's the status of my reimbursement?" | Status + policy clarity |
| Manager approver | "Anything to fix before I approve?" | Faster, cleaner approvals |
| Finance / AP | "Can routine questions stop reaching me?" | Deflection of routine load |
2 Knowledge organization
I structured the knowledge base the agent needs to answer accurately:
Claim status flow
Submitted → Pending manager approval → Finance review → Approved → Paid; plus Returned (fixable) and Rejected
Policy quick facts
Meals up to ¥150/person/day; intercity travel over ¥2,000 needs prior approval; a valid fapiao is required; submit within 60 days
Common return / reject reasons
Missing fapiao, over daily limit, duplicate submission, out-of-policy category, late submission
Required documents
Itemized fapiao / receipt, approval for amounts over ¥2,000, trip purpose for travel
Escalation path
To a finance specialist
3 Conversation flow design
The agent follows a branching dialogue:
flowchart TD
A[Employee opens chat] --> B{Claim ID or policy question?}
B -- Claim ID --> C[Look up claim]
C --> D{Claim status?}
D -- Pending approval --> E[Show approver + ETA + offer reminder]
D -- Finance review --> F[Show in review + ETA]
D -- Paid --> G[Show paid date + amount + reference]
D -- Returned/Rejected --> H[Explain reason + guide the fix]
H --> I{Resolved?}
I -- No --> J[Escalate to finance specialist]
B -- Policy question --> K[Answer from knowledge base]
E --> L[Anything else? / Close]
F --> L
G --> L
K --> L
4 Try the agent
A working prototype. Ask about a mock claim, ask a policy question, or request a human.
An interactive rule-based prototype with mock data, demonstrating the dialogue design above.
5 KPI framework (success measurement)
| KPI | Definition | Target |
|---|---|---|
| Self-service resolution rate | Inquiries closed without human handoff | ≥ 70% |
| Policy answer accuracy | Correct policy answers vs. the handbook | ≥ 95% |
| First-time-right submission rate | Claims accepted without a return | ↑ vs. baseline |
| Escalation rate | Inquiries routed to a specialist | ≤ 30% |
| Approval cycle time | Avg. time from submit to approve | ↓ vs. baseline |
| User satisfaction (CSAT) | Post-chat rating | ≥ 4.2 / 5 |
6 User Acceptance Testing (UAT) plan
Participants
2 finance staff, 3 employees, 1 manager
Scenarios tested
Pending approval, returned-fixable, paid, policy question (meal cap), invalid claim ID, escalation
Pass criteria
≥ 90% correct status; 0 wrong "paid" confirmations; clean handoff on all escalations
Feedback loop
Log misunderstood queries → refine intents → re-test
What this demonstrates: the same delivery pattern applied to a second finance function — knowledge organization, conversational flow, KPI definition, and UAT — showing these AI assistants generalize across the finance org.
Experience
Where I've applied AI & strategy.
Landor (WPP) — Strategy Consulting Intern
Shanghai · May–Aug 2025
- Conducted AI-assisted competitive research (Claude, ChatGPT, Nation AI) on the financial payments sector, cutting desk-research and reporting time significantly.
- Built a structured competitor database and comparison framework; produced quantitative gap analysis and heatmaps benchmarking UnionPay against Alipay and WeChat Pay.
- Synthesized findings into a strategy report (brand positioning, user insight, optimization recommendations) adopted by the UnionPay project team as a core reference for its 2026 brand strategy.
- Used AI tools (Claude, ChatGPT, Nation AI) to compress research and reporting cycles, and coordinated across Landor brand-design and WPP Marketing teams on the Huawei brand-renewal project.
Relevance — finance-domain analysis + applied AI + structured insight delivery
Credential — recommendation letter from Landor (WPP) China General Manager
UALCSSA — President
London · Feb 2025–present
- Led cross-functional coordination across 4 teams to deliver 17 large-scale events (100% execution rate).
- Negotiated and managed partnerships (Haidilao, Panda, EasyTransfer, London Square) — independent proposals and long-term stakeholder relationships.
Relevance — core liaison / cross-functional communication
3AM Entertainment — Marketing Intern
London · Sep 2024–Apr 2025
- Produced 13 creative proposals and ran offline events (gallery, KTV, bar, product launches) that grew the community by 650+ members and yielded reusable execution templates.
- Partnered with 200+ London KOLs / KOCs to boost brand exposure across the target market.
- Managed campaign budgets with tight venue and material cost control, lifting operational ROI by 16%.
Relevance — data-driven decision-making and performance tracking
Nanjing Dahui Group (Nanjing Impressions) — Business Development Intern
Nanjing · Jul–Aug 2024
- Analyzed food-sector trends and sales data across Douyin, Kuaishou, Video Accounts, and Taobao; flagged category-fit risks from influencers' track records to avoid ineffective partnerships and save on slot fees.
- Liaised with influencers and MCN agencies — supporting negotiation, scheduling, and client relationships — to land multi-platform partnerships efficiently.
Relevance — data-driven partner selection + business negotiation
Skills
Toolkit.
AI & conversational design
Data & analytics
Strategy
Coordination
Design & presentation
Contact