External
Amounts, due dates, and details must be found and copied by hand, and total spending is hard to see at a glance.
Case Study · Dear Me
Scattered statements,
one ledger that never loses its sources.
A real household ledger that helps you understand and manage statements and bills arriving by text, paper, and email — all in one place.
7월 지출
2,149,180원
이번 달 요약
쇼핑1,494,380원
생활600,000원
식비30,900원
✎ 직접 적기📷 영수증 찍기
최근 내역
쿠팡(주) 쇼핑 · 7.27-7,400원
테스트상점 기타 · 7.26-2,000원
월세 생활 · 7.26-600,000원
Status
Prototype
Role
Planning · Information architecture · Product design
Users
People who manage their own living expenses and bills
Updated
2026-07
Shared Context
The user's situation
Card statements arrive by text, telecom bills by email, maintenance fees on paper. Every month, users must find, re-read, copy, and double-check all of it.
| Layer | What Dear Me users want |
|---|---|
| Surface need | Collect text, email, and paper statements in one place and see amounts, due dates, and items. |
| Inner need | Stop worrying about missed or duplicated records, and understand where the money goes each month. |
| Latent need | Own the evidence and control of my own expense records, without depending on one bank or channel. |
Problem
The villain and three dimensions of the problem
The villain
The structure that scatters statements across different channels and formats
Internal
Anxiety about missed bills, and no confidence that records match the originals.
Philosophical
People should be able to understand their own money without chasing complicated documents.
Options & Objections
Alternatives compared
Dear Me doesn't dismiss other methods. After comparing each alternative's strengths and limits, it chose "statement inbox + AI structuring."
| Alternative | Strengths | Limits | Judgment |
|---|---|---|---|
| Manual ledger / spreadsheet | Full freedom and direct control of structure | Input burden, omissions, hard to trace originals | Baseline for comparison |
| Bank/card auto-sync apps | Collect transactions quickly | Can't cover text/paper/email statements, non-financial items, or details | Secondary data source |
| Statement inbox + AI structuring | One data model across channels and formats, linked to originals | Extraction errors, duplicates, sensitive-data risk | Dear Me's chosen direction |
| Fully automatic confirmation | Minimal user input | Wrong amounts or categories could be silently confirmed | Rejected: review step required |
Plan
The user's plan: a six-step flow
Collect
Capture texts, forward emails, photograph paper statements — everything lands in the Dear Me inbox.
Read
AI extracts the issuer, amount, period, due date, line items, and their location in the original.
Verify
See the original and extracted values side by side; edit, delete, or split.
Organize
Duplicate candidates are suggested; group by category, month, and payment status.
Understand
Monthly ledger, upcoming payments, per-category changes — with links to the originals.
Take it with you
Export any period; manage storage, deletion, and permissions yourself.
Solution · Information Architecture
Key screens and information architecture
| Screen | Content | Design principle it proves |
|---|---|---|
| Inbox | New statements, processing status, sources, error/duplicate alerts | Scattered inputs gathered in one place |
| Review | Original and extracted fields side by side; edit and confirm | Explainability and user control |
| Ledger | Monthly income, spending, upcoming payments, categories | Living expenses made understandable |
| Statement detail | Issuer, period, total, items, original, change history | Evidence and traceability |
| Source archive | Store/delete originals from text, email, paper | Never lose the original |
| Privacy settings | Storage location, permissions, export, delete, disconnect | Ownership of personal data |
Data field design
Source: text / email / paper / manual entry
Issuer · service name · account alias
Statement period · issue date · due date · payment status
Total · tax · discounts · line items · category
Original file · location in original · extraction confidence · user edit history
Duplicate group · linked transactions · memo · tags
AI vs User Control
What AI does, what users decide
AI assists
- Finding document regions, extracting text, mapping fields
- Suggesting categories, duplicate candidates, payment status
- Proposing monthly summaries and anomaly candidates
- Flagging low-confidence items
Users decide
- Verifying originals, editing values, final confirmation
- Merging, splitting, changing categories
- Interpreting meaning, taking action, deleting or archiving
- Choosing what auto-confirmation is allowed to do
Privacy & Trust
Privacy and trust principles
Open architecture
Because this is personal financial data, storage location, encryption, third-party transfer, and retention are published with the architecture.
Confidence shown
AI-extracted values show their source link and confidence; low-confidence values are never auto-confirmed.
Export & delete rights
Users can export and delete originals, structured data, and edit history at any time.
Anonymization
Example screens use samples with names, addresses, accounts, and amounts anonymized.
Honest labeling
Until the implementation is confirmed, claims like "local processing" or "fully encrypted" are never written as completed facts.
Evidence & Success Metrics
Evidence and success criteria
Unverified goals are never dressed up as numbers; what will be measured is published first.
| Evidence asset | Success criteria to measure |
|---|---|
| Anonymized statement set & field schema | Range of supported formats and types of extraction misses |
| Original ↔ extraction review screen | Edit rate; steps and time to confirmation |
| Before/after duplicate suggestions | Duplicate detection accuracy; preventing wrong merges |
| Monthly ledger prototype | Whether users understand total spending and upcoming payments |
| Personal data flow diagram | Clarity of storage, deletion, and export paths |
| User testing | Reduced anxiety about misses, trust in tracing, intent to use |
Status
Labeled "Prototype" because working screens and flows exist. It will move to MVP/Shipped once real users use it repeatedly.
Next Step & Research
Next steps and research questions
Next
Finalize the anonymized sample set and field schema
Build a sample set covering diverse statement formats, then run user tests on the extraction and review UI prototype.
Research
Document understanding and source preservation
How should AI show errors and sources while structuring diverse personal documents? — continued on the Research page.