AI Product Design & Innovation

LandCheck

AI-assisted land registration document review
Chu Hoai Linh · AI Product Design & Innovation
Proposed organization: Hanoi Land Registration Office
Less repetitive checkingClearer evidenceHuman decisions
⚠ Potential mismatch
✓Officer decides
Chu Hoai LinhProposed design · not a deployed service
The business problem

Repeated checking creates avoidable effort

Read documents Compare fields Draft corrections Recheck × N EVERY CASE, BY HAND
⟲

Officers repeat the same work

They read documents again and again, compare fields across files and draft correction instructions by hand.

?

Applicants need clarity

They need clear explanations of what information is missing or inconsistent — and how to fix it.

These are scenario-based challenges, not measured findings from a branch.
Chu Hoai LinhProposed design · not a deployed service
A focused design opportunity

One procedure, one controlled workflow

LANDCHECK SCOPE · ONE PROCEDURE STARTDocuments ready forpreliminary checking PROCEDURE IN SCOPE Land-use-rights change aftera transfer between individuals ENDOfficer records outcome& approves correction draft Automated legal validation Authenticity certification Official approval BLOCKED AT THE BOUNDARY
Chu Hoai LinhProposed design · not a deployed service
Users & design thinking

Research before committing to a solution

L

Ms. Lan

Illustrative officer persona
Needs fast evidence retrieval and clear controls.
M

Mr. Minh

Illustrative applicant persona
Needs understandable correction instructions.
5
officer interviews
2
supervisor interviews
5
applicant interviews
👁
authorized observation
Proposed research. Personas and pain points remain hypotheses — no fieldwork has been completed.
Chu Hoai LinhProposed design · not a deployed service
Current-state journey

Locate the friction in each step

1

Receive documents

FRICTIONUnclear labels
OPPORTUNITYSuggest document categories
2

Check completeness

FRICTIONRepeated checklist review
OPPORTUNITYLink files to approved rules
3

Compare information

FRICTIONSwitching across files
OPPORTUNITYSide-by-side source evidence
4

Prepare corrections

FRICTIONRepeated drafting
OPPORTUNITYUse confirmed findings
5

Recheck

FRICTIONUnclear changes
OPPORTUNITYPreserve version & issue history
Chu Hoai LinhProposed design · not a deployed service
Ideation & selection

Choose the solution that fits the problem

Digital checklist

Predictable, but limited cross-document support.

PARTIAL FIT

Citizen chatbot

Useful upstream, but indirect impact on officer review.

INDIRECT
SELECTED

Officer-facing review assistant

Evidence links and human controls, built around the officer's review.

BEST FIT ✓

Automated approval

Outside the scope and the accountability boundary.

OUT OF SCOPE ✕
Chu Hoai LinhProposed design · not a deployed service
Future-state workflow

AI prepares; officers validate

◆ AI
01Upload documents and suggest their categories
02Extract selected fields and link them to source passages
03Run approved completeness & consistency checks
05Draft instructions from confirmed issues
● Officer
01Confirms the categories
04Confirms, edits, rejects or escalates findings
05Approves before anything is communicated
Chu Hoai LinhProposed design · not a deployed service
Solution architecture

Use AI selectively; keep requirements rule-based

INPUTSPROCESSINGOUTPUTS Authorized documents Field definitions Versioned checklists Approved templates OCRAI Field extractionAI Rules & comparisonsRULES Evidence-linked findings AI READS · RULES DECIDE WHAT IS REQUIRED Officer review workspace Approved summary Audit history Manual reviewNEVER AN INVENTED ANSWER UNREADABLE / UNSUPPORTED
Chu Hoai LinhProposed design · not a deployed service
Prototype demonstration

Synthetic case DEMO-001

LandCheck · Review workspaceCASE DEMO-001 · synthetic
Documents
ApplicationPDF
Supporting files3
Checklistv2
Historylog
Source evidence
APPLICATION
DEMO-1234
SUPPORTING FILE
DEMO-1284
Issues
⚠ POTENTIAL MISMATCH

Identifier differs between the application and the supporting file. Inspect the highlighted original passages.

ConfirmRejectCorrect extractionManual review
A mismatch is a verification prompt, not proof of legal invalidity.
Chu Hoai LinhProposed design · not a deployed service
MVP & prioritized backlog

Prove the smallest useful workflow

MUSTMVP
Secure upload
Supported extraction
Approved checklist & evidence-linked alerts
Officer review controls
Audit record & correction draft
EXCLUDENEVER IN MVP
Autonomous messages
Official-record changes
Authenticity decisions
Automatic approval
Chu Hoai LinhProposed design · not a deployed service
Agile delivery

Test assumptions and revise the backlog

Small increment
User test
Classify failures
Reprioritize
Regression test
IterateEVERY INCREMENT
TESTIs comparison a major time cost?
TESTCan supported documents be extracted reliably?
TESTCan officers verify alerts efficiently — without overreliance?
Academic week deliverables
ScopeResearch planJourney mapPrototypeTest planReport
The academic deliverable is not a production-ready system.
Chu Hoai LinhProposed design · not a deployed service
Controlled pilot

Small, supervised, measurable

1
branch
~5
officers
≤100
cases
4
weeks, after authorization & preparation
WEEK 1

Training & baseline

WEEK 2

Shadow mode

WEEK 3

Supervised use

WEEK 4

Evaluation & revision

ENTRY GATESApproved data handling · Procedural validation · Offline testing
FAIR COMPARISONCompare cases by complexity and scan quality; reduce allocation bias where practical.
Chu Hoai LinhProposed design · not a deployed service
Five evaluation metrics

What success would look like

Illustrative thresholds — not achieved results

Review time

≥20%

Reduction in median total officer time, including verification.

Issue recall

≥90%

Detected reference issues ÷ all reference issues.

Alert precision

≥85%

Correct alerts ÷ all alerts.

Traceability

100%

Correct, accessible evidence for every displayed finding.

Independent verification

ALL

Critical seeded AI errors identified in controlled tests.

Chu Hoai LinhProposed design · not a deployed service
Governance, adoption & value

Safety and net benefit, assessed together

🔒

Controls

Restricted access, approved storage, retention rules, versioning and manual fallback.

⚖

Ownership

Procedural specialists approve rules; officers keep decision responsibility.

✦

Adoption

Co-design, training on limitations, pilot champions and feedback.

500CASES / MONTH
×
6NET MIN SAVED
=
50hRELEASED MONTHLY
Hypothetical. Released time is not automatically cash savings — account for operating and verification costs.
Chu Hoai LinhProposed design · not a deployed service
Conclusion
AI prepares the evidence.
Officers make the decisions.
01 · START

One procedure and an evidence-linked prototype.

02 · EVALUATE

Efficiency alongside accuracy and independent judgment.

03 · SCALE

Only after an authorized pilot meets agreed quality and governance gates.

AI disclosure: generative AI supported concept development, structure and drafting.
Source documents: Chu-Hoai-Linh.docx and Individual_Projects.pdf. No fieldwork or pilot completed.
Chu Hoai LinhProposed design · not a deployed service
Q&A appendix

Questions to prepare for

?Why not only a checklist?

Extraction supports varied layouts; approved rules still govern requirements.

?What if AI is wrong?

Inspect sources, reject findings and use the manual fallback.

?How do rules stay current?

A named procedural owner approves versioned updates.

?What about difficult cases?

Escalate outside the MVP to an authorized officer.

Chu Hoai LinhProposed design · not a deployed service
Pilot decision gates

Proceed, iterate or pause

REVISE & RE-TEST Pilot evaluation

PROCEED

Quality thresholds met, independent checks effective and net time savings demonstrated.

ITERATE

Excessive false alerts, confusing evidence links or high editing effort.

PAUSE

Sensitive-data incident, unsupported legal output or critical oversight failure.

Additional procedures require separate rules, evaluation and approval.
Chu Hoai LinhProposed design · not a deployed service
← → / space · F fullscreen