intesol.kr
Cases / Finance · Fintech
MVP CASE 06

Fraud losses down 91%, underwriting 82% faster

Anomaly detection, automated loan underwriting, and a financial advice chatbot built for a digital bank over 14 months. Half the project was the documentation needed to clear regulatory review.

91%
Fraud loss reduction
82%
Faster underwriting
67%
Lower support cost
Duration
14 months
Team
4 people
Started
2024 Q4
Market
Korea digital banking
SITUATION

What was blocking them

ISSUE 01
Rule-based detection had a ceiling
Fixed-threshold rules caught new fraud patterns only three to four weeks late, and losses accumulated in between.
ISSUE 02
Underwriting was the bottleneck
Loan decisions averaged 3.4 business days, the single largest cause of application drop-off.
ISSUE 03
Explainability was non-negotiable
Financial supervision requires a record of the reasoning behind every automated decision. A black-box model could never have been approved.
APPROACH

The order we worked in

Each phase began from what the previous phase measured. We kept that order because without the earlier step there is no way to verify the next one's effect.

01
Detection model and decision records
We fixed the design for structured reasoning records alongside each anomaly score before anything else. Audit readiness was the starting point, not an afterthought.
5 months
02
Automated underwriting pipeline
Document reading and credit checks were automated, with borderline applications passed to an underwriter together with the reasoning.
4 months
03
Financial advice chatbot
Product explanations were limited to answers grounded in regulatory documents, and anything readable as investment solicitation is blocked at generation.
3 months
04
Regulatory documentation and audit
Decision-log structure, model change history, and exception procedures were documented for internal audit and supervisory reporting.
2 months
RESULTS

Baseline at kickoff, measured after delivery

Metric
Before
After
Losses from fraud
baseline
→
-91%
Loan decision time
3.4 business days
→
7.3 hours
Customer support cost
baseline
→
-67%
New fraud pattern detection
3–4 weeks
→
within 2 days

How each figure is measured is set out in the measurement notes on the cases page. Revenue and cost figures come from the client's own accounting and are published only within the scope they approved.

“Proposing that we design the audit trail before tuning the model is what kept the approval schedule.”
Finance · Fintech · Head of risk management
Anomaly detection Automated underwriting Financial chatbot Decision logging Regulatory documentation
Free · 1 business day
This one started with a diagnostic too
The 6-Week MVP starts by measuring your baseline. Find out where you stand first — the diagnostic is free.