South KoreaFinance and insurance2024 – ongoingToss Bank
ML identity-document checks and capture feedback
Toss Bank reported about 1.5–2% lower subsequent identity-document rejection with ML capture feedback
Problem
Manual identity-document checks grew with customer numbers. Rejected photographs forced customers to submit another image and wait for verification.
Approach
Toss Bank deployed image-classification and object-detection models to score document quality. The system rejected clear failures and flagged uncertain cases for staff, who retained final approval. The bank reused the models to tell customers what to fix while taking photographs and selected thresholds through an A/B test.
Results
Takeaway
Earlier feedback reduced later rejection without a substantial change in capture-stage pass rates. The bank did not specify whether its reported percentage reduction meant relative change or percentage points.
Categories
- Size
- Large enterprise
- Industry
- Finance and insurance
- Business function
- Finance, risk and legalCustomer service
- Innovation type
- Process innovation
- AI technology
- Computer vision
- Adoption
- Built in-house
- Evidence
- Company's own claim
Sources
Source summaries
A Toss Bank engineer explained deployed document-quality models and capture feedback. The winning test variant lowered subsequent rejection by approximately 1.5–2%, with little difference in capture-stage pass rates.
Full details are in the original sources.