Credit decisioning rebuild: с manual underwriting к automated retail decisions

Контекст

Розничный банк УЗ. Consumer loan approval 5 days average. Underwriter team 30 человек, capacity bottleneck. NPL ratio 4.2% (industry 2.8%). Конкуренты-finтехи дают instant decisions.

CRO + Head of Lending — rebuild credit decisioning.

Диагностика (5 недель)

• Retail underwriting — manual, inconsistent quality across underwriters.
• Models exist, но not in production decision (advisory only).
• Data sources fragmented — credit bureau call manual.
• Audit trail incomplete для regulator.
• NPL by segment: highest у underbanked first-time borrowers (12%).

Подход (10 месяцев)

Месяцы 1-3. Architecture + decisioning platform selection.

Месяцы 3-5. Model rebuild. PD, LGD, EAD modes для retail. ML-augmented с traditional bureau data.

Месяцы 5-7. Decisioning platform launch. Real-time scoring + policy engine + automated decision.

Месяцы 7-9. Migration retail loan application к platform. Underwriter team transition к exception cases only.

Месяцы 9-10. Optimization, false positive tuning, edge case workflow.

Результаты

После 10 месяцев:

• Retail approval time: 5 days → 12 seconds.
• Approval rate: 38% → 52% (better discrimination, не permissiveness).
• NPL retail rate: 4.2% → 2.9% (12-month cohort).
• Underwriter team capacity freed: 70% redirected к SME / corporate.
• Customer NPS у retail loan applicants: +27.
• Regulator audit: passed first try.

Что критически важно

Model bias monitoring обязательно — disparate impact analysis caught early issues.

Underwriter team buy-in — they participated в edge case design.

Regulator dialogue continuous — not surprise.

Champion/challenger live — model continuously improved.

Связанное

• /solutions/banking-credit-decisioning-platform/ — credit platform
• /architecture/banking-mlops-architecture/ — MLOps
• /insights/banking-credit-risk-management/ — credit risk
• /expertise/banking-credit-discipline/ — credit discipline