A Ho Chi Minh City fintech cuts KYC time 91% and fraud loss 65% with an explainable trust stack

Context

Your Ho Chi Minh City fintech runs a digital-wallet and lending platform serving 2.8M consumers and 180,000 merchants across Vietnam. Customer onboarding ran a manual KYC review on 11% of applications flagged by basic rules: documents in non-Vietnamese script, mismatched selfies, or addresses in flagged provinces. The 90-person verification team handled the queue, with average time-to-onboard of 47 hours on flagged applications and 22% drop-off during the wait. Fraud loss on the lending product had risen from 1.4% to 2.6% of disbursed volume in 18 months, costing approximately VND 124B (US$5.1M) per year. The State Bank of Vietnam had also tightened reporting requirements on AML transaction monitoring, adding a compliance backlog the team could not clear.

Challenge

Three constraints. First, identity-document quality varied wildly: 14 acceptable ID formats, smartphone photo conditions ranging from professional to unusable. Second, the State Bank required explainability on any automated KYC decision, and the existing rules engine was already at the limit of what regulators would accept without case-by-case review. Third, fraud losses concentrated in two of the lending products, but the team did not know which features mattered most because a 2023 vendor "AI fraud" deployment had been a black box.

Approach

We ran a 4-phase model: foundation, KYC, fraud, hand-over. Foundation (4 weeks) built a unified identity-and-risk data store, ingesting application data, document images, device telemetry, transaction history, and the State Bank reporting feed. We named the methodology the AIMenta Trust Stack and produced a one-page framework the Chief Compliance Officer signed off in week three.

KYC (8 weeks) deployed a document-verification model handling all 14 ID formats with explicit confidence scoring. Below 0.85 confidence, the application routed to a human reviewer with the model reasoning attached. Above 0.85, the application auto-approved with the model rationale logged. Selfie liveness used an on-device check before server-side comparison.

Fraud (10 weeks) built a transaction-fraud model on the lending product, with feature attribution visible to every analyst. The model produced a risk score with the top three contributing features named in plain Vietnamese. Analysts could agree, override, or escalate. Every override fed back into retraining. Hand-over (parallel from week 16) trained two of your platform engineers and the head of compliance on model retraining, threshold calibration, and the State Bank reporting pipeline.

Results

Time-to-onboard on flagged applications fell from 47 hours to 4 hours 20 minutes, a 91% reduction. Drop-off during the wait fell from 22% to 4%, recovering an estimated 168,000 onboardings per year worth approximately VND 96B (US$4M) in lifetime contribution. Fraud loss on the lending product fell from 2.6% to 0.9% of disbursed volume, saving approximately VND 78B (US$3.2M) per year. Verification team throughput rose 3.4x at the same headcount; the team reallocated 28 people to AML transaction monitoring, clearing the State Bank backlog within six weeks.

The State Bank compliance review accepted the explainability framework without modification. Two new lending products launched on the same trust stack within four months of hand-over, with no AIMenta involvement.

Lessons

The 0.85 confidence threshold (auto-approve above, human-review below) was the design decision that made the regulator comfortable, and was the lever everything else hung on. Plain-Vietnamese feature attribution turned the fraud model from a black box into a tool analysts wanted to use. Reallocating freed verification capacity to AML cleared a separate compliance crisis at zero additional cost.