Artjoker
Case Studies
AI solutions
Agentic Underwriting Automation for a Microfinance Platform

Agentic Underwriting Automation for a Microfinance Platform

AI AGENTIC DEVELOPMENT

Project Overview

A fast-growing microfinance company offering short-term consumer loans (PDL and installment) processed up to 3,000 applications a day. Underwriting relied on bureau scores, manual document review, and a rules-based scoring model — an approach that couldn’t scale, produced inconsistent results across analysts, and created compliance risk.

Artjoker was tasked with replacing the manual underwriting pipeline with an autonomous AI agent system capable of processing applications end-to-end, adjusting risk thresholds in real time, and generating audit-ready decision rationale.

47sec

average decision time — down from 28 minutes
9,500+

applications/day — new daily throughput capacity
79%

fraud caught pre-disbursement — up from ~40%

The Challenge

25–40 minutes per application

for manual document review — unacceptable for a digital product promising instant decisions.
±18% variance

in approval rate among analysts evaluating similar risk profiles.
4.1% first-payment default rate

fraud detection was purely reactive, catching losses after the fact.
Incomplete documentation in ~35% of cases

despite regulatory requirements for full rationale on every decline.
Rising bureau API costs,

with many requests returning partial or outdated data.
14 months without retraining

model drift was becoming visible in rising NPL.

The Solution

Artjoker built a fully agentic underwriting pipeline where every application moves through a chain of specialized agents:

Agent 1

Document Intelligence Agent

Receives uploaded documents (PDF, photo) and uses vision-language models to extract, validate, and cross-check data. Detects anomalies (photo manipulation, font inconsistencies, metadata mismatches) and flags them for fraud review.
Agent 2

Data Enrichment Agent

Orchestrates parallel calls to credit bureaus (UBKI, Equifax Ukraine), Open Banking API, and proprietary behavioral databases. When data is incomplete, triggers collection of alternative signals and synthesizes a unified borrower profile.
Agent 3

Risk Scoring Agent

Applies an ensemble multi-model assessment (gradient boosting + LLM-based qualitative scoring). Dynamically adjusts decision thresholds based on current portfolio composition. Retraining triggers are automated — weekly.
Agent 4

Decision & Compliance Agent

Generates a structured decision with natural-language rationale that meets regulatory requirements. For declines, automatically produces the mandatory notice. All decisions are logged to an immutable audit trail.
Agent 5

Fraud Triage Agent

Runs in parallel with the main pipeline. Detects synthetic identities, velocity patterns, and collusion signals between applicants. Suspicious applications are passed to an analyst with a ready-made briefing report.

Results

Metric	Before	After	Change
Average underwriting time	28 min	47 sec	↓ 97%
Approval rate variance	±18%	±2.3%	↓ 87%
First-payment default rate	4.1%	2.2%	↓ 46%
Compliance documentation completeness	65%	100%	↑ 35 pp
Bureau API cost per application	$0.38	$0.19	↓ 50%
Daily throughput (applications/day)	3,000	9,500+	↑ 3.2×
Fraud caught pre-disbursement	~40%	~79%	↑ 2×