01 PROBLEM STATEMENT
The Bottleneck Slowing Every Loan Decision
A leading private sector bank processing 8,000+ retail loan applications per month depended entirely on manual document handling. Loan officers read every KYC form, income proof, and property paper by hand, then re-keyed data into their IBM i DB2 Loan Origination System. Errors only surfaced days later at credit review. Application volumes were growing 18% year-on-year. Headcount was not.
02 current challenges
Loan Document Processing Challenges
2-3 Days
Validation cycle
Applications sat in batches, rarely reaching credit review on the same day.
12-18%
Rework rate
Transcription errors caught only at credit review, restarting the entire cycle.
Zero Trail
Audit evidence
No structured record of document decisions. Two RBI audits flagged this as a risk.
Varied
Classification standard
Each underwriter applied different rules, creating inconsistency across hubs.
03 SOLUTION OVERVIEW
STAR’s Approach – AINE Loan Document Intelligence
STAR Systems and IBM Technology Expert Labs deployed an AI pipeline that automatically reads, classifies, and extracts data from every loan document, then writes validated records directly into the bank’s IBM i DB2 LOS via a standard JDBC adapter. No migration. No LOS changes. Live in 8 weeks.
AI PATTERN
RAG + Vision + Extraction
04 WORKFLOW PROCESS
Step-By-Step: How a Loan Document Moves Through AINE
Step 1 – Document Received: Loan file arrives via portal, email, or branch scan in any format.
Step 2 – Classify: Gemini Vision identifies document type automatically KYC, income proof, property paper, etc.
Step 3 – Extract + Score: With AI confidence scores attached to every extraction.
Step 4 – Validate + Flag: Low-confidence fields flagged for targeted officer review. Rest proceed automatically.
Step 5 – LOS Write: Clean, validated data written to IBM i DB2 via JDBC. No manual keying.
Step 6 – Audit Log: Every decision stored with timestamp, model version, and confidence score for RBI.
05 KEY FEATURES
What the Platform Does
Mixed-Format Document Reading:
Handles scanned paper, digital PDFs, and phone photographs in one pipeline no pre-sorting by staff.
Confidence-Scored Extraction:
Every field extracted with a confidence score. Borderline cases flagged for human review without stopping the application.
12+ Document Types Automated:
KYC, income proofs, bank statements, property papers, insurance certificates, co-applicant declarations and more.
Zero-Migration LOS Integration:
Connects to IBM i DB2 via standard JDBC adapter. No changes to the LOS. No disruption to live processing.
Full Audit Trail:
Every extraction decision logged with timestamp, model version, and confidence score. Fully explainable to RBI.
Quarterly Model Retraining:
Bank’s team labels edge cases. Model retrained on bank-specific data every quarter. STAR manages L2/L3 support.
06 BUSINESS OUTCOMES
What Changes After Go-Live
70%+
Reduction in validation time
40-60%
Processing cost reduction per loan
85%+
Reduction in data entry errors
CFO
- 40-60% cost reduction per loan.
- Staff redeployed from document handling to credit decisioning.
Operations Head
- Validation drops from 2-3 days to under 2 hours.
- 100% audit trail for every compliance reviewer.
CXO / Risk
- Uniform classification across all hubs.
- Every decision explainable and defensible to RBI.
07 REAL-WORLD SCENARIO
A Day in the Life – Before and After
| Before | After |
|---|
| Customer asks about eligibility for a new MSME loan scheme. Teller searches SharePoint for 20 minutes, cannot find the guideline, escalates to branch manager. | Teller asks the assistant ‘MSME loan eligibility criteria’. Answer appears in 10 seconds with citation to the official product guideline. Customer served on the spot. |
| Branch operations officer needs the SOP for foreign remittance documentation. Searches file share and email archives. Finds an outdated version from 2 years ago. | Officer asks ‘foreign remittance SOP’. Assistant retrieves the current version updated last month. Operational error avoided. |
| New relationship manager joins the branch. Spends first 3 weeks reading through hundreds of policy documents with no structured way to learn on the job. | New RM uses the assistant to ask policy questions as they arise during customer interactions. Onboarding time cut significantly. RM productive faster. |
| Regional ops team receives 50+ policy query escalations per week from branches. Response time averages 2 days. Staff and customers both wait. | Escalation volume drops by 80%. Most queries resolved instantly by the assistant. Ops team handles only genuine edge cases and policy clarifications. |
08 ROI AND VALUE JUSTIFICATION
Why this Numbers Work
| Value Driver | Indicative Impact | How It Is Realised |
|---|
| Policy query resolution time | 2 days reduced to 2 minutes | Branch staff get instant cited answers instead of escalating to regional ops. More time with customers, less time searching. |
| New staff training cost | Significantly reduced | Assistant bridges knowledge gaps during onboarding. New hires productive faster with on-demand policy and SOP access. |
| Operational error rate | Reduced through accurate, cited answers | Staff act on correct, up to date policy information instead of outdated documents or memory. Fewer compliance and service errors. |
| Positive ROI timeline | Within 6 months of go-live | Staff productivity gain and training cost reduction across 400+ branches exceeds platform and managed service costs within one quarter. |
09 NEXT STEPS
01
Discovery Call
30-min call to map your collections workflow, CBS setup, and DPD distribution.
02
Pilot Scoping
Identify 2–3 DPD cohorts from your portfolio for a structured pilot campaign.
03
Pilot Delivery
AI scoring and outreach workflows run on your data before commercial rollout.
04
Business Case
Recovery uplift, operational savings, and ROI measured from pilot outcomes.
