Random sampling is the wrong strategy for quality control in lending — not because sampling is inadequate in principle, but because lending file risk is not randomly distributed. A loan originated by an experienced RM with a clean error history, on a standard product, with a prime borrower, from a reliable DSA, is statistically unlikely to contain a material error. A loan originated by a new RM, on a complex product, with a borderline credit profile, from a DSA with elevated exception rates, is statistically very likely to contain at least one error that needs to catch before disbursement. A random sample treats these two loans with equal review probability, which means it under-reviews the high-risk files and over-reviews the low-risk ones. The Quality Control Agent AI applies a risk-stratified sampling strategy — concentrating human QC attention where errors are most likely, and using automated review for the files where the error probability is low enough that automated checks are sufficient.
Random sampling is the wrong strategy for quality control in lending — not because sampling is inadequate in principle, but because lending file risk is not randomly distributed. A loan originated by an experienced RM with a clean error history, on a standard product, with a prime borrower, from a reliable DSA, is statistically unlikely to contain a material error. A loan originated by a new RM, on a complex product, with a borderline credit profile, from a DSA with elevated exception rates, is statistically very likely to contain at least one error that needs to catch before disbursement. A random sample treats these two loans with equal review probability, which means it under-reviews the high-risk files and over-reviews the low-risk ones. The Quality Control Agent AI applies a risk-stratified sampling strategy — concentrating human QC attention where errors are most likely, and using automated review for the files where the error probability is low enough that automated checks are sufficient.
Why risk-stratified sampling outperforms random sampling for lending QC
The operational argument for random sampling — that it is statistically valid and unbiased — is correct but incomplete. Statistical validity requires that the population being sampled is homogeneous, which lending files are not. A 15% random sample of a 500-file population will, by the central limit theorem, produce an accurate estimate of the population error rate. What it will not do is concentrate review effort on the files where errors are most consequential. A high-risk file with a potential compliance finding that escapes the 85% of files that were not reviewed creates a problem that a 15% sample rate estimate cannot prevent — it can only describe, retrospectively, how common such problems were.
The Quality Control Agent AI's risk stratification uses six signals to assign each file to one of three review tiers: Tier 1 (full automated + human review), Tier 2 (full automated review, human review if AI flags), and Tier 3 (automated spot-check, human review only if AI flags a critical error). The six signals are: originating RM error history, DSA exception rate in the last 90 days, product complexity, borrower credit profile volatility, loan amount relative to product segment average, and the presence of any exception approval in the file. Every file is assessed on all six signals at the point of origination, and its tier assignment is updated if any signal changes materially before disbursement.
"A 15% random sample tells you that approximately 8% of your files have errors. A risk-stratified 100% review tells you exactly which 8% have errors and what those errors are."
The three review tiers: criteria, coverage, and human involvement
Risk Tier Classification — November 2025 · 284 Active Files
All 284 files classified at origination · Tier assignment dynamic · Updated daily as signals change
Tier 1 (full review)84 files (29.6%)
Tier 2 (auto + conditional)138 files (48.6%)
Tier 3 (spot-check)62 files (21.8%)
Human reviews triggered94 of 284 (33.1%)
Every Tier 1 file receives the full 60-point automated QC checklist and a mandatory human review by the QC officer regardless of the automated result. Tier 1 files are those where one or more high-risk signals are active. The error rate in this tier is 41.7% (35 of 84 files flagged with at least one error) — substantially higher than the institution's 19.8% overall error rate. The resource concentration is justified: these 84 files represent 61% of all errors found despite being only 29.6% of the volume.
RM error rate >20% trailing 90d
DSA exception rate >25%
Loan amount >2× product average
Any credit policy exception in file
Complex product (LAP / co-lending / NRI)
Borrower DPD in prior 24 months
Every Tier 2 file receives the full 60-point automated QC checklist. Human review is triggered only if the automated check finds a failing result on one or more checks. Of the 138 Tier 2 files, 47 (34.1%) triggered at least one automated flag and received human review. The remaining 91 files (65.9%) passed automated review without any flag and were cleared without human involvement. Error rate in human-reviewed Tier 2 files: 22.2% (10 of 45 reviewed). Tier 2 represents the largest share of volume but a mid-range risk profile — the automated review is the primary gate, with human review as the escalation for flagged cases.
RM error rate 10–20%
DSA exception rate 10–25%
Standard products
Loan amount within normal range
No exception in file
Borrower prime profile
Tier 3 files receive a 20-point critical-check subset of the full QC checklist — focusing on compliance-critical items (KFS date, NACH account match, AML check, sanctions screening) rather than the full operational quality suite. Human review is triggered only if a critical check fails. Of the 62 Tier 3 files this month, 7 (11.3%) triggered a critical flag and received human review. The remaining 55 (88.7%) were cleared by the automated critical-check subset. These are the institution's lowest-risk files: long-tenure RMs with clean error histories, standard products, prime borrowers, no exceptions, amounts within normal range.
RM error rate <10% trailing 90d
DSA exception rate <10%
Standard salaried / MSME product
Loan amount below product median
No exception in file
Prime borrower, zero prior DPD
The full file review: a Tier 1 file walkthrough — LA-2025-9481 · LAP ₹38L
Tier 1 Full Review — LA-2025-9481 · LAP ₹38L · Mysuru · Nov 14, 2025
Risk signals: new RM (8% error rate) · DSA VS-018 (28% exception rate) · LAP complex product · Amount 1.8× average
ProductLAP · ₹38L · 12 years
RMKiran M. · Joined Jun 2025 · 8% error rate
DSA sourceVS-018 · Exception rate 28%
Review tierTier 1 · Full review + mandatory human
60-point review — critical items shown
✅Aadhaar identity — OTP verifiedUIDAI OTP verified Nov 10 · Name match confirmed · DOB confirmed
✅PAN validation — NSDL confirmedPAN AANPM8291K · Active · Name matches Aadhaar
❌Encumbrance certificate — EXPIREDEC dated April 18, 2025 · Review date Nov 14 · EC is 209 days old · Maximum permitted: 180 days · Error code: ERR-PROP-EC-EXPIRED · Routes to: RM for fresh EC
✅CERSAI search — completed within 30 daysCERSAI search Nov 3 · No prior charge · Clear
⚠Property valuation — valuer not on current approved panelValuation report by Ravi Associates · Ravi Associates removed from approved panel Oct 1, 2025 · Report dated Oct 22 — 3 weeks after removal · Error code: ERR-PROP-VAL-PANEL · Routes to: Legal team to obtain fresh valuation from approved valuer
✅KFS issued before sanction — confirmedKFS generated Oct 28 · Borrower eSign Oct 29 · Sanction letter Nov 1 · Sequence correct
❌NACH mandate — bank details mismatchNACH mandate: HDFC Bank XXXX-4421 · Bank statement: HDFC Bank XXXX-8814 · Different account numbers · Error code: ERR-LOAN-NACH-MISMATCH · Routes to: Ops team to re-execute NACH with correct account
✅AML / sanctions screening — clearBorrower not on OFAC, UN, CIBIL negative list · Not a PEP
✅Processing fee — matches scheduleFee charged: ₹28,500 · Schedule: 0.75% of ₹38L = ₹28,500 · Correct
Review outcome
HOLD — 3 errors · 2 critical, 1 advisory
ERR-PROP-EC-EXPIRED (critical) · ERR-LOAN-NACH-MISMATCH (critical) · ERR-PROP-VAL-PANEL (advisory — fresh valuation required)
Disbursement blocked until
Fresh EC obtained + NACH corrected + new valuation
Estimated: 5–7 working days
RM and DSA notified immediately
100%Automated coverage — every file receives at minimum the critical-check subset · Tier 1: full 60-point review + human · Tier 3: 20-point critical checks
41.7%Tier 1 error rate — vs 19.8% overall · Risk stratification concentrates human review where error probability is highest · Justified resource allocation
33.1%Human reviews triggered — 94 of 284 files · Down from 100% that a non-stratified 100% review would require · QC officer capacity preserved for high-risk files
6Risk signals per file — RM error rate, DSA exception rate, product complexity, loan amount, exceptions in file, borrower credit volatility · All dynamic
The Tier 1 file with 3 errors that was caught before disbursement was not caught because of good luck. It was caught because a new RM, a high-exception DSA, a complex product, and an above-average loan amount are the four signals that predict "this file needs a full review" — and the QC AI computed all four at origination.
LA-2025-9481 had an expired EC, a non-panel valuer, and a NACH account mismatch — three errors that would have reached disbursement if the file had been assigned to Tier 3 and reviewed only on the critical-check subset. But the Tier 3 assignment is reserved for experienced RMs with clean histories, on standard products, with prime borrowers. Kiran M. has been with the institution since June 2025. The assigned DSA (VS-018) has a 28% exception rate. The product is LAP — a complex product with property documentation requirements that generate errors at above-average rates. The loan amount is 1.8× the LAP product average. Four of six risk signals pointed to Tier 1. The full review found three errors and held the disbursement. Risk stratification is not a shortcut — it is the mechanism that makes 100% coverage resource-efficient by concentrating human review where it creates the most value.
