The Credit Decision Agent AI auto-approves 68% of applications and auto-declines 14%. The remaining 18% are referred to a human underwriter — not because the system failed to process them, but because they contain a specific, identified ambiguity that requires human judgment to resolve. The referral is not a failure state. It is the AI's most precise output: a complete case brief, a specific question, and a defined set of options for the underwriter to choose from.
The referral is not a middle ground — it is a specific decision
A poorly designed credit decisioning system uses the referral as a catch-all for uncertainty. When the score is borderline, refer. When there are flags, refer. When the underwriter might disagree, refer. This approach produces high refer rates that defeat the purpose of automation and create underwriter workload that is heavier and less structured than what existed before.
The Credit Decision Agent AI's referral is a specific, triggered outcome. Each of the six referral conditions is defined in the scorecard logic with a precise threshold. When a referral is triggered, the system has already identified exactly what requires human judgment — it provides the underwriter with a brief that takes them directly to the specific question, skipping everything the AI has already resolved. The underwriter is not reviewing the application from scratch. They are answering one or two specific questions, supported by the evidence the AI has assembled, and selecting from a defined set of options.
The six referral trigger conditions
FOIR
45–55%
FOIR between 45% and 55% — above policy ceiling but below decline threshold
The FOIR at the proposed loan amount is between 45% and 55% — above the standard policy ceiling but below the hard knock-out. Human underwriter reviews: what is the nature of the existing EMI obligations (are any nearing end of tenure?), is the income trend strongly upward (making the FOIR a temporary constraint), and is there a co-applicant or additional income source that would bring the ratio within policy?
→ Underwriter decision: approve with FOIR covenant, counter-offer at lower amount, or declineScore
55–64
Weighted scorecard in the refer zone — no KO failures, but multiple soft negatives
The application cleared all knock-out gates and refer triggers but accumulated a combination of soft negatives that brought the total below the 65-point auto-approve threshold. No single factor failed definitively — the concern is the pattern. The underwriter reviews: is the combination of soft negatives coincidental or structural? Is there context from the application or the borrower's communication that changes the assessment?
→ Underwriter decision: approve at reduced amount, approve with conditions, or declineDPD
60+ in
12 months
One DPD event of 60–89 days in the last 12 months — borderline credit history
The credit bureau shows one DPD event between 60 and 89 days in the last 12 months. This is below the 90-day KO threshold but above the 30-day soft-flag level. The underwriter reviews: is the DPD event on an account that has since been closed, on a small credit card versus a major secured loan, explained by a documented hardship event? The nature of the DPD event matters to the assessment in a way that a threshold alone cannot capture.
→ Underwriter decision: approve (DPD explained/context satisfactory), approve with conditions, or declineIncome
Gap 2+
months
Two or more consecutive months with no income credit in the bank statement
The bank statement shows a gap of two or more consecutive months with no eligible income credit — indicating a period of unemployment, leave without pay, or business disruption. The AI scores this negatively but refers rather than auto-declines because the context matters: a sabbatical is different from redundancy; a business closure is different from a seasonal inventory gap. The underwriter reviews the gap period and any explanation the borrower has provided.
→ Underwriter decision: approve (gap explained, current income confirmed), approve at reduced amount, or declineDocument
Amber
Flag
Document verification returned a soft flag — not a hard forgery signal, but an anomaly requiring review
The Document Verification AI returned an amber flag on one or more submitted documents — authenticity score between 75 and 89, one or two soft signals present but below the hard flag threshold. The credit underwriter reviews the specific flag (e.g., minor bank corroboration discrepancy, employment letter age), the document context, and decides whether the flag changes the credit assessment or whether the explanation in the file resolves it.
→ Underwriter decision: proceed (flag satisfactorily explained), request additional document, or declinePolicy
Exception
Required
One parameter is outside credit policy but within the exception authority of a senior underwriter
A specific parameter falls outside the standard credit policy — LTV slightly above ceiling, employment sector with restricted category but strong institutional backing, loan amount above standard auto-approve limit. The exception cannot be made by the auto-decision engine but is within the exception authority of a senior underwriter or credit committee, depending on the materiality. The refer brief includes the specific exception required, the policy reference, and the evidence supporting a positive determination.
→ Underwriter decision: grant exception (with documented rationale), decline exception, or escalate to credit committeeThe referral brief: what the human underwriter receives
Every referral sends the underwriter not a flagged file but a structured brief: the application summary (5 key metrics), the specific trigger that caused the referral, the evidence for and against approval on that specific issue, the AI's assessment of which outcome is most supported by the evidence, and three decision options with the implications of each. The underwriter's work is not to review the application from the beginning — it is to read the brief, review the referenced section of the file, and make a determination on the specific question the AI has identified.
At the Credit Decision Agent AI's average referral brief quality, underwriters spend an average of 8 minutes per referred application rather than the 45–60 minutes a full manual review would require. The 18% referral rate, at 8 minutes per referral, requires approximately 0.3 full-time equivalent underwriter per 100 applications per day — compared to 1.5 FTE per 100 applications in a fully manual operation.
The 10 most referred application scenarios
FOIR 47–50% — borderline range, strong income trend
FOIR 47% at proposed amount, but borrower's income has grown 28% in 12 months. The improving trajectory suggests FOIR will reduce organically. Underwriter reviews: can a 2-year income covenant be set, so the FOIR at year 2 projected income is within policy?
→ Typical outcome: approve with income covenant (82% of these referrals)Score 58 — clean bureau, stable income, but low savings rate
No KO failures. CIBIL 718. 0 DPD. FOIR 38%. But cash flow score 62 and savings rate 9% — two soft negatives pulling score below auto-approve. Underwriter reviews: is the low savings rate structural (high rent city, dependents) or behavioural?
→ Typical outcome: approve with NACH mandate strengthening (67% of these referrals)DPD 65 days on credit card — 8 months ago, since closed
A 65-day DPD event on a credit card that was closed 6 months ago. The account is not active; the DPD has not recurred. The rest of the credit history is clean. Underwriter reviews: is a closed account DPD from 8 months ago material to a secured home loan applicant?
→ Typical outcome: approve (DPD on closed account, not predictive for secured loan) — 74% of these2-month income gap — sabbatical, now re-employed with higher salary
Income gap November–December 2024 (2 months). Salary credits resume in January 2025 at ₹1,10,000/month vs ₹82,000/month before the gap — borrower changed employer. The gap corresponds to a notice period. Underwriter reviews: the gap is explained and current income is higher.
→ Typical outcome: approve (gap explained, income improved) — 79% of these referralsEmployment letter 94 days old — 4 days outside 90-day window
Document Verification flagged the employment letter as 4 days outside the 90-day recency requirement. The letter is genuine; the employer is verified. The substance of the letter is current — the borrower is employed at the same role, same salary. Underwriter reviews: is a 4-day window breach material?
→ Typical outcome: proceed (4-day window is de minimis, request fresh letter as condition) — 88%LTV 77% — 2pp above 75% ceiling, strong borrower profile
LTV at requested amount is 77% against a 75% policy ceiling. Borrower: CIBIL 762, FOIR 32%, 0 DPD, income improving. Exception is 2pp — within the 5pp exception authority of a senior underwriter. Brief includes: exception required, evidence for approval, precedent cases.
→ Typical outcome: exception granted with standard conditions — 71% of these referralsScore 61 with DPD 30 12 months ago AND FOIR 44%
Two separate soft negatives: a 30-day DPD event 12 months ago (just within the 12-month window) and FOIR at 44% (1pp below the 45% ceiling). Score 61 — borderline. Neither is a KO alone; together they push the score into refer zone. This is a genuine judgment call on whether the combination represents acceptable risk.
→ Typical outcome: divided — 52% approve at lower amount, 48% declineFreelancer — high income, high variance, no Form 16
Self-employed borrower: 12-month average eligible income ₹1.8L/month after haircut. CV 0.44 (high). No Form 16 — full-time freelancer. FOIR 34%. The income is substantial but variable, and the income proof is bank statement only. Underwriter reviews: is 12 months of bank statement income sufficient without a tax filing?
→ Typical outcome: approve with additional bank statement + GST return condition — 64%7 bureau enquiries in 6 months — all for home loan (comparison shopping)
7 enquiries in 6 months triggers the refer threshold. However, all 7 enquiries are from housing finance companies or banks — the borrower has been comparison-shopping for a home loan, not accumulating unsecured credit. The enquiry pattern is the expected behaviour of a diligent home loan applicant.
→ Typical outcome: approve (enquiry pattern explained by product type — 86% of these)Income from 3 different sources — no dominant primary income
Eligible income ₹1.4L/month: salary ₹48K, consulting ₹52K, rental ₹18K (after haircut: ₹9K). No single income source is dominant. The income is diversified — arguably more resilient than single-source — but the credit policy income stability model was designed for salaried borrowers. Underwriter reviews the appropriateness of the haircut structure for this profile.
→ Typical outcome: approve with income source documentation condition — 72%The AI/Human boundary — what the Credit Decision AI never decides
The 18% referral rate is not a system limitation — it is the system working correctly
A credit decisioning system with a 0% referral rate has not automated credit decisions — it has replaced human judgment with a rigid rule set that cannot handle ambiguity. A system with a 60% referral rate has not improved on manual underwriting — it has created a costly triage layer with no efficiency gain. The 18% referral rate is the target: the precise proportion of applications that contain a specific, identified ambiguity that a well-briefed human underwriter can resolve in 8 minutes with the evidence the AI has assembled. The Credit Decision Agent AI's value is measured not by how few applications it refers, but by how precisely it identifies which applications require human judgment — and how completely it prepares the human to make that judgment in the shortest time.