- Why AI in Loan Underwriting Is a Competitive Necessity
- Why Manual Underwriting Costs More Than You Realise
- What AI Underwriting Actually Does
- India's INDIA-5 Credit Intelligence Stack
- ML, GenAI, and Agentic AI: Three Tools, Three Jobs
- What RBI Actually Requires Before Go-Live
- Where AI Underwriting Should NOT Go Yet
- AI Across the Full NBFC Lifecycle
- The 4-Phase Implementation Roadmap
- Build, Buy, or Partner?
- How APPWRK Builds AI Underwriting for NBFCs
- Frequently Asked Questions
- Why AI in Loan Underwriting Is a Competitive Necessity
- Why Manual Underwriting Costs More Than You Realise
- What AI Underwriting Actually Does
- India's INDIA-5 Credit Intelligence Stack
- ML, GenAI, and Agentic AI: Three Tools, Three Jobs
- What RBI Actually Requires Before Go-Live
- Where AI Underwriting Should NOT Go Yet
- AI Across the Full NBFC Lifecycle
- The 4-Phase Implementation Roadmap
- Build, Buy, or Partner?
- How APPWRK Builds AI Underwriting for NBFCs
- Frequently Asked Questions
Key Takeaways
- 50% TAT reduction is verified, not aspirational. L&T Finance Project Helios delivered 30% TAT reduction across 5,000+ SME cases. Bajaj Finance disbursed Rs.1,600 crore through AI-powered channels in a single quarter (Q3 FY26). This is earnings-call data.
- India's NBFCs are growing 17% annually vs. 12% for banks and AI underwriting is the primary mechanism compounding that lead every quarter.
- Speed is the byproduct, not the point. The real value is credit expansion: reaching the 160-190 million thin-file borrowers that manual underwriting systematically excludes and the $330 billion MSME credit gap.
- India's INDIA-5 data stack is a global competitive advantage. Account Aggregator, UPI, GST, Bureau, and NACH give Indian NBFCs consent-based, real-time credit intelligence unavailable in most markets.
- RBI compliance is non-negotiable before go-live. The August 2024 Draft Circular requires independent model validation, SHAP explainability, DPDP-compliant granular consent, and a 6-week parallel run. Governance documentation costs Rs.15-40 lakh separately.
- The 4-phase roadmap separates success from failure. Most NBFC AI implementations fail because Phase 3 (the parallel run) is skipped.
This guide covers how AI-powered loan underwriting works for Indian NBFCs in practice: the real outcomes being delivered, the INDIA-5 data stack powering it, what RBI compliance actually requires, and a phased roadmap that preserves credit quality throughout.
An MSME owner in Nagpur applies for a Rs.12 lakh working capital loan. In 2022, this meant 18 days of paperwork, field visits, manual CIBIL checks, and a credit committee meeting. In 2026, the same application triggers an AI that pulls bank statements via Account Aggregator in 4 seconds, scores 5,000 data signals in under a minute, generates a pre-filled credit memo, and routes a recommendation to a credit officer. Decision: 6 hours. Disbursal: same day.
This is not a future prediction. L&T Finance confirmed a 30% TAT reduction across SME underwriting via Project Helios. Bajaj Finance disbursed Rs.1,600 crore through AI-powered channels in Q3 FY26 alone. The Economic Times has reported that AI-driven loan decisions are now 50% faster across India's lending sector. The 50% headline is Q3 FY26 earnings-call data from named, listed NBFCs.
This guide explains, in plain decision-maker language, how AI-powered loan underwriting for NBFCs in India works in practice: the outcomes being delivered, the data architecture behind it, what RBI compliance actually requires, where AI should not go, and how to implement it in a phased way without disrupting credit quality. Written for CTOs, CROs, and business heads at Indian NBFCs evaluating or accelerating AI underwriting deployment.
Why AI in Loan Underwriting Is Now a Competitive Necessity for Indian NBFCs
AI-powered loan underwriting is not a technology experiment for Indian NBFCs in 2026. It is the primary mechanism through which India's fastest-growing lenders are winning market share, expanding credit access to the $330 billion MSME credit gap, and building operational moats that compound every quarter.
India's NBFC sector is growing at 17% annually versus 12% for banks (Nomura, 2026). The gap is not accidental. NBFCs deploying AI across underwriting and servicing are approving creditworthy borrowers that manual processes exclude, processing volumes their headcount could never sustain, and doing it at a cost-per-file that keeps becoming more competitive.
India's NBFC AI Adoption Surge
The competitive urgency comes from two directions simultaneously. From above: digital-first NBFCs and fintech lenders with native AI are approving thin-file borrowers in hours. From below: 160-190 million new-to-credit Indian adults who have UPI history, GST filings, and AA-accessible bank statements are waiting for a lender fast enough to reach them.
Manual underwriting cannot serve both pressures. It is too slow to compete with digital-first NBFCs at the top, and too expensive to economically serve small-ticket MSME borrowers at the bottom. AI underwriting solves both simultaneously, compressing TAT from weeks to hours while making small-ticket lending economically viable at scale.
The real strategic argument for AI underwriting: Speed is the byproduct, not the point. NBFCs deploying AI purely for TAT reduction capture roughly 20% of the available value. Those using it to expand their addressable market, reaching the 160-190 million NTC borrowers with UPI and GST history, capture the full 100%. The business case is credit expansion, not just cost reduction.
Named NBFC Results: The Proof
The outcomes below are from named Indian NBFCs, sourced from earnings calls, analyst reports, and official press releases.
Tata Capital deployed AI credit memos and KAI Voice collection calls in 11 Indian languages. Poonawalla Fincorp launched 5 new AI solutions in 2026, including an underwriting partnership with IIT Bombay and an Early Warning System predicting portfolio risk at micro-market level. The pattern across India's large NBFCs is consistent: AI started in one department and became enterprise-wide within two years.
Why Manual Underwriting Is Costing Indian NBFCs More Than They Realise
Manual loan underwriting at most Indian NBFCs involves 7 to 21 days of processing time per file, costing between Rs.800 and Rs.2,500 per application in staff and operational overhead. The visible cost is time. The invisible cost is addressable market.
The 5 Hidden Costs of Manual Underwriting
-
1
Time Cost
Each file requires 4 to 7 touchpoints: document collection, CIBIL pull, bank statement analysis, income verification, field visit, and credit committee prep. Average TAT: 7 to 21 days. For MSMEs with working capital urgency, this hands the borrower to a faster competitor.
-
2
Staff Cost
A loan officer can manually process 40 to 60 files per month. An AI system handles thousands per day. The headcount-per-crore-disbursed ratio is the metric that separates scalable NBFCs from those that hit an AUM ceiling.
-
3
Opportunity Cost
While a file sits in a manual queue, a faster NBFC competitor approves the same borrower. Digital-first NBFCs are winning the thin-file segment by default because their TAT is measured in hours, not weeks.
-
4
Inconsistency Cost
Two credit officers reviewing the same file make different decisions 30 to 40% of the time due to subjective judgment variance. AI applies identical logic to identical inputs, every time. Consistency is also a regulatory asset when RBI examines your underwriting process.
-
5
Exclusion Cost
India's MSME credit gap exceeds $330 billion (World Bank, 2023). Manual underwriting cannot economically assess small-ticket, thin-file borrowers. Every creditworthy borrower declined due to process economics is lost revenue and lost financial inclusion impact.
What AI Underwriting Actually Does (And What It Does Not Replace)
AI underwriting automates the data-intensive, document-heavy steps of credit assessment. It does not replace human credit judgment. It makes that judgment faster, more consistent, and better-informed by eliminating the 80% of processing time spent gathering data before a decision is made.
Common misconception: "AI underwriting means fully automated loan approval with no human in the loop." For most Indian NBFC loan products above Rs.2 lakh, the credit decision still requires human sign-off. What AI eliminates is the data-gathering bottleneck. The 50% TAT reduction headline is conservative for many standardised loan types.
Three Layers of AI in Underwriting
| Layer | What it does | Technology | Time saved |
|---|---|---|---|
| Layer 1: Document Intelligence | Reads, classifies, and extracts data from bank statements, ITR, GST returns, KYC documents | OCR + NLP + Computer Vision | 2-5 days to minutes |
| Layer 2: Risk Scoring | Runs 500-5,000+ variables through predictive models to generate a credit risk score | ML models (XGBoost, LightGBM) | Hours to seconds |
| Layer 3: Decision Support | Generates credit memo, flags policy exceptions, routes to human for final call | GenAI / LLM + Rules engine | 3-4 hours to 20 minutes |
Manual vs. AI: Step-by-Step Comparison
| Step | Manual | AI-Powered | Improvement |
|---|---|---|---|
| Bank statement analysis | 2-3 days (manual review) | 30 seconds (AA pull + OCR) | 99% faster |
| Bureau pull | 1-2 days | 5 seconds (API) | 99% faster |
| Credit scoring | 2-4 hours (human judgment) | Under 60 seconds (ML model) | 95% faster |
| Credit memo prep | 3-4 hours (officer writing) | 15-20 minutes (GenAI draft) | 75% faster |
| Committee decision | Human (unchanged) | Human with 70% less prep | 70% faster prep |
India's Data Advantage: The INDIA-5 Credit Intelligence Stack
India's digital infrastructure, Aadhaar, UPI, GST, Account Aggregator, and NACH, gives Indian lenders a unique, consent-based data ecosystem that makes AI underwriting more powerful here than almost anywhere else in the world. No other emerging market has this combination at this scale, and it is India's largest structural advantage in the global AI lending race.
The INDIA-5 Stack: Five-Layer Breakdown
Layer 1: Bureau Intelligence (Foundation). CIBIL, Equifax, and CRIF Highmark provide past credit history, existing loans, and repayment track record. Critical limitation: approximately 40% of Indians have no bureau history, making this layer insufficient as a standalone data source for any NBFC targeting financial inclusion.
Layer 2: Behavioral Intelligence (UPI). With borrower consent, UPI transaction data reveals payment frequency, merchant categories, spending patterns, and income consistency. A borrower making 80+ UPI transactions per month to diverse merchants demonstrates financial activity that no bureau score captures. Available for approximately 600 million UPI users in India.
Layer 3: Business Intelligence (GST). GST returns (GSTR-1, GSTR-3B) and e-invoice data reveal business revenue, turnover trends, filing regularity, and tax discipline. For MSMEs, GST filing consistency is a stronger proxy for business health than 3-year-old financial statements. Available for 1.4+ crore GST-registered businesses.
Layer 4: Cash Flow Intelligence (Account Aggregator). The AA framework enables consent-based bank statement pulls in 4 to 8 seconds instead of 5 to 7 days. By end of FY25, Rs.1.67 lakh crore in loans were disbursed via Account Aggregator across 189 lakh accounts, a 208% year-on-year increase in disbursement value. Source: Sahamati AA Impact Report H2 FY25.
Layer 5: Compliance Signals (NACH + EPFO + Utilities). NACH mandate history reveals EMI payment discipline. EPFO records confirm employment stability. NACH mandate bounces are one of the earliest predictive signals for impending default, typically appearing 60-90 days before a formal delinquency event.
Account Aggregator: The Coverage Gap Nobody Mentions
Reality check: AA is powerful but not universal. Rural cooperative banks, many Regional Rural Banks, and smaller UCBs are not yet fully integrated as Financial Information Providers. For NBFCs targeting Tier 3+ rural borrowers, AA data coverage can drop to 60-70% of applications. Build fallback data paths on Day 1: if AA pull fails, auto-trigger GST pull via DigiLocker plus alternate bank statement upload. Never build a single-point AA dependency into your architecture.
Architecture Decision: Feature Store for high-volume NBFCs. For NBFCs processing 500+ applications per day, a centralised feature store with pre-computed UPI signals, GST flags, and AA-derived income ratios reduces AI scoring latency from 8-12 seconds to under 2 seconds. Below 500 applications per day, on-demand computation is sufficient.
ML, GenAI, and Agentic AI: Three Distinct Tools, Three Distinct Jobs
Most NBFC CTOs think of AI underwriting as a single system. In practice, it is three layers working in sequence. ML handles the numbers. GenAI handles the language. Agentic AI handles the workflow. Confusing these three leads to both technology mismatches and vendor overselling.
How Each Tool Fits Into the Underwriting Stack
| Tool | What it does in underwriting | India NBFC Example |
|---|---|---|
| ML / Machine Learning | Scores creditworthiness from structured data (bureau + UPI + GST + AA). XGBoost, LightGBM. Handles 500-5,000 variables simultaneously. | Lendingkart: 5,000-variable MSME scoring in 3.5 seconds |
| GenAI / LLMs | Reads unstructured documents (ITR PDFs, handwritten income statements, vernacular text), drafts credit memos, generates plain-language rejection explanations for RBI compliance | Tata Capital: AI credit memos reducing committee prep by 70% |
| Agentic AI | Multiple AI agents collaborate: one validates KYC, one extracts GST data, one runs the credit model, one compiles the memo, all orchestrated end-to-end with human escalation built in | L&T Finance Project Helios: 30% TAT reduction, 1.5 hours saved per SME case |
Architecture Decision: Use GenAI for unstructured documents only. For structured documents (bank statements, ITR, GST returns), fine-tuned OCR plus NLP is faster, cheaper, and more accurate. GenAI adds value for genuinely unstructured content: handwritten income statements in vernacular languages, franchise agreements, varied-format property documents. A hybrid approach prevents hallucination risk in high-stakes structured data extraction.
What RBI Actually Requires Before Your AI Model Goes Live
The RBI's August 2024 Draft Circular on Model Risk Management classifies AI lending models as high-risk financial instruments, requiring documented governance before any model enters production, regardless of vendor or platform. This is examination-ready compliance, not aspirational guidance.
The 5 Non-Negotiable Requirements
-
1
Model Risk Documentation
Every AI credit model must have: model development documentation, a validation report by an independent team, and a governance approval trail. Any material model change resets this requirement. Plain language: Before go-live, you need proof that someone other than the model developer has tested and approved it.
-
2
Explainability: No Black Boxes
Credit decisions must be explainable to the borrower in plain language. SHAP (Shapley values) for gradient boosting models is the industry standard. For customer-facing rejection notices, translate SHAP output: "Your application was declined primarily due to high existing EMI obligation relative to income." This dual-layer approach satisfies both model performance and regulatory requirements. Plain language: Your model must explain its "no" the way a loan officer would.
-
3
Human-in-the-Loop: Mandatory
Fully automated decisions for complex or high-value loans are not permitted without human review pathways and escalation protocols. AI recommends; humans decide, especially above Rs.10 lakh ticket sizes. Plain language: AI speeds up the process; a human still signs off.
-
4
DPDP Act 2023: Granular Consent Architecture
The Digital Personal Data Protection Act 2023 requires separate, explicit borrower consent for each data source. AA data, UPI data, and GST data each require their own consent flow. Bundled consent is non-compliant. Each consent must be granular and produce an auditable record. Plain language: You cannot use a borrower's UPI data unless they specifically consented to that use, separately from other data sharing.
-
5
Model Retraining Governance
Post-deployment, every material model change requires a fresh validation cycle before the updated model goes live. This applies to changes triggered by performance drift, regulatory change, or new data sources. Plain language: Deploying the model is not the end. You need a quarterly monitoring process and a protocol for when accuracy starts dropping.
Hidden cost: RBI compliance documentation is expensive and excluded from vendor quotes. Preparing these five compliance artifacts for an upper-layer NBFC typically costs Rs.15 to 40 lakh in consulting and documentation effort before deployment. Budget for this separately before signing any AI underwriting contract.
Algorithmic bias is now an examination requirement. RBI examination requires documented bias testing across gender, geography, and employment type before deployment. Models trained on historical bureau data systematically disadvantage self-employed, rural, and women borrowers. An AI model that declines self-employed women in rural areas at higher rates than comparable urban male borrowers is a compliance liability, even if its default prediction accuracy is high.
Counter-narrative: AI models do not self-improve. A deployed credit model degrades silently. Economic cycles, new fraud tactics, and RBI rate changes cause model accuracy to drop over time. A model at 92% accuracy at launch may operate at 78% accuracy 18 months later without triggering any alert, unless you have continuous monitoring. The RBI retraining governance requirement exists precisely because this failure mode is real and documented.
Where AI Underwriting Should NOT Go Yet: Guardrails That Protect Your Portfolio
Not every borrower segment is ready for AI underwriting in 2026. Deploying AI where it lacks usable signal is the fastest way to increase NPA ratios, not reduce them. A vendor who claims AI can underwrite every Indian borrower equally well is overselling the technology.
Three Borrower Segments Where AI Falls Short Today
Caution Zone 1: Borrowers with Zero Digital Footprint. Approximately 160-190 million thin-file adults have no UPI history, no GST registration, and no formal bank account activity. AI has no usable signal. What to do instead: human assessment with field-based income verification. As India Stack expands, many of these borrowers will become AI-scorable within 2-3 years, but forcing AI-only decisions today will increase NPAs.
Caution Zone 2: Ultra-Small-Ticket Loans (Below Rs.5,000). The compliance cost of running an AI model, including documentation, audit trail, and RBI governance, can approach the economics of the loan itself at this ticket size. Group lending models or NBFC-MFI frameworks are better suited for this segment.
Caution Zone 3: Seasonal or Agricultural Income Patterns. AI models trained on regular monthly inflow patterns consistently misread lump-sum post-harvest income as irregular or high-risk, generating false negatives for creditworthy agricultural borrowers. Use a hybrid: AI for document processing and KYC, human credit officers for agricultural income assessment.
India requires separate model pipelines per borrower segment. A single unified model trained on urban salaried bureau-positive borrowers performs poorly on rural self-employed MSMEs. Build separate training pipelines for at minimum: (1) salaried consumer, (2) self-employed consumer, (3) MSME working capital, (4) rural or agricultural. A single unified model is a vendor convenience, not a credit engineering decision.
AI Across the Full NBFC Lifecycle: Underwriting Is Just the Beginning
Loan underwriting is where AI delivers the most visible TAT improvement. But across the NBFC lending lifecycle, AI transforms every stage from the first borrower inquiry to the last recovery call. NBFCs that deploy AI only in underwriting capture perhaps 30% of the available efficiency. Full lifecycle deployment builds a compounding advantage every quarter.
The 6-Stage AI Lending Lifecycle Breakdown
Stages 1-2 (Lead Qualification + KYC): AI chatbots qualify leads and run soft bureau checks without agent involvement. Computer vision and OCR reduce onboarding TAT from 48 hours to 6 hours. Sixty to eighty percent of top-of-funnel queries are automated.
Stage 4 (Servicing): AI chatbots handle EMI queries, statement requests, and foreclosure inquiries 24/7 on WhatsApp and voice. RBI's 30-day grievance window compliance becomes manageable at scale.
Stage 5 (Collections): AI voice bots call pre-due accounts in the borrower's preferred language, capture promises-to-pay, and send UPI payment links. Predictive models identify default risk 60-90 days before formal delinquency. Tata Capital's KAI Voice handles calls in 11 Indian languages.
Stage 6 (Portfolio Monitoring): L&T Finance's Project Nostradamus delivers real-time micro-market cluster-level insights. Poonawalla Fincorp's Early Warning System flags at-risk segments before defaults materialise.
The 4-Phase AI Underwriting Implementation Roadmap for Indian NBFCs
Most NBFC AI underwriting projects fail not because the technology is wrong, but because the implementation sequence is wrong. This roadmap is built specifically for India's regulatory environment and data infrastructure.
Phase-by-Phase Rollout Plan
Phase 1: Readiness and Data Audit (Month 1-2)
- Map all existing data sources: LOS data, bureau history, bank statement formats, field assessment data
- Check Account Aggregator integration status (if not connected, this is Day 1 priority)
- Identify your pilot product: personal loans or MSME working capital (high volume, structured data)
- Assess internal model risk governance capability versus external consulting need
- Budget Rs.15-40 lakh for RBI compliance documentation before any vendor quote
Output: Data readiness report + pilot product + vendor shortlist
Hidden cost alert: Data cleaning and labelling typically costs 30-50% of total AI build cost and is almost always excluded from vendor quotes.
Phase 2: Model Development and Compliance Prep (Month 3-4)
- Feature engineering from AA + GST + UPI + bureau data (separate pipelines per borrower segment)
- Model development: XGBoost baseline, SHAP explainability layer, DPDP-compliant consent architecture
- Draft Model Risk Management documentation per RBI August 2024 Draft Circular
- Bias testing across gender, geography, and employment type
- Design granular consent flows for each data source separately
Output: Model v1 + compliance documentation package + consent flow design
Phase 3: Parallel Run and Validation (Month 5-6)
CRITICAL: Most commonly skipped. Most common cause of NBFC AI underwriting failure. Do not compress or skip this phase.
- Run AI model alongside manual underwriting on every file. Compare outputs in real time
- Track: approval rate difference, false positive rate, processing time delta
- Human review of every AI recommendation during this phase
- Tune model on every case where AI and human disagree
- Calibrate escalation threshold: at what confidence score does AI recommend versus route to human?
- Minimum 6-week parallel run on highest-volume product before going live
Output: Model validation report + live or no-go decision + escalation protocol
Phase 4: Scale and Optimise (Month 7-12+)
- Roll out to full product volume
- Add GenAI credit memo layer (reduces committee prep time by approximately 70%)
- Deploy portfolio monitoring dashboard with model drift detection
- Quarterly retraining cycle triggered by performance benchmarks
- Expand AI to onboarding, servicing, and collections stages
Output: Full lifecycle AI deployment + quarterly model review process
Ongoing cost alert: Model drift monitoring is a recurring infrastructure cost. A model at 92% accuracy at launch may be at 78% eighteen months later without active monitoring.
Build, Buy, or Partner? The Decision Framework for NBFC CTOs
The right answer depends on AUM, internal data science capability, and timeline to compliance-ready deployment. Most NBFCs default to "buy a platform" without understanding that platform selection does not transfer the compliance documentation burden to the vendor.
Cost, Timeline, and Compliance: Side-by-Side
| Option | Timeline | Cost | RBI Compliance Readiness | Best For |
|---|---|---|---|---|
| Build In-House | 12-24 months | Rs.1-5 crore+ | Full internal governance build required (MLOps, SHAP, bias testing) | Large NBFCs (AUM above Rs.10,000 crore) with existing data science teams |
| Buy a Platform | 3-6 months | Rs.25-80 lakh/yr (SaaS) | Platform handles tech; governance documentation remains NBFC's responsibility | Mid-size NBFCs wanting speed; must validate India-specific model performance |
| Partner with Specialist | 6-12 months | Rs.40 lakh to Rs.2 crore (project) | Partner and NBFC share documentation responsibility; fastest path to RBI-compliant deployment | Most NBFCs: fastest path to India-specific, compliance-ready deployment |
For NBFCs with AUM below Rs.5,000 crore, partner plus configure is the fastest, most compliance-ready path. Building from scratch requires MLOps, bias testing, and SHAP infrastructure that most NBFCs do not yet have internally. When buying a SaaS platform, ask for reference clients at comparable NBFC segments, and confirm Model Risk Management documentation is included in the engagement scope, not billed separately.
Two questions every NBFC CTO should ask any AI underwriting vendor: (1) Does your model include separate training pipelines for salaried consumer, self-employed consumer, MSME working capital, and agricultural segments? A vendor who says yes to a single unified model is a red flag. (2) Do you provide a Model Risk Management documentation package that satisfies the RBI August 2024 Draft Circular? If they hesitate, your NBFC will discover the compliance gap during an RBI examination.
How APPWRK Builds AI Underwriting Platforms for Indian NBFCs
At APPWRK IT Solutions, we have built AI-powered credit assessment and lending technology platforms for financial institutions across India and Southeast Asia. Our engagements span ML model development from Account Aggregator and LOS data, RBI Model Risk Management documentation, SHAP-based explainability layers, DPDP-compliant consent architecture, and phased deployment support through the mandatory parallel run validation phase.
Our Three-Principle Engineering Approach
Our engineering approach follows three principles that most vendor engagements miss: India-first model segmentation with separate training pipelines per borrower segment, compliance documentation as part of delivery rather than a post-launch concern, and the 6-week parallel run embedded in the project plan by design. We have delivered this for clients in digital lending, consumer finance, and MSME lending verticals.
Whether you are assessing your data infrastructure readiness, evaluating build vs. buy vs. partner options, or accelerating a stalled AI underwriting rollout, APPWRK's engineering team will help you build it correctly from the outset. Talk to our AI team today.
Explore APPWRK's AI Development Services to see how we structure phased, compliance-ready AI engagements for regulated financial institutions in India.
Frequently Asked Questions
Q: How does AI reduce loan approval time for NBFCs in India?
AI automates the data-heavy pre-decision steps: Account Aggregator pulls bank statements in 4-8 seconds instead of 5-7 days, ML models score credit risk in under 60 seconds, and GenAI drafts credit memos in 15-20 minutes instead of 3-4 hours. Indian NBFCs have documented 30-80% TAT reductions, with L&T Finance achieving 30% on 5,000+ SME cases (Business Standard, February 2026).
Q: What is the Account Aggregator framework and how does it improve NBFC credit scoring?
The Account Aggregator (AA) framework is an RBI-regulated, consent-based system that lets borrowers share verified bank statement data with NBFCs in under 10 seconds. Before AA, NBFCs waited 5-7 days for physical statements. By end of FY25, Rs.1.67 lakh crore in loans were disbursed via AA across 189 lakh accounts, a 208% year-on-year increase in disbursement value (Sahamati AA Impact Report H2 FY25).
Q: Which Indian NBFCs are using AI for loan underwriting in 2026?
Documented deployments include L&T Finance (Project Helios: 30% SME TAT reduction, 5,000+ cases), Bajaj Finance (Rs.1,600 crore disbursed via AI call centres in Q3 FY26), Tata Capital (AI credit memos, KAI Voice in 11 languages), and Poonawalla Fincorp (IIT Bombay underwriting partnership plus Early Warning System). EY India reports 74% of Indian FIs have active AI proof-of-concepts.
Q: What does RBI require for AI loan underwriting compliance in India?
The RBI August 2024 Draft Circular on Model Risk Management requires: (1) independent model validation documentation, (2) SHAP-based explainability for every credit decision, (3) mandatory human-in-the-loop for high-value loans, (4) DPDP Act 2023 granular consent per data source (AA, UPI, GST separately), and (5) post-deployment retraining governance. Compliance documentation typically costs Rs.15-40 lakh before deployment.
Q: What is the ROI of AI underwriting for an Indian NBFC?
Documented outcomes include 30-80% TAT reduction, 25-40% approval rate uplift at the same or lower default rate, 20-50% delinquency reduction, and 30-50% operational cost reduction. ROI is highest when AI expands credit access to the $330 billion MSME credit gap by approving thin-file borrowers previously excluded by manual underwriting economics.
Q: What alternative data sources does AI use for credit scoring in India?
Beyond CIBIL: UPI transaction data (payment frequency, income inflows), GST returns (GSTR-1, GSTR-3B for business revenue), Account Aggregator bank statements (cash flow, EMI servicing history), NACH mandate history (EMI payment discipline), and EPFO records (employment stability). Combining these adds 25%+ prediction accuracy for thin-file borrowers (World Bank, 2024).
Q: What is agentic AI in NBFC lending?
Agentic AI refers to multiple AI models collaborating in an orchestrated workflow where one validates KYC, one extracts GST data, one runs the credit model, and one compiles the memo, without manual handoffs. L&T Finance's Project Helios is India's most cited example: 30% TAT reduction, 1.5 hours saved per case, across 5,000+ SME files.
Q: How is GenAI underwriting different from ML underwriting?
ML models score creditworthiness from structured data (bureau, UPI, GST, AA) using gradient boosting algorithms. GenAI handles language-heavy tasks: reading unstructured documents, drafting credit memos, and generating plain-language rejection explanations for RBI compliance. Most NBFC deployments use both in sequence, with ML for scoring and GenAI for document processing and memo generation.
Q: How do I implement AI underwriting without disrupting credit quality?
Follow the 4-phase approach: (1) data readiness and AA integration audit, (2) model development with RBI compliance documentation, (3) mandatory 6-week parallel run alongside manual underwriting on all files, then (4) phased scale-up with quarterly drift monitoring. Phase 3, the parallel run, is most commonly skipped and most commonly causes implementation failure.
Q: Where should NBFCs not use AI underwriting in 2026?
Three segments remain high-risk: borrowers with zero digital footprint (no UPI, GST, or formal banking), ultra-small-ticket loans below Rs.5,000 where compliance cost approaches loan value, and seasonal or agricultural income borrowers whose lump-sum post-harvest inflows are misread as irregular income by models trained on salaried patterns.
Q: What does AI underwriting cost for an Indian NBFC?
Build in-house: Rs.1-5 crore+ over 12-24 months. SaaS platform: Rs.25-80 lakh per year. Specialist partner: Rs.40 lakh to Rs.2 crore as a project. All options require Rs.15-40 lakh for RBI compliance documentation (rarely included in vendor quotes). Data cleaning and labelling adds 30-50% to total build cost, also typically excluded from vendor scopes.
Q: Can AI underwriting be fully automated for Indian NBFCs?
Not for most loan products above Rs.2 lakh. RBI guidelines require human-in-the-loop for complex or high-value credit decisions. What AI fully automates is the 80% of processing time spent gathering, verifying, and organising data before the human decision. The credit committee still decides; AI makes that decision faster, more consistent, and better-informed.
About The Author
