The top of the origination funnel used to be the most expensive part. Telecallers, DSAs, branch outreach, partner referrals. Each lead carried an acquisition cost embedded in the eventual loan margin. AI voice and text bots have effectively rewritten that math.

Bajaj Finance has built the most documented operation here. Leads generated directly by Voice and Text AI Bots produced Rs 1,895 Cr in disbursals in Q4 FY26 alone. Cumulative disbursals attributed to AI voice bots in the personal loan business have crossed Rs 15 Bn.

This is not a chatbot answering FAQs. It is an agentic outreach system that initiates contact, handles objections, captures consent, qualifies the lead, and hands a hot prospect into the application flow without a human sales agent in the chain.

Two things make this work at scale. First, the voice bot is trained on the lender's own product policy, so it does not surface eligibility criteria the borrower will fail at the next stage. Second, it is integrated upstream into the AA consent flow and downstream into the underwriting engine, so a qualified lead does not get re-interviewed by a human three days later.

The cost implication is direct. Telecaller-driven outreach at an NBFC runs at roughly Rs 50 to Rs 150 per qualified lead. Voice bot-driven outreach, after amortising the model build cost, runs at a fraction of that. No shift differential. No attrition cost. No quality variance across the day.

Tata Capital has built the same capability but for a different problem. Through its partnership with Sarvam AI, the company runs 20 Lakh+ monthly customer interactions through Samvaad, the voice AI platform, across 11 Indian languages. Bajaj Finance's voice bots are conversion-engineered. They optimise for lead-to-disbursal in a single language stream, mostly English and Hindi. Tata Capital's are reach-engineered. They serve customer segments that English-only digital channels could never have engaged at scale. (For a lender expanding into semi-urban and rural geographies, multilingual voice AI is not a customer experience overlay. It is a market access strategy.) The Samvaad framework keeps a human in the loop for nuanced or judgement-heavy conversations.

Voice AI in origination has split into two architectures. Conversion at scale, and reach at scale. The lender running neither in 2026 is running last decade's stack with an AI label on top..

This is the layer where the gap between lenders is widest. Pulling AA data is table stakes; most lenders are doing it. Building a model that does something useful with that data is where the architecture choice shows up.

L&T Finance's Project Cyclops is the most measurable example in the market. The system runs 17 to 20 ML models in production, each trained on a specific underwriting question. Cash flow regularity. Income seasonality. GST filing consistency. Geographic risk weighting. The collective output is a three-dimensional credit score drawing on bureau data, AA liabilities data, and what the company calls trust signals. Payment gateway data, geolocation intelligence, banking transaction velocity.

The point of the architecture is not to approve more borrowers indiscriminately. It is to identify safe borrowers inside segments that traditional rules would have rejected, and route them back into the approval queue. “Swap In” Project Cyclops boosted new-to-credit underwriting by 34% versus non-Cyclops methods. Indexed Net Non Starters dropped significantly as well in both TW and Farm Equipment finance

The test of an underwriting engine is not the sanction it approves… it is the NPA it prevents 6 months later.
(Wait for the next edition for the deep dive into Credit)

For decades, document QC at an NBFC meant a desk full of officers cross-checking PAN photos against KYC forms, salary slips against bank statements, and addresses against utility bills. It is one of the most repetitive, error-prone, and expensive steps in origination. It is also the step where Vision AI has produced the most direct cost takeout.

Bajaj Finance's progression is the cleanest data set in the industry. Auto-QC of documents reached 44% in Q4 FY26, against a stated target of 90% by FY27. Cumulative document processing through AI-enabled flows has eliminated manual verification for over 35 Mn documents. (The 90% target is deliberate. The remaining 10% is reserved for the cases where regulators expect human eyes in the loop. Identity document anomalies, video KYC face mismatches, signature variances. The model flags. The human decides.)

Poonawalla Fincorp has taken a parallel route through its IIT Bombay partnership. The company uses GenAI models for short-term and prime personal loans, compressing time-to-sanction and improving approval quality. The Central KYC AI Platform alone reduced manual intervention in KYC processing by approximately 15%.

Models are trained over large data sets that read a salary slip and confirms it matches the borrower's stated employer, the bank account credited matches the AA-pulled primary account, and the dates align with the application month. These models are reconciling across data sources. That reconciliation rule library takes years to build and gets sharper with every reviewed exception.

The cost implication is direct. A large NBFC processing 50,000 applications per month, with 4 to 6 documents per application, is processing 200,000 to 300,000 documents monthly. At even 30% auto-QC, that is 60,000 to 90,000 documents that no longer require a human reviewer.

ICICI Bank's design choice takes this further. If the bank statement comes from AA, the GST filing comes from the GSTN API, and the KYC comes from CKYC and Aadhaar, the document upload step does not exist. The QC step does not exist either, because the data was already verified at source.

Manual document QC is the first origination cost line that disappears entirely under an AI-first stack. Lenders still scaling QC/backoffice in 2026 are scaling a function that will cease to exist soon.

The credit memo is the document that captures the underwriter's analysis and recommendation for sanction. Historically it has been written manually after a Personal Discussion with the borrower, drawing on financial statements, collateral details, bureau output, and the underwriter's own notes. The process is judgement-heavy, but a substantial part of it is documentation. That is the layer GenAI is taking over.

L&T Finance's Project Helios is the AI Underwriting Co-Pilot live in SME Finance. The system synthesises financial statements, collateral details, AA data, and PD notes into a draft credit memo at the click of a button. The underwriter still owns the recommendation. The co-pilot owns the first draft.

The more interesting feature is forward-looking. Helios suggests specific questions the underwriter should ask the applicant during the PD, generated from gaps and anomalies the model has identified in the application data. The underwriter walks into the conversation with a tailored interview prompt rather than a generic checklist.

Bajaj Finance has gone after the same problem from the documentation side. The company is using AI to summarise PD notes taken during origination, with a stated target of 30% improvement in underwriter productivity by FY27. (PD note summarisation is one of the most time-consuming non-judgement parts of credit review. Removing it does not remove the credit officer. It reduces the time the credit officer spends on documentation versus actual decisioning.)

The architectural point is the same in both cases. The co-pilot does not replace the underwriter. It compresses the part of the underwriter's day that is administrative, so more of the day goes into the part that requires judgement.

Same headcount. More sanctions. Sharper questions. Faster turnaround. The co-pilot is a re-allocation of the underwriter's hours, not a replacement of the underwriter.

Origination fraud has historically been caught after the fact. A loan disburses, defaults at month one, the post-mortem reveals a manipulated PDF or a stolen identity, the lender writes off the loss and tightens the next quarter's policy. AI is moving fraud detection from retrospective to inline.

Bajaj Finance has the most documented inline fraud stack in the market. Vision AI examines document authenticity, detecting tampered PDFs, mismatched signatures, and template anomalies. Voice AI analyses voice biometrics during the application call to flag impersonation attempts. Network anomaly detection identifies coordinated application patterns that suggest a fraud ring rather than a single bad applicant. All three layers run at the point of origination, before sanction.

The biometric layer has scaled visibly. Facial recognition cameras are deployed at 60 branches, with 6.3 Mn existing customer face matches facilitated in Q4 FY26 to streamline point-of-sale interactions. The same infrastructure that confirms an existing customer at the counter also catches an impersonation attempt at onboarding.

The geolocation layer is where it gets sharper. If an application claims residence in Pune but the device originating the application is in a tier-3 town with a known fraud signature, the model flags it before the bureau pull is even paid for. (Bureau pulls cost real money. A fraudulent application that triggers a bureau pull is not a write-off risk alone. It is also a wasted operating cost.)

L&T Finance and Poonawalla Fincorp have built parallel fraud layers. Poonawalla's 57-project AI programme covers fraud detection across multiple workflows. L&T Finance's data architecture flags inconsistencies in trust signals at application stage rather than at portfolio review.

The regulatory framing matters here. RBI's FREE-AI principles require explainability in fraud decisions. A borrower whose application is rejected on a fraud flag must, in principle, be informed in plain language why. That is harder than it sounds. A black-box neural network that flags fraud but cannot say why is not deployable under FREE-AI. The fraud models being built into Indian origination stacks today are, by design, interpretable. The model says fraud and points to the specific signal. Signature anomaly score X. Geolocation mismatch Y. Document tampering confidence Z.

Inline fraud detection is the only origination AI use case where the bar is set by the regulator, not the lender. FREE-AI makes interpretability a deployment condition. The lenders who built explainability in from day one will not have to retrofit it.

Each step in the origination process has its own AI use case, its own data dependency, its own measurable outcome. Voice bots at the top of the funnel. Multi-model underwriting in the middle. Vision AI on the documents. GenAI on the credit memo. Inline fraud at every stage. The lenders furthest along on this stack run all five in production. Most lenders run one or two and call it AI-first.

The infrastructure to close the gap is already public. AA is live. ULI is being scaled under Utkarsh 2029. The GSTN API is accessible. CKYC and Aadhaar are standardised. What separates the lenders building durable cost advantages from the lenders catching up is not access to data. It is the decisioning logic sitting on top of it.

That logic is built over time and not bought.

Sources:
Bajaj Finance Q4 FY26 FINAI Update;
L&T Finance Q4 FY26 Earnings and Project Cyclops / Project Helios documentation; Poonawalla Fincorp Q3 FY26 Update and Five AI Solutions Launch (January 2026); I
CICI Bank, Celent 2025 Model Bank, Analytics-Led Decision Making for SME Lending; Tata Capital, Sarvam AI Customer Story on Samvaad multilingual voice AI deployment (January 2025);
RBI Utkarsh 2029 (March 2026);
PwC / Dvara Research / FACE, Principles of Responsible and Trustworthy AI in Digital Lending (March 2026);
Jan Samarth Portal Public Data.