India's personal loan AUM reached Rs 15.9 lakh crore as of September 2025, growing 13% year-on-year. Fintech-led NBFCs alone sanctioned 10.9 crore personal loans worth Rs 1,06,548 crore in FY 2024-25. Every loan generates a servicing tail: EMI queries, payment confirmations, balance statements, part-prepayment requests, NOC applications, and grievance escalations. The volume is not small.

The cost of a human-handled inbound servicing call at an Indian bank or NBFC ranges from approximately Rs 80 to Rs 150 per interaction, depending on call duration and agent tier. At the scale of a large NBFC with 30 to 40 million active loan accounts, even a 10% reduction in human-handled interactions saves hundreds of crores annually. The cost argument for AI servicing does not require a strategic vision document. It just requires a spreadsheet.

Bajaj Finance has the most publicly documented numbers. In Q4 FY26, AI voice and text bots handled 72% of DIY customer servicing. AI voice bots replaced 1,500 calling agents. Rajeev Jain stated on the Q4 FY26 earnings call that an AI call centre agent costs one-third of a human agent. The FY27 target is that any customer communication across any channel routes through AI first, with human intervention reserved for genuine complexity.

The broader cost signal is the opex-to-NTI ratio. Bajaj Finance's opex-to-NTI improved to 32.8% in Q3 FY26 versus 33.1% in the prior year, with management attributing the improvement to AI adoption across servicing, underwriting, and collections. The 25 to 40 basis point annual improvement target is the financial expression of what AI servicing produces at the cost line.

The servicing cost advantage of AI is not theoretical. It is being realised, quarter by quarter, in basis point improvements on the income statement.

Not all servicing interactions are equal in complexity or risk. The AI servicing architecture in Indian lending has settled into three distinct tiers, each with a different human involvement model.

Tier 1 is fully automated: EMI amount queries, due date reminders, outstanding balance checks, statement downloads, part-prepayment calculations, NOC status requests, and loan closure confirmation. These are closed-ended, data-lookup interactions with deterministic answers.
AI handles them at high accuracy with no human in the loop. The bot either retrieves the right number or it does not. There is no judgement required. A lot of these tasks are already handled by IVR- the key would be to perform with a higher effectiveness and achieve higher C-SAT.

Tier 2 is AI-assisted: top-up eligibility checks, loan restructuring information, interest rate queries on floating rate products, and complex grievance triaging. AI retrieves the relevant data and drafts a response. A human reviews before the final communication goes out. The AI reduces the human's workload from research plus drafting to review plus send. This is AI- augmented layer, where the human gives in the key inputs based on their experience.

Tier 3 is human-required: legal notices, deceased account management, serious complaints, and RBI ombudsman referrals. AI triages and routes these to the correct human team with the relevant account context pre-populated. The human does not start from scratch.

India's major bank servicing bots map across these tiers with varying depth.

None of these bots handles all three tiers. Each is optimised for a capability and channel subset. The architecture question for a lender deploying servicing AI is not which bot to pick. It is how to design the handoff between tiers so that Tier 1 deflection is maximised, Tier 2 throughput is improved, and Tier 3 routing is fast and accurate.

The bot that deflects a Tier 1 query saves Rs 100. The bot that misroutes a Tier 3 grievance and delays an RBI ombudsman referral costs Rs 10,000 in potential fine exposure. Tier design matters more than bot selection.

India's lending market is not English-first, and it is not Hindi-first either. A two-wheeler loan borrower in Coimbatore wants to query their EMI in Tamil. A gold loan customer in Surat expects to be understood in Gujarati. A microfinance borrower in Bhubaneswar does not necessarily read Roman script.

A servicing bot that cannot respond in the borrower's preferred language defaults to two outcomes: it routes to a human agent, which eliminates the cost advantage, or it produces an answer in a language the borrower cannot fully understand, which is a Fair Practices Code accessibility concern.

SBI's SIA can respond in 14 languages via speech or text. Bank of Baroda's bob SAMVAD covers 22 Indian languages in real-time communication, making it the most linguistically capable deployment in the Indian banking system as of March 2026. Kotak Keya operates in English and Hindi only. ICICI iPal supports vernacular languages with voice capability, though the depth of language model quality across all supported languages is not publicly disclosed.

The conversational AI market in India was valued at Rs 38.1 billion in 2024 and is projected to reach Rs 152.3 billion by 2030 at a 26.22% compound annual growth rate. That growth is being driven partly by the recognition that linguistic depth is not optional for a market this linguistically diverse. The lenders whose servicing bots cover the full language range of their borrower base will have a structural advantage in servicing cost and borrower retention over those whose bots cover English and Hindi and route everything else to a human.

A servicing bot that works in multiple languages is not just a cost-saving tool for a lender but also increases effectiveness multiple times. As most lenders are restricted from offering multiple languages due to their sub-scale nature in centralised call centres.

.

The cost reduction case for AI servicing is clear and already being demonstrated. The revenue generation case is where the architecture gets more interesting, and more regulated.

Cross-sell in lending is the offer of a related financial product to an existing borrower: a personal loan top-up to a home loan customer, a credit card to a two-wheeler loan repayer, a fixed deposit to a business loan borrower who has maintained clean repayment for 18 months. The signal that makes the offer relevant is the same signal that makes it possible: the lender already has the borrower's repayment history, transaction data, and AA cash flows.

The AI-first cross-sell architecture has three components.

The propensity model analyses transaction behaviour, repayment history, product holding pattern, life-stage signals, and AA cash flows to predict which product a borrower is most likely to accept and when. The model does not offer a personal loan to a borrower who took one six weeks ago. It does not offer a credit card to a borrower whose spending patterns suggest they are managing existing obligations at the margin. Done correctly, the propensity model filters out the offers that would increase the lender's risk rather than the borrower's debt burden.

The pre-approved offer engine generates a credit line offer for existing borrowers with clean repayment history, without requiring a new application or human underwriting. HDFC Bank offers instant disbursal of up to Rs 40 lakh in 10 seconds for eligible existing customers. SBI's YONO platform generates pre-approved personal loans within minutes by analysing transaction history and credit scores. The offer is calibrated to demonstrated repayment capacity, not to a manual credit officer's availability.

The Next Best Actions system is the most sophisticated layer. HDFC Bank's NBA system analyses transaction and digital behaviour data for each customer and recommends the most relevant product or intervention at the optimal moment. The recommendation is not static. A customer whose grocery spend pattern has shifted upward over three months receives a different next best action than one whose discretionary spend has declined. The model reads the signal and times the offer accordingly.

Bajaj Finance's cross-sell targets make the revenue potential explicit. The cross-sell ratio is targeted above 60% by FY29, products-per-customer targeted at 6 to 7 from 6.15 currently, and AI bots are targeted to generate Rs 12,103 crore in disbursals in FY27, up from Rs 1,895 crore in Q4 FY26 alone. Muthoot Finance, with gold loan AUM of Rs 1,39,658 crore as of Q3 FY26, is investing in AI-powered analytics specifically for cross-selling financial products to its existing gold loan customer base.

The DPDP constraint runs through all of this, and it is not a footnote. Data collected for loan servicing e.g. transaction queries, EMI payment behaviour, interaction logs, bot conversation history cannot be used for cross-sell propensity modelling without a separate, specific consent record stating that purpose. The lender whose privacy policy contains a broadly worded consent that covers both servicing and marketing is not compliant with the DPDP Act's purpose limitation principle. The consent for the servicing stack and the consent for the cross-sell propensity model must be distinct, with distinct retention schedules.

There is also a right-to-opt-out requirement. A borrower has the right to opt out of AI-driven communication and of automated profiling for marketing purposes. A cross-sell architecture that does not honour that opt-out, or that does not make it easily accessible, creates a regulatory exposure under the DPDP Act that compounds with every unsolicited offer generated.

The practical architecture consequence: the cross-sell propensity model cannot draw from the same data pool as the servicing model without a documented consent bridge. Many lenders building AI cross-sell capabilities have not yet separated these two consent layers. That is a build-now problem, not a wait-and-see one.

The AI cross-sell engine is not a marketing tool attached to a servicing platform. It is a separate data architecture with its own consent foundation, its own model governance, and its own regulatory exposure if either is missing.

Parts 3 and 4 described the credit underwriting stack and the fraud and compliance stack as four-layer architectures sharing the same data foundation. The servicing stack maps onto the same logic, with two additional components specific to the servicing function.

The interaction layer handles the borrower-facing interface: multilingual AI voice and text bots for Tier 1 and Tier 2 queries, with a mandatory human escalation path for Tier 3. The FREE-AI framework's human oversight principle requires that no servicing function be fully automated without a documented human review path for high-stakes interactions. An automated loan closure confirmation is low-stakes. An automated response to a disputed EMI deduction is not.

The data layer is the same unified customer data model that feeds the credit model and the fraud model. Repayment history, transaction patterns, product holding, AA flows, and servicing interaction logs all sit on the same architecture. The consent record for each data element specifies its permitted use: credit assessment, fraud detection, or servicing. Cross-sell requires a separate consent record with a separately stated purpose.

The propensity and personalisation layer runs the Next Best Actions models, the pre-approved offer engine, and the dynamic offer timing logic. This layer is governed by the separate cross-sell consent, operates on a separate data partition from the servicing layer, and has its own FREE-AI governance requirements: explainability of offer logic, opt-out mechanics, and a bias audit to ensure the propensity model is not systematically directing high-margin offers away from segments that are equally creditworthy.

The feedback layer closes the loop in two directions. Servicing interaction outcomes : resolved at Tier 1, escalated to Tier 2, routed to Tier 3, bot-failed and feeds that back into bot training on a documented retraining schedule. Cross-sell acceptance and rejection rates feed back into the propensity model. Both loops are governed by the FREE-AI retraining schedule requirement from the board-approved AI governance policy.

The connection to credit and fraud stacks is not incidental. The same data layer governs credit, fraud, compliance, and servicing. The same governance architecture applies across all four functions. An AI-first lender does not build four separate data stacks and four separate governance frameworks. It builds one stack, one governance architecture, and deploys separate models against each function with function-specific consent records and accountability owners.

The servicing stack is the fourth application of the same architecture that first appeared in Part 3. The lender that has built the data layer correctly finds servicing AI straightforward to deploy. The lender that has not built it finds every new function requires a new foundation.

The servicing function is where the AI investment from Parts 2, 3, and 4 compounds. The credit model that approves the right borrower creates a cleaner servicing portfolio. The fraud model that catches bad applications at origination reduces dispute volume post-disbursement. The compliance architecture that classifies SMA-0 in real time gives the servicing team a day's head start on a deteriorating account.

This is why the series has been arguing for a single data architecture across all four functions rather than function-specific stacks. The compounding is only possible if the layers talk to each other.

Part 6 will cover collections. By that point in the series, everything built in Parts 2 through 5 — the origination data, the credit model, the fraud flags, the servicing interaction logs, the SMA classification — feeds into the collections model's first decision about how to reach a borrower who has missed a payment. The lender that has built each layer correctly arrives at collections with a significant advantage. The lender that has patched each layer separately arrives with a significant debt.

Sources:

 Bajaj Finance Q4 FY26 FINAI Update (provided);
Bajaj Finance Q3 FY26 Investor Presentation (Medianama);
RBI Financial Stability Report 2025;
Analytics Vidhya HDFC Bank Case Study (August 2025);
IJFMR Research — AI in Indian Banking (SBI, HDFC, ICICI, Kotak, 2025);
IndianWeb2 — From Call Centers to Chatbots (May 2026);
RoboticsBiz — Top Banking Chatbots India;
Biznessidea — ML Applications in Indian Banking Sector (May 2026);
Appwrk — AI in NBFC Lending (March 2026);
ResearchAndMarkets — Conversational AI India 2025-2030 (August 2025);
DPDP Act 2023; RBI FREE-AI Committee Report (August 2025)