A dispute AI agent integrates with the bank's core banking system (Finacle, BaNCS, FLEXCUBE, T24) and payment network APIs to pull real-time transaction data the moment a customer calls. The agent verifies customer identity via registered mobile OTP or MPIN, then retrieves transaction details — merchant name, amount, date, MCC code, and network — within 8 seconds.

Based on the dispute type identified through natural language ('I didn't make this payment', 'the ATM didn't give me cash', 'the merchant charged me twice'), the agent applies the correct regulatory framework automatically: RBI Zero Liability Circular for unauthorized card transactions, NPCI Dispute Resolution System for UPI failures, NPCI ATM Dispute Resolution for cash dispensing failures, or NACH dispute process for unauthorized ECS debits.

The agent collects all required intake fields — dispute amount, transaction date, merchant/beneficiary details, customer-stated reason, and whether the card/UPI instrument is still in customer possession — and generates a structured dispute record with a unique reference number. A confirmation SMS and WhatsApp message with the reference number and expected resolution timeline is sent immediately.

For eligible disputes under RBI Zero Liability, the agent triggers provisional credit initiation automatically. Escalation logic routes complex disputes — high-value fraud above ₹50,000, merchant representment cases, Ombudsman-bound complaints — to a specialist queue.

Once authenticated, the agent proactively reads out the last five transactions on the account, or allows the customer to describe the disputed transaction. Natural language understanding handles varied phrasings: 'the ₹2,400 Swiggy charge yesterday I didn't make', 'my ATM at Connaught Place didn't give me money but debited', 'I was charged twice for the same Amazon order.'

For card disputes, the agent confirms three RBI-required intake fields: whether the physical card is still in customer possession, whether the card was shared with anyone, and when the customer first noticed the discrepancy. For UPI disputes, the agent captures the UTR number or the recipient UPI handle.

The agent then presents a dispute summary for verbal confirmation: merchant name, amount, date, and dispute category. Upon confirmation, the dispute is logged with a microsecond-accurate timestamp — critical for establishing that the customer reported within the 3-working-day window for RBI Zero Liability eligibility.

For customers with multiple disputed transactions — common in subscription fraud patterns — the agent handles multi-transaction dispute batching in a single call, filing each transaction as a separate dispute record with linked reference numbers.

Transaction disputes span multiple payment rails, each with distinct regulatory frameworks and resolution timelines. Card disputes (debit and credit) fall under Visa, Mastercard, or RuPay network chargeback rules, with reason codes mapped to specific dispute categories. Unauthorized transactions trigger the RBI Zero/Limited Liability framework (July 2017 circular). Merchant disputes — goods not received, wrong amount, duplicate billing, quality disputes — follow standard chargeback reason codes (Visa 13.x series, Mastercard 4853, RuPay equivalent).

UPI disputes are governed by the NPCI Dispute Resolution System (DRS), with separate flows for failed transactions (bank error), fraudulent payments (social engineering), and merchant disputes (incorrect collection). The NPCI UDGAM (Unified Grievance Portal) is the escalation path for unresolved UPI complaints.

ATM disputes cover two scenarios: cash not dispensed (technical failure) where RBI mandates credit within 5 working days with ₹100/day penalty beyond that, and partial cash dispensing. NACH/ECS disputes cover unauthorized mandates and unauthorized debits against existing mandates — the agent verifies mandate status in real time via NACH mandate registry lookup.

NEFT/RTGS/IMPS failed transfers have a distinct flow under RBI payment system regulations: RTGS mandates 2-hour credit, NEFT mandates same-day credit, and IMPS is near-real-time with a 30-minute SLA for dispute acknowledgement.

The RBI framework establishes three liability tiers based on when the customer reports the unauthorized transaction relative to the bank's communication (SMS/email alert). Zero liability applies when the fraud is due to the bank's negligence or a third-party breach, regardless of reporting time, and when the customer reports within 3 working days of receiving the bank alert.

Limited liability applies when reported between 4 and 7 working days. Maximum customer liability caps are: ₹5,000 for Basic Savings Bank Deposit (BSBD) accounts; ₹10,000 for savings accounts, prepaid instruments, debit cards, and credit cards with limits up to ₹5 lakh; and ₹25,000 for credit cards above ₹5 lakh and corporate accounts. Beyond 7 working days, liability is per the bank's board-approved policy.

The AI agent enforces this by: capturing the exact timestamp when the customer reports; cross-referencing the bank's SMS alert timestamp in the communications log to calculate the reporting window; auto-classifying the applicable liability tier; communicating the provisional credit amount to the customer immediately; and initiating the provisional credit workflow with all regulatory fields pre-populated.

For cases where the bank's SMS alert was not delivered (network failure, number change), the agent applies the transaction date as the reference point and flags the case for compliance team review — ensuring customers are not penalized for bank-side communication failures.

Speed improvements come from eliminating three time-consuming steps: hold queue wait (average 4–7 minutes in peak periods), manual CBS transaction lookup (2–4 minutes by a human agent reading account statements), and manual form completion for dispute logging (3–5 minutes for a complete intake record).

The AI agent does all three simultaneously during the call: authentication happens during the opening greeting; CBS lookup fires the moment the customer describes the transaction; and dispute record fields populate automatically from the transaction data. By the time the customer finishes describing the issue, the agent already has 70–80% of the dispute record complete.

For complex cases — where the customer disputes multiple transactions, provides unclear descriptions, or the CBS lookup returns multiple potential matches — the agent handles disambiguation naturally, taking 5–8 minutes rather than 3–5, still well below the 12–18 minute human average.

The operational benefit extends beyond per-call time. Because dispute records are 100% structured and complete — no free-text notes requiring back-office interpretation — chargeback teams process cases 40% faster, and RBI TAT compliance improves from typical industry averages of 78–84% to 96%+ with Kallix deployment.

Provisional credit is a temporary credit to the customer's account during the dispute investigation period, mandatory under RBI guidelines for zero liability cases and regulated for ATM disputes, protecting customers from cash flow disruption while the bank investigates.

The AI agent's role is to ensure the triggering workflow is initiated correctly and completely. When an eligible dispute is identified, the agent submits a provisional credit request to the dispute management system with all required fields: customer account number, disputed transaction details, dispute category, RBI liability tier, complaint timestamp, and estimated credit amount.

The agent communicates the provisional credit timeline to the customer clearly: 'For unauthorized card transactions reported within 3 working days, your bank will credit ₹[amount] to your account within 10 working days as per RBI guidelines.' This sets expectations and reduces follow-up calls by 30–40%.

For ATM disputes, the 5-working-day TAT is tracked by the dispute management system. At T+3 working days without resolution, an internal escalation fires to the ATM operations team. At T+4, a senior manager is notified. If credit is not processed by T+5, the ₹100/day penalty is automatically added to the pending credit amount for each subsequent day — per RBI mandate.

UPI disputes fall into four main categories with distinct resolution paths. Failed transactions — where the customer's account is debited but the beneficiary doesn't receive funds — are the most common: the NPCI DRS automatically initiates a reversal if flagged within T+5 working days. The AI agent files the complaint, confirms the UTR number, and tracks reversal status.

Fraudulent UPI payments — where the customer was socially engineered via fake customer care, QR code scam, or collect request fraud — require different handling. The agent collects the recipient UPI handle, the circumstances of the payment (collect request vs push payment), and whether the customer approved a collect request unknowingly. This classification is critical for NPCI DRS categorization and potential IT Act Section 66C/66D referral.

Merchant UPI disputes (paid but service not delivered, wrong amount collected) follow a merchant dispute path via NPCI DRS, with the TPAP (PhonePe, Google Pay, Paytm) handling merchant-side investigation within 5 business days.

For wrong beneficiary payments — where the customer entered the wrong UPI handle — reversal is not guaranteed as NPCI treats UPI as irrevocable. The agent explains this clearly, initiates the formal dispute process, and advises the customer to send a written request to the beneficiary bank to freeze funds. The NPCI UDGAM portal is the escalation path if bank-level resolution fails within 30 days.

Card disputes trigger a formal chargeback process governed by Visa, Mastercard, or RuPay network rules. The bank (as card issuer) files a chargeback against the merchant's acquiring bank using a specific reason code: Visa 10.4 for card-absent fraud, Visa 13.1 for merchandise not received, Mastercard 4837 for no cardholder authorization, Mastercard 4853 for cardholder disputes. The chargeback shifts liability to the acquiring bank, which contacts the merchant for evidence.

If the merchant provides compelling evidence, they can re-present the chargeback (representment). If the issuer disagrees, the case escalates to arbitration by the card network. This multi-stage process takes 45–120 days, but the customer's provisional credit is maintained throughout. Card fraud disputes have a 65–80% recovery rate because the chargeback mechanism forces merchant liability.

UPI disputes are resolved through NPCI's DRS — a bilateral process between the sending bank and receiving bank. There is no equivalent to card network arbitration. For failed transactions, NPCI's system confirms the failure and initiates automated reversal. For fraud, funds can only be recovered if the beneficiary account is frozen before withdrawal — which depends on police action timelines and beneficiary bank cooperation.

The practical implication for customers: card fraud disputes have a materially higher recovery probability. Banks should communicate this distinction clearly during dispute intake so customers have realistic expectations about UPI fraud outcomes.

Correct reason code assignment at intake is one of the highest-value functions of an AI dispute agent. Misclassified reason codes are the primary cause of chargeback rejections — when a bank files Visa 13.1 (merchandise not received) for a transaction that should be Visa 10.4 (card-absent fraud), the merchant can defeat the chargeback by providing a delivery confirmation, and the customer loses.

The AI agent uses a decision tree driven by natural language signals. If the customer says 'I didn't make this purchase' or 'someone used my card details online' → fraud category (Visa 10.4, MC 4837). If 'I ordered but never received' → consumer dispute (Visa 13.1, MC 4853.1). 'They charged me the wrong amount' → Visa 13.3, MC 4853.3. 'They charged me after I cancelled' → Visa 13.2 (recurring cancellation). 'The quality was unacceptable' → Visa 13.5, MC 4853.4.

For RuPay, NPCI's dispute reason code framework maps to equivalent categories: unauthorized (U), merchant dispute (MD), ATM (AT), and technical decline (TD). The agent applies the NPCI-specific flow for RuPay cards, which differs from the Visa/Mastercard international chargeback cycle.

Correct reason code assignment at intake improves chargeback win rates by 18–25% compared to manual intake, because structured data prevents back-office agents from guessing categories after the fact. All AI reason code assignments are reviewable and overridable by the chargeback team before formal network submission.

ATM disputes are among the most time-sensitive under RBI regulations. The RBI ATM transaction failure guidelines mandate that banks credit the customer's account within 5 working days of the dispute being raised. Failure triggers compensation of ₹100 per day for each day beyond the deadline — payable automatically by the bank without the customer needing to claim it separately.

The terminal ID is the most critical piece of data the agent collects. It allows the bank to pull the ATM switch log and cassette reconciliation data — the primary evidence for ATM dispute resolution. The agent asks for the terminal ID visible on the ATM screen or transaction slip, or accepts the branch name and city as a fallback for customers who don't have the slip.

For partial cash dispensing (customer requested ₹10,000, received ₹8,000), the dispute is for the difference amount only. The agent calculates this automatically when the customer states both amounts.

The dispute management system tracks the 5-working-day TAT from the complaint timestamp. At T+3 working days without resolution, an internal alert fires to the ATM operations team. At T+4, a senior operations manager is notified. If credit is not processed by T+5, the ₹100/day compensation is auto-calculated and appended to the pending credit for each subsequent day. The customer receives a proactive call informing them of the delay and the compensation being accrued.

NACH and ECS disputes fall into two categories: debits made without any mandate (completely unauthorized), and debits that exceed the mandate amount or occur after mandate cancellation. Both violate NPCI NACH operating rules.

The agent's first action is a real-time NACH mandate registry lookup using the customer's account number and either the UMRN (Unique Mandate Reference Number) or the originator company name. This lookup confirms whether a valid mandate exists and what the authorized debit amount and frequency are. For business loans and insurance premiums with standing mandates, this prevents false disputes while ensuring unauthorized debits are caught immediately.

If no valid mandate exists, or the debit is clearly unauthorized, the agent simultaneously files a dispute in the dispute management system and triggers a stop-payment instruction to the NACH processing team. The stop-payment prevents any future debits from the same originator until mandate status is clarified.

For mandates that were cancelled by the customer but the debit continued, the agent retrieves the mandate cancellation record and files a chargeback against the originator through the NPCI NACH return process. For NBFC loan repayment NACH disputes, the agent distinguishes between a technical payment failure (where the customer wanted to pay but NACH failed) and a genuine dispute (unauthorized or excess debit) — a critical distinction that determines whether to route the customer to alternative payment or dispute filing.

Merchant disputes are governed by card network chargeback rules that require customers to attempt resolution with the merchant before a chargeback is filed for most dispute categories. The AI agent handles this requirement with a structured question: 'Have you contacted the merchant about this issue?' If yes, the agent collects the date, channel, and merchant response. If no, the agent advises merchant contact within 7–15 days while filing an initial dispute record as a precautionary measure.

For duplicate charges — the most common merchant dispute by volume — the agent verifies both transaction records in CBS, confirms the same merchant, same amount, and same or near-same date, and files Visa 12.6 / MC 4834 (duplicate processing). These disputes have among the highest win rates (85–90%) because duplicate transaction evidence is unambiguous.

For goods/services not received (Visa 13.1, MC 4853.1), the agent sets a 7-day window for the merchant to provide delivery confirmation before chargeback is escalated. If the merchant provides tracking showing delivery, the agent informs the customer and requests counter-evidence. This prevents chargebacks on delayed deliveries and reduces instances of friendly fraud that would otherwise result in merchant representment victories.

For wrong amount charges (Visa 13.3, MC 4853.3), the agent calculates the dispute amount as the difference between what was charged and what was agreed — not the full transaction. This prevents over-claiming and improves chargeback acceptance rates, since filing for the full amount when only a portion is disputed is grounds for network rejection.

The RBI Integrated Ombudsman Scheme (IOS), launched November 2021, merged the Banking, NBFC, and Digital Payments Ombudsman schemes into a single Centralized Receipt and Processing Centre (CRPC). Customers can file with the Ombudsman if the bank has not resolved the complaint within 30 days, or if the customer is not satisfied with the resolution.

The AI agent's TAT monitoring system flags complaints approaching the 30-day window at T+25 days — giving the bank a 5-day window to resolve before Ombudsman escalation becomes available. An internal alert reaches the dispute resolution team and the complaint manager simultaneously.

At T+30 days, if the dispute is unresolved, the agent initiates a proactive outbound call explaining: (1) the 30-day pre-complaint period has elapsed; (2) the customer is now eligible to file with the RBI Integrated Ombudsman; (3) no fees apply; (4) the SACHET portal is the online filing platform. The agent sends the SACHET portal link via WhatsApp and reads out the information the customer will need: complaint reference number, disputed amount, bank name, and branch IFSC.

The agent also generates a pre-filled complaint summary — a structured narrative with all transaction details, dispute timeline, bank communications, and outstanding amount — that the customer can attach to the Ombudsman complaint. Ombudsman complaints submitted with complete documentation packages resolve 40% faster than those filed without.

Failed and pending transactions are the highest-volume complaint category for most retail banks, accounting for 35–45% of all dispute calls. The AI agent handles these through real-time CBS integration, pulling current transaction status before the customer finishes explaining the issue.

For NEFT: the agent queries the UTR number in the NEFT outward message log. If the transaction was rejected by the beneficiary bank (invalid account, dormant account, account closed), the sending bank must return funds by end of business day. If the status is 'transmitted' but credit not yet confirmed, the agent advises the standard processing window and sets a monitoring alert for same-day follow-up.

For RTGS: the 2-hour credit mandate means any RTGS transfer showing unconfirmed credit after 2 hours is immediately complaint-eligible. The agent files the complaint at this point and the bank initiates enquiry with the beneficiary bank through the RBI RTGS system.

For IMPS: real-time credit is expected, but 0.3–0.8% of transactions remain stuck due to technical failures. The agent checks IMPS transaction status via NPCI API. For stuck IMPS transactions, the NPCI IMPS dispute portal is used for bilateral resolution between banks.

For card pre-authorization holds (hotels, fuel pumps, rental cars) showing as 'pending' beyond 7 days, the agent identifies the MCC-based hold expiry window (7–30 days depending on merchant category) and advises accordingly. For stale holds older than 30 days, a hold release request is filed with the acquirer.

Documentation completeness at intake is the single biggest driver of chargeback win rates and regulatory TAT compliance. Incomplete dispute records require back-office follow-up, which adds 2–5 working days to resolution time and introduces errors as customer memory of the incident fades.

The AI agent collects documentation in a structured sequence. Identity verification via OTP or MPIN is completed first and logged with a timestamp — establishing authenticated complaint initiation. Transaction details are auto-populated from CBS: merchant name, MCC code, transaction amount, currency, date, time, authorization code, and network. The customer's verbal description is transcribed and stored as the dispute narrative.

For fraud disputes, the agent additionally collects: whether any OTP was shared with a third party, whether the card was in the customer's physical possession at the time of the transaction, and the approximate time the customer first noticed the suspicious activity. These are the three liability determination fields under RBI Zero Liability guidelines.

Post-call delivery includes SMS with reference number and bank-specific dispute TAT within 60 seconds of call end; WhatsApp PDF with complaint summary, dispute category, liability tier (if fraud), expected provisional credit date, and bank escalation contact; and a formal email dispute acknowledgement as required under RBI customer communication guidelines. The WhatsApp delivery achieves 91–94% read rates within 2 hours, compared to 35–40% for email alone.

Status inquiries after dispute filing account for 40–50% of all dispute-related call volume. Customers typically call 2–3 times after filing: at day 5 ('has anything happened?'), day 15 ('still no update'), and day 25 ('this is taking too long'). Each of these calls is predictable and preventable with proactive communication.

The AI agent sends proactive status updates at five milestones: T+1 working day ('Your dispute is under investigation and has been assigned') — prevents first check-in call; provisional credit issuance date — voice call informing the customer of the exact credited amount, the highest positive NPS event in the dispute journey; chargeback filed with network — customer informed investigation is proceeding and why it takes 45–120 days; merchant response received — customer told whether the merchant contested or accepted; and final resolution with the outcome and any permanent debit or credit adjustment.

For inbound status checks, customers call and quote their reference number. The agent retrieves live dispute status from the complaint management system and reads out the current stage, last update, next expected action, and target resolution date — no human agent involvement.

For disputes approaching the 30-day Ombudsman window, the status update call proactively informs the customer of their rights — converting a potential escalation into a managed process. Banks deploying proactive dispute status calling see 35–40% reductions in repeat inbound calls and NPS improvements of 18–24 points across the full dispute journey.

Subscription disputes are a growing category as digital subscription services proliferate. Two sub-types are common: charges after cancellation (most common) and free trial conversions without adequate disclosure. Both have established chargeback paths but different win rate profiles.

For post-cancellation charges, the key intake fields are: when the customer cancelled, through which channel (app/email/phone), and whether the customer has a written cancellation confirmation. A written confirmation constitutes compelling evidence for the chargeback. The agent files Visa 13.2 / MC 4853.3 with the cancellation evidence attached.

For subscriptions billed via NACH mandate, the agent simultaneously checks whether the mandate is still active and, if so, initiates cancellation through NPCI NACH. This is communicated clearly to the customer: 'I'm cancelling your standing instruction to [merchant] right now, and separately filing a dispute to recover the unauthorized charge. These are independent processes — the cancellation happens immediately regardless of the dispute outcome.'

For free trial conversion disputes ('I didn't know it would auto-renew'), the resolution depends on the merchant's disclosure at trial sign-up. If renewal terms were clearly disclosed, the chargeback success rate is low. The agent informs the customer honestly, files Visa 13.7 (cancelled merchandise/services) noting inadequate disclosure as the customer's stated reason, and advises contacting the merchant for a goodwill refund — which merchants often grant to avoid the chargeback fee entirely.

High-value fraud disputes require faster intervention, law enforcement coordination, and specialist human review. The ₹50,000 threshold aligns with the amount at which Indian banks classify frauds as requiring senior review under RBI fraud reporting guidelines.

For transactions above this threshold, the AI agent's escalation protocol activates: the dispute is tagged 'high-value fraud priority' and routed to the bank's dedicated fraud investigation unit (FIU) with a 48-hour first-response SLA. A specialist analyst calls the customer within 30 minutes to gather additional forensic details not collected in the automated intake.

The agent provides the customer with a clear advisory on FIR filing: for cyber fraud above ₹50,000, filing an FIR at the local cyber crime police station and registering at cybercrime.gov.in (National Cyber Crime Reporting Portal, MHA) is strongly recommended. The FIR number, once obtained, should be submitted to the bank — it significantly improves chargeback outcomes because card network fraud liability frameworks prioritize cases with law enforcement documentation.

For SIM swap fraud (fraudster ports the customer's number to intercept OTPs), the agent provides a telecom provider escalation advisory: the customer should call the original telecom provider's fraud line to report the unauthorized SIM port and request service restoration — a step most fraud victims don't know to take, and one that stops the fraudster's access immediately.

Representment is the merchant's right to contest a chargeback by providing evidence that the transaction was valid. Common merchant evidence includes delivery confirmation with signature, customer IP address matching stated location, 3DS authentication logs showing OTP verification, recurring transaction terms accepted at sign-up, and authorization logs showing PIN entry for card-present disputes.

The AI agent's role in the representment phase is customer communication and evidence coordination. The bank's chargeback analyst makes the technical determination on whether the merchant's evidence is compelling, but the customer needs to be informed and given the opportunity to respond — particularly for consumer disputes where the customer may have counter-evidence.

The agent calls with a plain-language summary of the merchant's claim: 'The merchant says they have a delivery confirmation showing your name and address. Do you have any evidence that you didn't receive the package — photos of an empty letterbox, a conversation with the delivery service, anything like that?' This conversational approach extracts useful counter-evidence three times more effectively than email requests.

If the customer provides counter-evidence, the agent uploads it to the dispute case and triggers second chargeback (pre-arbitration) filing. If the customer cannot counter the merchant's evidence, the agent explains that the chargeback will be reversed, the disputed amount will be debited back, and advises on alternative resolution paths — consumer forum, FEMA, or FIR if fraud is genuinely suspected. Pre-arbitration win rates with AI-collected counter-evidence: 55–65% versus 35–40% without.

Complaint management integration is critical to the AI dispute agent's value proposition. Without integration, efficiency gains from the AI intake are partially offset by manual re-entry into the bank's CRM or complaint management system. With integration, the dispute record is created, structured, and routable the moment the call ends.

For banks using Salesforce Service Cloud (common in private sector banks), the agent creates a Case object with all standard dispute fields mapped — Account ID, contact, transaction reference, dispute category, reason code, dispute amount, regulatory TAT deadline, and priority flag. High-value fraud disputes are auto-assigned to the fraud investigation unit queue with a 48-hour SLA.

For banks on Freshdesk or Zendesk (common in newer fintechs and NBFCs), the agent creates a ticket via API with custom fields for banking-specific dispute metadata. Tags applied automatically based on dispute type enable intelligent routing without manual review.

For ServiceNow deployments (enterprise banks with complex ITSM setups), the agent creates an Incident or Request record in the financial services module, with escalation workflows pre-configured for regulatory TAT monitoring.

CBS integration is read-only during intake: the agent reads from Finacle, BaNCS, T24, and FLEXCUBE for transaction verification but does not write directly. All provisional credit and account adjustments flow through the bank's dispute management system with human approval before CBS posting — maintaining audit trail integrity and CBS controls.

Compliance in AI-handled dispute resolution operates across four dimensions: customer data processing, call recording and retention, outbound communication, and regulatory reporting.

Under DPDP Act 2023, dispute-related personal data requires explicit consent for processing. The AI agent collects DPDP consent at the start of every dispute call: 'This call will be recorded for quality and regulatory compliance purposes. Your dispute data will be processed to investigate and resolve your complaint. Do you consent?' This consent is logged with a timestamp against the dispute record for audit purposes.

Call recordings for dispute cases are retained for 8 years under RBI internal audit requirements — longer than the standard 2-year DPDP retention for general customer interaction recordings. The bank maintains separate data retention policies covering dispute records versus marketing calls.

For outbound dispute status update calls, TRAI TCCCPR 2018 applies. Dispute status calls are classified as transactional — not commercial — communications, exempting them from NCPR (National Customer Preference Register) restrictions and allowing calls to DND-registered numbers. Call timing is restricted to 9 AM–8 PM with the registered bank name as caller ID.

Under RBI Outsourcing Guidelines (2006, updated 2022), the bank retains full regulatory responsibility for dispute resolution quality even when an AI vendor handles intake. The bank's compliance team must approve the AI agent's dispute flow scripts, conduct quarterly audits of intake records, and maintain the right to override any AI-handled dispute.

Chargeback win rate is the percentage of filed chargebacks where the issuing bank succeeds in recovering funds from the acquiring bank. Indian retail bank industry average is 52–58%, with variation by category: ATM disputes (95%+ — unambiguous technical failure), card fraud (60–70%), and consumer disputes like goods not received (45–55%, most contested by merchants).

AI-assisted intake improves win rates through three mechanisms. First, correct reason code assignment: when a human agent logs a fraud dispute as a consumer dispute because the customer's language is ambiguous, the chargeback is vulnerable to merchant evidence that would not succeed against the correct fraud reason code. AI agents classify based on intent signals with 92–94% accuracy, reducing misclassification from the industry average of 18–22%.

Second, documentation completeness: chargebacks with incomplete intake records — missing possession status, no OTP sharing declaration, unclear cancellation date — are abandoned by the bank's own back-office as unfiled-able. AI intake eliminates incomplete records, increasing the proportion of disputes actually filed by 15–20%.

Third, speed: Visa's standard chargeback filing window is 30 days from transaction date in most categories; Mastercard allows up to 120 days. Human intake backlogs cause 8–12% of valid disputes to miss filing windows entirely. AI intake with same-call registration eliminates filing window misses.

Fraud incidents often repeat — either because the fraudster still has access to account credentials, or because the customer behavior that led to the initial fraud has not changed. The AI agent's post-dispute prevention protocol addresses both vectors.

Immediate protective actions triggered upon fraud dispute filing: (1) card hot-listing in the bank's CMS (TSYS/FIS/FSS) within 30 seconds — the physical and virtual card are blocked for all future transactions. The customer is informed and a replacement card request is initiated for delivery within 3–5 working days. (2) For UPI fraud, the customer's bank-linked UPI handle suspension is requested via NPCI API — preventing further UPI transactions from the compromised mobile number. (3) A fraud advisory WhatsApp message is sent within 60 seconds covering the specific fraud type and protective steps.

For social engineering fraud (customer tricked into sharing OTPs, approving collect requests, or installing remote access apps), the advisory is tailored to the specific vector: 'Don't share OTPs with anyone claiming to be bank staff — we will never ask for this' (vishing); 'Check the URL before entering card details' (phishing); 'No remote access apps are needed for banking — delete AnyDesk or TeamViewer immediately' (screen sharing fraud).

The T+24 hour follow-up call for social engineering fraud is a Kallix-specific protocol based on production data: 12–18% of social engineering fraud victims are contacted again within 24–48 hours by the same or associated fraudster, posing as a fraud recovery service ('We're from the bank's anti-fraud unit, we can recover your money if you pay ₹500 processing fee'). The proactive follow-up call explicitly warns against this secondary scam.

Language accessibility is a meaningful compliance and customer experience issue in Indian banking. A significant portion of bank customers — particularly in Tier 2–4 cities — are more comfortable describing complex financial grievances in their regional language. Forcing English-only dispute intake creates both an accessibility barrier and a documentation quality problem: customers describing fraud in a language they're not fluent in provide less specific, less actionable accounts.

The AI agent handles regional language dispute conversations with the same structured intake logic as English. The agent understands regional language financial terminology — Tamil banking customers use specific vocabulary for transaction types that doesn't map directly to English equivalents — and applies the same dispute categorization and reason code logic regardless of the language of the conversation.

For digital documentation, all formal dispute records submitted to card networks, CBS, and regulators are in English as required by Visa, Mastercard, NPCI, and RBI reporting standards. The customer-facing WhatsApp summary is generated in the customer's conversation language, ensuring they understand what has been filed on their behalf.

Language routing integrates with the bank's existing multilingual IVR infrastructure: if the customer has an established language preference in the bank's CRM, the dispute agent initiates in that language. Otherwise, auto-detection from the first sentence is confirmed with: 'I'll continue in Tamil / मैं हिंदी में बात करूँगा — is that okay?'

EMI and loan repayment disputes are distinct from card or UPI disputes because the resolution pathway is internal — within the same bank's CBS rather than across a card network or NPCI infrastructure. The agent's approach is verification-first rather than dispute-first.

The most common EMI dispute scenarios: (1) EMI debited but loan account shows outstanding — usually a NACH reconciliation posting lag; (2) double debit of EMI — duplicate NACH execution, common on month-end processing; (3) customer paid EMI via NEFT/UPI but NACH also executed — both payments collected.

For scenario (1), the agent pulls the loan account ledger in real time and checks posting status. Most apparent EMI disputes are resolved in the call itself — the agent shows the customer that the debit is pending CBS posting and will reflect within 24 hours. This resolves 40–50% of EMI dispute calls without any formal filing.

For scenario (2) and (3), the agent files an internal rectification request with the loan operations team, tags both duplicate transaction reference numbers, and initiates a credit back of the excess amount within 3 working days. For double NACH executions, the case is additionally flagged to the NACH operations team to investigate and prevent recurrence in the next cycle.

For pre-closure excess debits — where the customer closed a loan but NACH continued debiting — the case is escalated as a mandate cancellation failure. The loan operations team confirms whether the mandate was instructed to cancel at pre-closure, and the excess amount plus interest differential is returned within 7 working days.

RBI's Card-on-File (CoF) Tokenization mandate (effective October 2022) requires all card-on-file transactions at merchants to use network tokens instead of actual PANs. This improves security but creates complexity in dispute resolution because CBS displays a token reference rather than the traditional PAN.

The AI agent pulls token metadata from the card management system — the token maps to the actual PAN, card expiry, and the token requestor ID (the merchant or wallet app that created the token). This mapping is required for chargeback filing, which must reference the actual PAN per Visa and Mastercard network protocols.

For tokenized fraud disputes, an additional investigation step applies: the agent flags the token requestor for security review, as a token database breach at the token requestor's end is a potential fraud vector. If the breach is confirmed at the token requestor, fraud liability may rest with them under card network tokenization operating rules.

For disputes where the customer claims a merchant's stored card token was used without their knowledge for a new purchase, the agent collects whether the customer has revoked or deleted the stored card at the merchant — an important authorization distinction. A token the customer actively revoked being used for a subsequent charge is a strong unauthorized transaction claim; a token the customer left active is weaker.

NEFT, RTGS, and IMPS disputes involve interbank settlement rather than card network chargeback mechanisms. Resolution depends on bilateral communication between the sending and receiving banks, mediated by NPCI (for NEFT/IMPS) or RBI (for RTGS).

For NEFT: the agent queries UTR status in the NEFT outward message log. Transactions rejected by the beneficiary bank (invalid account, dormant account, account closed) generate a return message, and the sending bank must credit the customer by end of business day. NEFT now runs 24x7 in 30-minute clearing cycles since the RBI's 2019 round-the-clock upgrade — the agent informs customers that transfers work outside business hours.

For RTGS: the 2-hour credit mandate is strict. Any RTGS transfer showing unconfirmed credit after 2 hours of transmission is immediately complaint-eligible. The agent files the complaint at this threshold and the bank initiates enquiry through the RBI RTGS system.

For IMPS: real-time credit is the design intent, but 0.3–0.8% of IMPS transactions fail due to technical issues. The agent checks IMPS transaction status via NPCI API. For stuck IMPS, the NPCI IMPS dispute portal is used for bilateral bank resolution.

For wrong beneficiary payments, the agent explains clearly: 'UPI and IMPS transfers are irrevocable — there is no automatic mechanism to reverse funds once credited to a different account. I am formally requesting the beneficiary bank to return the funds with the account holder's consent. This process takes 7–30 days and depends on the beneficiary's cooperation. I also recommend filing a police complaint if the amount is above ₹10,000.'

Dispute management is one of the highest-cost functions in retail banking operations, combining high call volumes (12–18 minutes per dispute call), specialist staff requirements, and significant back-office processing. AI automation creates value across three cost centres: intake staff, back-office documentation, and compliance management.

Cost per dispute with AI: ₹120–₹200 covering all call handling, structured record generation, status update calls, and TAT monitoring. Cost per dispute with traditional call centre: ₹600–₹900 covering agent time at ₹30–₹50/minute loaded cost, quality review, and re-contact for missing intake information — the 20–30% quality rejection rate that requires callbacks is a major cost driver in the traditional model.

Back-office savings are significant: complete structured intake records reduce back-office processing time by 40%. A bank processing 5,000 disputes per month can redeploy 4–6 back-office FTEs from data entry and re-contact to investigation and resolution work — higher-value activities with better outcomes.

Chargeback recovery uplift generates ₹1.2–₹2.8 crore annually for a bank with ₹15–₹25 crore in annual disputed transaction value, based on the 12–18 percentage point win rate improvement. This revenue recovery partially or fully offsets the Kallix deployment cost on its own.

RBI ATM compliance: banks pay ₹100/day for every ATM dispute not resolved within 5 working days. For banks with 1,000+ monthly ATM disputes, even a 5% improvement in TAT compliance saves ₹18–₹25 lakh annually in mandatory compensation payments that would otherwise accrue.

Dispute resolution is among the more complex AI voice agent deployments due to the number of system integrations required: CBS (transaction lookup), CMS (card block), NPCI APIs (UPI DRS and NACH dispute filing), card network dispute systems, CRM (complaint logging), and the bank's dispute management platform.

Weeks 1–2 cover the integration layer: CBS read-only API for transaction lookup, CMS write access for card blocking (requires security review and approval from the bank's information security team), NPCI API credentials for UPI DRS and NACH dispute endpoints, and CRM dispute case creation API. This is the longest phase because it requires coordinated involvement from technology and information security teams.

Week 3 covers dispute flow scripting: the complete decision tree for all eight dispute categories, reason code mapping logic, escalation triggers (high-value threshold, Ombudsman TAT), and language variants for all 10 supported languages. Kallix provides pre-built templates for all major Indian banking dispute categories that are customized to the client bank's specific policies and thresholds.

Week 4 covers compliance review: the bank's legal and compliance team reviews dispute intake scripts, DPDP consent language, liability tier communication, and Ombudsman escalation advisory. RBI Outsourcing Guidelines require formal compliance sign-off before any production launch.

Weeks 5–6 cover UAT with the dispute operations team and a live pilot at 10–20% traffic. Kallix dispute specialists monitor all pilot calls in real time during the first week. Weeks 7–8 complete full rollout with ongoing quality monitoring and weekly performance reporting.

The comparison is not binary — it is a hybrid model where AI handles the high-volume, structured-intake portion, and human experts handle judgment-intensive, ambiguous, and high-stakes cases.

For the 60–75% of disputes handled end-to-end by AI: these are predominantly failed transactions (straightforward credit reversal), ATM disputes (technical failure confirmed by switch log), and duplicate charges (unambiguous CBS evidence). For these categories, AI handling is not just faster and cheaper — it is more consistent, because the decision logic is uniform and regulatory compliance is built in.

For the 25–40% escalated to human investigators: these are fraud disputes requiring forensic analysis, representment cases requiring legal judgment, high-value disputes above ₹50,000, and complex multi-account fraud patterns. Because AI has already completed intake, the human investigator receives a complete, structured case record — not a call recording to listen to and transcribe. This improves investigator productivity by 35–40% on the cases that actually need them.

Customer satisfaction: dispute journey NPS with AI handling (proactive status updates, 3–5 minute intake, same-day reference number) averages 42–48, compared to 22–28 for traditional call centre handling (hold times, incomplete documentation requests, status uncertainty). The NPS improvement is the second-largest customer satisfaction lever after the actual dispute win rate.