These query types represent 60–75% of total inbound call volume for most retail banks and NBFCs. By handling them through AI, banks eliminate hold time and the after-hours availability gaps that currently drive CSAT scores down. The agent reads live from your CBS via API — not from a cached dataset — so the balance it reads is identical to what your branch teller would see at the same moment.

Beyond read-only inquiry, the agent can trigger post-call actions configured by your policy: send a mini-statement to the registered mobile via SMS or WhatsApp, flag an unusual transaction for human review, or schedule a callback from a human agent for queries outside its scope. Every call is fully transcribed and logged with timestamps for your audit and regulatory reporting needs.

Authentication design is agreed with your compliance and IT security team during onboarding. The most common configuration for Indian banks is: caller ID must match the registered mobile number (first factor), one-time password sent to that number and confirmed verbally (second factor). A third factor is recommended for high-sensitivity queries such as large transaction history or full account statements — options include last 4 digits of debit card, date of birth, or a custom voice PIN.

The agent enforces a lockout after a configurable number of failed authentication attempts — typically three — and flags the account in your CBS for a security review. All authentication events are logged with timestamps and attempt counts for your audit trail. OTPs and PINs are verified against your authentication service in real time and never stored by the Kallix platform.

Containment rates vary significantly by query type. Account balance is the most contained (92% resolved by AI — nearly all callers get their answer without escalation). UPI transaction status is high at 85% because the status lookup is deterministic. EMI queries sit at 78% — some callers want to discuss restructuring, which requires human judgment. Disputes and fraud reports are intentionally low-containment at 35%: the AI triages and captures structured details, but resolution requires human authority.

The scope of what the AI handles fully versus what it triages is defined and approved by your operations and compliance team before go-live. We do not expand AI scope without explicit sign-off. The goal is a clearly bounded AI layer that your team trusts and can audit — not an AI that attempts everything and occasionally fails in ways that create regulatory exposure.

Integration is set up in onboarding weeks 1–2 in close collaboration with your IT and API banking team. The connection is read-only by design: the AI agent has no ability to initiate transactions, modify account records or access data outside the specific fields mapped in the integration specification. That spec is reviewed and signed off by your IT security and compliance teams before any live calls are made.

For banks without an existing API layer on their CBS, we build a lightweight middleware service that exposes specific account inquiry endpoints. This middleware sits within your infrastructure — on-premise or private cloud — and is not accessible from the public internet. We have completed this type of greenfield integration for four CBS platforms in production. Time to complete a new API layer is typically 3–4 additional weeks and is quoted separately.

The RBI outsourcing framework treats Kallix as an IT service vendor. The regulated entity — your bank — retains full oversight responsibility: you must maintain audit logs, ensure India data residency, include Kallix in your IT vendor risk assessment, and contractually require us to meet the security and confidentiality standards specified in the Master Direction. We make all of this operationally straightforward: our DPA is pre-aligned to the outsourcing direction's requirements, India data residency is the default for all banking customers, and we provide a vendor risk documentation pack for your assessment process.

The Digital Payment Security Controls direction applies specifically to payment status queries and any interaction involving account credentials. We comply with its requirements on caller authentication before sharing account data, call recording and audit trail, customer notification for sensitive data access, and incident reporting timelines. Our compliance documentation pack covering both Master Directions is provided during onboarding.

PCI DSS compliance for voice channels is governed primarily by Requirement 3 (protection of stored cardholder data) and Requirement 12 (information security policy). In a voice banking context the most common scope question is whether the AI captures and stores primary account numbers or card verification values spoken verbally. The Kallix platform automatically detects and masks these patterns in real-time transcription — they are never written to storage in plain text.

PCI DSS compliance is a shared responsibility. Kallix provides the technical controls; your bank holds the merchant or acquirer compliance posture. We have supported three banking customers through their annual QSA audit cycles, and our scope documentation is designed to make the QSA's assessment straightforward. We are clear about where our boundary ends and the bank's begins.

The DPDP Act 2023 designates the bank as the 'data fiduciary' — the entity that determines the purpose and means of processing customer data. Kallix is the 'data processor', acting strictly under the bank's instructions. Our DPA is structured accordingly: the bank retains full control over data access policies, retention periods, and the right to instruct deletion at any time.

For banking specifically, purpose limitation is the critical compliance requirement: account data accessed for a balance inquiry cannot be retained beyond the call session or used for any other purpose — marketing, analytics, model training — without explicit customer consent. We enforce this at the architectural level: each API call to the CBS is scoped to the specific fields needed for the active query, logged with purpose, and not retained after the session ends. DSARs are fulfilled within 72 hours.

Banking deployments run in a dedicated VPC with no shared compute with non-banking customers. The API connection to the CBS runs over an encrypted private tunnel or AWS Direct Connect — not the public internet. Role-based access control governs who within your organisation can access call recordings and transcripts: compliance teams see all; branch managers see their region only.

Penetration testing is conducted on all banking deployments before go-live and annually thereafter, with reports available to your IT security team. We hold ISO 27001 alignment (full certification in progress) and our security controls are reviewed as part of the banking onboarding process. For banks with regulatory requirements for dedicated infrastructure, a private deployment option is available on the Enterprise plan.

Fraud report calls receive elevated handling. When the AI detects keywords like 'fraud', 'unauthorised transaction' or 'didn't make this payment', it interrupts the standard authentication flow, requests minimal identity verification (registered mobile match only), and immediately begins capturing dispute details. The faster a fraud report is captured and escalated, the faster your team can initiate a card block or account freeze — speed matters for chargeback timelines.

For dispute calls where the customer is contesting a legitimate charge rather than reporting fraud, the AI completes full authentication, looks up the transaction in the CBS, reads back the details, and asks whether the customer wants to dispute or just verify the amount. This single step resolves a large proportion of 'dispute' calls that are actually balance confusion, reducing escalation volume for your disputes team by 20–30%.

Failed payment calls are the most common payment-related inquiry and among the most frustrating when handled poorly. The agent is configured with plain-language translations of the 20–30 most common UPI and NEFT error codes. Instead of reading 'U30 — Declined by beneficiary bank', it says 'The payment was declined by the receiving bank. This typically resolves within 2–4 hours if their bank is processing a backlog, and your money has not left your account.' That level of explanation eliminates most unnecessary escalations.

For debit-credit discrepancies — where the sender's account was debited but the beneficiary has not received funds — the agent confirms the debit, reads back the UTR, explains that NPCI has a defined reconciliation window for disputed UPI transactions, and offers to file a chargeback request or schedule a callback from the payments team. This scenario generates the most customer distress, and resolving it clearly on the first call is a primary CSAT driver.

In production, most customers don't know which payment rail was used — they know 'I sent money and it hasn't arrived'. The agent handles this ambiguity by asking the date, approximate amount, and whether it was sent through a UPI app, net banking or branch, then routes to the correct CBS query. The caller experience is 'describe your problem in your own words', not 'press 1 for UPI, press 2 for NEFT'.

RTGS queries — typically for transfers above ₹2 lakh — receive elevated handling given the amounts involved. The agent reads back full transaction details after completing full MFA and offers to connect to a relationship manager for any follow-up. For IMPS, which settles 24/7 in near real time, most queries resolve immediately because IMPS transactions either complete or fail within 30 minutes of initiation.

Loan account queries are the second most common inquiry type for NBFCs and third most common for retail banks. The vast majority — EMI due date, outstanding principal, next instalment amount — are deterministic lookups that the AI resolves in under 90 seconds. Customers particularly value 24/7 availability here: EMI anxiety peaks on evenings and weekends when call centres are understaffed.

For prepayment and foreclosure queries, the agent provides the current outstanding principal plus applicable prepayment penalty read from the loan terms in your LMS, clearly discloses that the figure is indicative as of today and accrues interest daily, and offers to send the exact amount to the registered mobile. It then offers to connect to the loans team to initiate the prepayment — handling the inquiry stage without human involvement while ensuring the transaction stage reaches a human.

Credit card queries are among the most time-sensitive banking inquiries: payment due dates are hard stops and confusion about minimum versus total due is a leading cause of avoidable late fees and credit score impact. The AI resolves these clearly — confirming both the minimum due and the full outstanding, explaining the interest implications of partial payment — and offers to send a statement to the registered channel for the caller's records.

For suspicious transaction flags, the agent follows the dispute triage flow: captures details, asks whether the caller wants to raise a dispute or verify the charge, and creates the appropriate case. For lost or stolen card reports, the agent immediately escalates to the card services team with the caller's authentication status and card details already captured, so the agent can initiate a block without asking the caller to repeat their information.

Containment rate is the percentage of calls the AI resolves without human transfer, and it is the primary metric banks track after deployment because it directly determines cost per call. The 65–80% range comes from production data across our banking and NBFC customers; variation is driven by query mix (a bank with a higher proportion of dispute calls will see lower overall containment), script quality, and how broadly the AI's scope is configured.

Containment improves over time. In month 1, it sits at the lower end of the range as the AI encounters edge cases not fully addressed during onboarding. By month 3, the weekly tuning cycle has resolved 90%+ of the original edge cases and containment typically reaches the upper end. We track containment rate as a formal SLA metric in every banking engagement.

The AHT reduction has three components. First, the AI completes authentication, CBS query and response in a single continuous interaction — no hold while a human agent waits for a system to load, no internal transfer to a specialist for a straightforward lookup. Second, post-call work is automated: transcript, case notes and CBS access log are generated and written to the CRM without any agent effort. Third, callers who reach the AI know their query will be resolved without hold time, which eliminates the time spent managing caller anxiety.

For your human agents, the indirect benefit is also significant: when AI handles 65–80% of routine inquiries, the 20–35% that reach humans are disproportionately complex cases. Agents develop deeper expertise in those case types rather than spending 80% of their day on balance inquiries. Human-handled AHT also drops as agents become specialists.

After-hours availability is the most immediately valued feature for banking customers in India. Banking contact centres typically run 8am–8pm or 9am–9pm, leaving a significant gap — particularly during festivals, month-end, and weekend afternoons — when customers cannot reach an agent. In most banking deployments we have run, 25–35% of total inbound inquiry call volume arrives outside standard business hours.

For after-hours escalations that cannot wait — fraud reports, urgent card blocks, account freezes — we configure a direct alert to the on-call officer via SMS and CRM notification, separate from the general inquiry queue. A fraud report received at midnight creates a high-priority case and immediately notifies the on-call fraud officer, so the case is actioned within the hour or at the start of business hours at the latest.

Language is a primary driver of banking CSAT in India, particularly for tier-2 and tier-3 city customers and rural segments served by NBFCs and MFIs. A customer asking 'mera balance kya hai' should get the same service experience as one asking 'what is my account balance'. Our Hindi and Hinglish models handle banking terminology — 'EMI', 'NEFT', 'passbook', 'KYC' — naturally because they were trained on real Indian banking call transcripts, not translated from American English.

For NBFCs and MFIs serving rural borrowers, regional language support is not a feature but a functional requirement. A microfinance customer in rural Maharashtra whose preferred language is Marathi cannot be adequately served by a Hindi-only system. Our beta regional language support is prioritised specifically for NBFC and MFI customers in tier-2/3 and rural markets. Language readiness for your target region is confirmed before any deployment commitment.

India has a large and growing segment of banking customers — senior citizens, rural depositors, first-time borrowers — for whom voice is the only practical channel for account inquiries. For these customers, the AI's consistent patience, willingness to repeat and confirm, and unlimited call duration is genuinely superior to navigating a DTMF IVR or waiting in a queue. The agent never rushes, never puts them on hold, and responds without frustration to a question asked three times.

For banking deployments targeting elderly or rural segments, we configure specific interaction parameters: slower speech rate, simplified vocabulary (banking jargon avoided where possible), confirmation prompts after every data point ('Your balance is ₹24,532. Shall I repeat that?'), and a lower threshold for human escalation — two failed interactions rather than three. These parameters are set per customer segment during onboarding.

Proactive outbound is one of the highest-ROI banking AI use cases, particularly for NBFCs with large EMI portfolios. An AI that calls every borrower 3 days before their EMI due date, confirms the amount, and offers payment options reduces late payment rates by 15–25% in production without adding headcount to the collections team. The calls are informational and helpful in tone — not collections-style pressure — which protects the customer relationship while improving repayment rates.

Large transaction alerts operate differently: the agent calls the account holder within 60 seconds of a transaction above the configured threshold, reads back the transaction details, and asks whether the customer authorised it. An affirmative response ends the call in 30 seconds. A denial immediately escalates to the fraud team with the call flagged as a potential unauthorised transaction — a significant fraud prevention capability that also improves customer trust.

WhatsApp delivery after a banking call significantly improves the practical value of the interaction: a customer who called to hear their last 5 transactions can also receive those transactions in a readable, shareable format without logging into net banking or visiting a branch. WhatsApp message open rates in India exceed 90%, making it far more effective than email for customers who don't regularly check their inbox.

All WhatsApp banking messages go through the WhatsApp Business API registered to your bank's number and verified badge — the customer sees the bank's official name, not an unknown number. Message content follows your security policy: account numbers truncated, transaction amounts shown in full, merchant names included. Full account numbers and sensitive personal data are never included in WhatsApp messages. WhatsApp template approval is handled during onboarding.

Most banks arrive at Kallix with an existing IVR handling basic call routing and first-factor authentication. We don't require replacing it. The conservative integration approach — IVR stays, does initial routing and caller ID verification, passes the call and auth status to Kallix for the conversation — limits the change surface and reduces the risk of disrupting existing call flows. This is what we recommend for the first 90 days.

For banks ready to replace the IVR, Kallix becomes the entry point for all inbound calls and handles language detection, authentication, inquiry resolution and escalation routing in one continuous interaction. CSAT typically improves significantly because callers reach the AI by speaking naturally rather than navigating menus. We validate AI performance in the IVR-as-front-door phase before any IVR decommissioning conversation.

The structural limitation of a DTMF IVR is that it can only handle options explicitly built into its menu tree. A caller who wants to know why a specific NEFT transfer is pending must navigate to the correct menu branch — or call back when a human agent is available. An AI agent understands the intent directly from natural language and executes the appropriate CBS query without any menu navigation.

The measurable outcome of this difference is containment rate: IVR systems typically contain 20–30% of inbound inquiry calls — callers who navigate to the right option, get the information they need, and don't request a human. AI agents contain 65–80% of the same call volume, a 3–4x improvement, because they understand intent rather than requiring callers to map their need onto a menu structure. IVR containment is capped by menu design; AI containment expands continuously with the knowledge base.

The typical NBFC serves thousands to hundreds of thousands of borrowers with a contact centre of 10–50 agents. The ratio of borrowers per agent makes routine inquiry handling operationally unsustainable at scale without AI. The most common NBFC use cases are: EMI due date confirmation before payment, payment receipt confirmation ('did my transfer go through?'), outstanding principal lookup, and NOC request status for closed loans.

For MFIs serving rural borrowers, the language dimension is a functional requirement, not a feature. A microfinance customer in Telangana who speaks Telugu cannot be adequately served by a Hindi-only contact centre. We prioritise regional language support specifically for NBFC and MFI customers in tier-2/3 and rural markets, and work with them on localised voice personas that are familiar and trusted in their target communities.

Banking deployments take longer than standard business deployments because of two additional work streams: CBS integration (setting up the read-only API connection, which requires your IT team's involvement) and compliance review (every script and authentication flow is reviewed by your legal and compliance team before any live calls). We coordinate these parallel streams during onboarding rather than running them sequentially, which keeps the 4–6 week timeline achievable.

The biggest variable is CBS API readiness. Banks with a mature API banking layer — typically those that have already built open banking or fintech integrations — complete the CBS connection in 1–2 weeks. Banks where the API layer needs to be provisioned or extended add 3–4 weeks to the timeline. We assess CBS API readiness in the first discovery call and provide a realistic timeline before any commitment is made.

In banking, a poor escalation experience is worse than no AI at all: a customer who authenticated successfully, explained their problem, and gets transferred to a human who knows nothing and asks them to start over is more frustrated than if they had reached the human directly. We treat warm handoff as a first-order requirement. The human agent's screen shows: caller name, account (last 4 digits), query type, authentication result, CBS data accessed, and a 3-sentence summary of what was asked and what the AI found.

Transfer routing matches your department structure. A caller disputing a UPI transaction goes to the payments team queue, not general enquiries. A caller asking about a loan modification goes to the loans servicing queue. If the AI detects fraud indicators, the call goes to the fraud team regardless of how the caller originally framed their query. Routing rules are defined and approved by your operations team during onboarding.

RBI's Integrated Ombudsman Scheme requires banks to have a documented complaint handling process with defined resolution timelines. The AI's role is triage and documentation, not resolution: it ensures every complaint is captured accurately, assigned a reference number before the call ends (giving the customer a tracking number they can follow up on), and routed to the right resolution team with complete context. This eliminates the most common failure mode in complaint handling — a misdirected transfer where the receiving agent has no context.

For customers in financial distress — overdue borrowers, customers affected by fraud, those disputing charges that impacted their ability to pay bills — the agent is configured to de-escalate in tone, acknowledge the stress explicitly, and prioritise connection to a human who can take action rather than providing only informational responses. A 'distress flag' triggers priority routing in the human agent queue for these calls.

Reporting serves two audiences in a banking context. Operations teams track containment rate, AHT, escalation rate and peak hour coverage — the metrics that determine how well the AI is performing and where to invest tuning effort. Compliance teams track authentication success and failure rates, CBS data access logs, call recording availability, and DND and consent adherence — the metrics that matter for regulatory review and internal audit.

The most operationally valuable output is the uncontained call analysis: every call that escalated to a human is tagged with the reason — out-of-knowledge-base question, authentication failure, escalation trigger, customer request. Reviewing this weekly reveals gaps in the knowledge base and authentication flow and drives the tuning agenda. By month 3, most banks have resolved 80%+ of the original escalation reasons through this iterative process.

ROI has three drivers. Cost savings: if the AI contains 70% of 10,000 monthly inquiry calls at ₹10/call versus ₹60/call for a human agent, the monthly saving is ₹3.5L. At Kallix's India pricing, full payback is typically achieved within 3–4 months. Scale savings: as call volume grows with customer base, AI cost per call stays constant while human cost scales linearly — the ROI improves as the business grows.

CSAT improvement is the second driver. In post-call surveys, customers who reached their banking query answered by AI within 90 seconds give higher satisfaction scores than customers who waited 5 minutes in a human queue for the same information. Lower CSAT translates to measurable customer retention improvement — and in banking, a percentage point of churn reduction has significant LTV impact. The third driver is compliance: a fully documented audit trail on every call reduces regulatory and audit cost.

The full cost picture: plan subscription plus per-minute overage if applicable, plus a one-time banking onboarding fee of ₹80,000–₹1.5L depending on CBS integration complexity. For a bank running 10,000 inquiry calls per month at an average AI handle time of 90 seconds (approximately 2,500 AI minutes per month), the Growth plan's included minutes cover a significant share, with modest per-minute overage. Total all-in: approximately ₹1.5L–₹2L per month.

Compare that to staffing a human contact centre team for the same volume: at Indian call centre rates of ₹18,000–₹35,000 per agent per month fully loaded, handling 10,000 calls at 4-minute average handle time requires 4–6 full-time agents — ₹72,000–₹2.1L per month — before supervision overhead, attrition cost, training cost, and without any after-hours coverage. The AI unit economics strengthen further as call volume grows.

We structure the banking pilot around two written success criteria: a containment rate target (typically 60%+ on scoped query types) and an authentication success rate target (typically 85%+). Both are agreed before work begins. At the end of the pilot, if both criteria are met, the bank has data to justify a production rollout to the full inquiry queue. If they are not met, we document the gaps, fix them for a second pilot run, or refund the pilot fee. We have had to do the latter twice across all banking engagements.

The pilot queue is deliberately narrow: a specific query type (balance only, or UPI status only) or a geographic segment (a single circle or branch cluster). Starting narrow allows us to validate the system rigorously before expanding. Pilot fee for banking is ₹1.5L–₹4L depending on CBS integration complexity and scope, credited in full toward the first month of production if you proceed.