azyware
User experience

Conversational banking: agents for balance, cards and disputes

EZ
Eazyware
· 7 min read
Quick answer

What should a bank know about conversational banking AI before launching an agent?

Conversational banking agents handle balances, card controls and disputes with identity checks and policy-gated actions. Read-only requests are answered after authentication; card actions run inside limits the bank sets; disputes are captured fully and handed to people with a summary, never decided by the agent.

Conversational banking AI is a set of agents, on WhatsApp, in the app, on the web and on the phone, that let a customer ask what their balance is, block a card, change a limit, or report a transaction they do not recognise, and get it done in the conversation. Each of those requests has a different risk. A balance enquiry is read-only after authentication. A card block is a write action that is safe to automate because it protects the customer. A dispute is a regulated process with timelines and evidence, so the agent captures it fully and hands it to people. The design follows that risk ladder.

This article explains what a conversational banking agent should and should not do, how identity and policy gates work across channels, why disputes are handled differently, and what the first deployment looks like for a bank or NBFC in India. It is written for digital, contact-centre and product heads.

Why conversational banking AI is different from a bank chatbot

The bank chatbot of the last decade answered FAQs and pushed everything else to a form or a call. Customers learned to type "agent" immediately. A conversational banking agent has access to the customer's accounts through scoped APIs, can take actions within policy, and escalates with context rather than a transcript. The difference is resolution: the customer leaves with the card blocked or the dispute filed, not with a link. We set out the general argument in AI agent vs chatbot; in banking the gap is wider because the actions matter more. Our broader approach to the sector is on the fintech industry page.

Intents by risk: what the agent may do

IntentRiskIdentity checkWhat the agent does
Balance, recent transactions, statementRead-onlyAuthenticated session or OTPAnswers directly from the core or middleware API
Block or freeze a cardWrite, protectiveAuthenticated plus confirmationExecutes immediately; confirms; logs
Unblock a card, raise a limitWrite, exposure-increasingStep-up authenticationExecutes within a policy limit; otherwise routes to a person
Change contact detailsWrite, fraud-sensitiveStep-up authentication and cooling periodCaptures the request; the change follows the bank's existing control
Report an unrecognised transactionRegulated processAuthenticatedCaptures every required field, blocks the card if the customer wants, files the dispute, hands off with a summary
Loan or product questionsRead-only, advisoryOptionalAnswers from approved product documents with citations; never quotes a rate it cannot source

Identity across channels

The same customer arrives on WhatsApp, the mobile app and the phone, and identity works differently on each. Inside the authenticated app the session is the identity. On WhatsApp the number is a hint, not proof, so the agent binds the number to the customer during a one-time enrolment and uses an OTP or app-based approval for anything beyond generic information. On voice the agent verifies with registered details and an OTP to the registered number, and never accepts a full card number or PIN. Step-up authentication is required for exposure-increasing actions on every channel.

The rule we apply is simple: the agent never has more authority than the customer has proven. If identity is weak, the agent is read-only on public information and offers the secure channel.

Policy-gated actions

Every write action goes through a policy gate that sits between the agent and the banking API. The gate encodes what the bank has decided: a card can be blocked at any time; a temporary limit increase is allowed up to a defined amount and duration; an unblock requires step-up; a contact change follows the existing control. The agent proposes the action with the evidence from the conversation; the gate checks it and either executes or routes to a person. The model never calls the core banking API directly, and the gate's decision is logged with the reason. The pattern is described in policy-gated actions.

Disputes: capture completely, decide by people

A transaction dispute has regulatory timelines, evidence requirements and a customer who is usually anxious. The agent's job is to make the first ten minutes good: confirm the transaction the customer means, ask whether the card should be blocked, collect the required fields (what happened, when the customer noticed, whether they shared credentials, whether they have contacted the merchant), file the dispute in the system of record, and tell the customer the reference number and what happens next. The disputes team receives a complete, structured file with a summary instead of a half-filled form and a call back. The agent does not assess liability or promise an outcome.

Fraud-in-progress

If the conversation indicates fraud in progress, such as multiple unrecognised transactions in the last hour, the agent prioritises blocking the card and connecting the customer to a person immediately, over collecting fields. This is a designed escalation path, tested in evaluation, not something the model decides on its own.

Language, tone and the Indian customer

Customers write in English, Hindi, regional languages and mixtures of them. The agent should understand code-switched text and reply in the customer's language, with numbers, dates and amounts formatted the way the bank's statements format them. Tone is plain and calm; no exclamation marks in a dispute conversation. Multilingual handling and the models we evaluate for it are covered in multilingual customer support with AI for Indian businesses.

What the contact centre gets

Escalations arrive as a case with the transcript, the identity level reached, the actions already taken, and a two-line summary. Agents stop asking customers to repeat themselves. The weekly escalation review, owned by a contact-centre manager, decides which reasons become new intents and which stay human. Metrics are per intent: resolution with no reopen, CSAT on agent-handled conversations, escalations and reasons, and actions outside policy, which should be zero.

A worked example

A bank wanted to move card-control and dispute-intake volume off its phone lines. The first release covered balance and transaction enquiries, card block, and dispute capture on WhatsApp and in-app chat, with voice added later on the same intent logic. Identity was bound at enrolment with OTP step-up for anything beyond read-only. Card blocks executed instantly; unblocks and limit changes routed to a person in the first release while the policy gate was observed in shadow mode. Dispute files arrived in the disputes queue complete, with a reference number already given to the customer. The qualitative outcome was that the disputes team spent less time on call-backs for missing information, phone volume on "block my card" fell to the point that the IVR option was demoted, and the compliance team could show a per-conversation record of identity level, action and gate decision. The design borrowed from our WhatsApp agent for a D2C brand for channel handling and from the multilingual voice agent for a hospital network for language coverage.

Team and timeline

Conversational banking agents are built under the customer service agent service from $12,500 / ₹8L for a single-channel text agent, rising with intents, channels and integrations; voice is added under voice agents from $17,500 / ₹11.2L plus usage at $0.05–0.15 per minute. Your side provides a product owner, a contact-centre lead, API access through middleware, an identity and fraud contact, and sample conversations. We provide an AI engineer, an integration engineer and a conversation designer. A first release with three to five intents on one or two channels typically takes eight to twelve weeks including shadow mode. All figures and Care Plan tiers are on the pricing page.

Before you start: a checklist

  • List intents and classify each as read-only, protective write, exposure-increasing write or regulated process
  • Define identity levels per channel and which actions each level unlocks
  • Write the policy for every write action: limits, step-up, cooling periods
  • Confirm middleware or API access to accounts, cards and disputes systems
  • Collect real conversations and call transcripts per intent for evaluation
  • Agree the dispute capture fields with the disputes and compliance teams
  • Decide the languages for launch and the fallback for unsupported ones
  • Name the contact-centre manager who owns the weekly escalation review

Questions clients ask

  • Should we launch on WhatsApp or in the app first? In-app has the strongest identity; WhatsApp has the most reach. Most banks launch read-only and card block on both and keep exposure-increasing actions in-app.
  • Can the agent handle UPI disputes? It can capture them completely and file them; liability assessment and the resolution timeline remain with the disputes team.
  • What about customers who just want a person? They get one, with context, and the transfer reason is counted so the intent can be improved or left human.
  • How do we stop the agent inventing product terms? It answers product questions only from approved documents with citations and refuses when it cannot source an answer.

The WhatsApp Business Platform documentation is the primary reference for what the channel permits. Related posts: WhatsApp AI chatbot for business, how to build an AI agent that is safe to run unattended and the fintech industry page.

Match authority to proven identity, gate every action, capture disputes completely and hand them to people, and conversational banking becomes a channel customers prefer rather than avoid.

Frequently asked questions

Is a conversational banking agent safe for card actions?

▾

Blocking a card is protective and safe to automate after authentication. Unblocking or raising limits increases exposure and needs step-up authentication and a policy limit; outside that limit a person decides. Every action is logged with the gate decision.

Can the agent resolve a dispute?

▾

No. It captures the dispute completely, blocks the card if the customer wants, files it in the system of record and gives a reference. Liability and resolution remain with the disputes team under regulatory timelines.

Which channels should a bank support first?

▾

In-app chat and WhatsApp for text, then voice on the same intent logic. Identity strength differs per channel, so exposure-increasing actions usually stay in-app in the first release.