Banking & NBFI
A Bangla voice brain on the hotline that verifies the caller, reads the balance from core banking, and never invents a figure.
A retail bank hotline in Bangladesh absorbs an enormous amount of traffic that is not advice and not a complaint. It is balance checks, last-transaction checks, card blocking, cheque-book status, branch and ATM locating, EMI dates and reset requests. Those calls are short, highly repetitive, and they arrive in Bangla, in English and in romanised Bangla inside the same sentence. They are also the calls that pin your agents during the hours when a genuine dispute or a fraud report needs someone experienced on the line. The engine we deploy answers that repetitive layer in Bangla, verifies who is calling before it says anything account-specific, and reads every figure it states out of core banking on that call rather than generating it. When the caller needs a human, the handover carries the verified identity and the transcript so the customer does not start again.
The problem before the product.
The hotline closes, the customer does not stop calling
Card blocking, fraud reports and balance anxiety do not observe business hours. An after-hours IVR that can only take a message converts a two-minute resolution into a next-day callback, and in the fraud case into a real loss.
A model that guesses a balance is a regulatory incident
Any system that states a figure on a recorded line has to source that figure from the system of record. Generic assistants will happily produce a plausible number. That is the single fastest way to turn an automation project into an audit finding.
Code-switching breaks generic speech stacks
Real callers say an account type in English, a complaint in Bangla and a digit string in either. Speech recognition tuned for one language at a time mis-hears the parts that matter most, which are the numbers.
Collections and reminders never reach the whole book
Outbound tele-calling floors work a list top-down and reach a fraction of it. The unreached remainder is not a calling problem, it is a capacity problem.
Built against your
systems of record.
Tiered identity verification
Low-risk intents answer after light verification. Anything account-specific requires a stronger tier before a single figure is spoken. The tiers are yours to set, and they are enforced in the flow rather than requested in a prompt.
Grounded figures, never generated ones
Balances, due dates, last transactions and card status are read from core banking through your integration layer at the moment of the call. If the system of record is unreachable, the engine says so and offers a handover. It does not estimate.
Outbound campaigns with consent and quiet hours
OTP confirmation, EMI and payment reminders, renewal and card-activation nudges, running against your list with retry policy, consent state and quiet hours honoured, and every outcome written back before the call ends.
Warm handover with context
When a call escalates, the agent receives the verified identity, the intent and the transcript. The customer is not asked to repeat the account number they already gave.
Append-only recording and full transcripts
Every call is recorded and transcribed to your storage, append-only, with the verification tier and the data sources used on that call attached to the record.
Your systems stay yours.
Nothing here is a replacement platform. The engine reads from and writes to what you already run.
- Core banking (through your existing integration or middleware layer)
- Card management and switch systems
- Existing PBX, contact centre and SIP trunks
- CRM and complaint management
- Collections and recovery systems
- bKash, Nagad and Rocket for payment confirmation flows
What it will not do.
The constraints are the product. A system that cannot be audited cannot be approved.
- Deployable on-premise or in your own cloud tenancy, which is what makes this viable for institutions with data residency obligations under Bangladesh Bank guidance.
- Customer PII is masked before any prompt leaves your boundary and re-hydrated only inside your systems.
- Append-only call recording and transcript retention, with retention windows set by you.
- Consent state is carried on every outbound campaign, with quiet hours and retry caps enforced by the engine.
- Every spoken figure is traceable to the system of record call that produced it.
Why this sector
works differently here.
The regulators, the systems and the operating reality that decide whether a deployment in this sector actually survives contact with your environment.
Bangladesh Bank sets the approval bar
The central bank regulates banks and NBFIs and issues the ICT security expectations your project will be assessed against. In practice this is why on-premise or own-tenancy deployment is not a preference here but the condition of getting approved at all, and why we design for it from the first diagram rather than retrofitting it.
MFS changed what customers expect from a hotline
bKash and Nagad trained the entire market to expect instant, self-service, round-the-clock money operations in Bangla. A bank hotline that closes at night is now measured against that, not against other banks.
Agent banking pushed the customer base rural
Agent banking extended formal accounts well beyond branch networks, which means a growing share of callers are Bangla-only, on mobile networks of varying quality, and unfamiliar with English banking vocabulary.
Digit strings are the hard part, not the grammar
Account numbers, card digits, amounts and dates spoken in mixed Bangla and English are where generic speech stacks fail, and they are exactly the tokens that must be exact. Recognition here is tuned for digits first.
Core banking is reached through middleware, not directly
Most institutions here front their core with an integration or ESB layer, and your security team will insist on it. We connect there, read-scoped to the specific fields a call needs, which is also what makes the access reviewable.
The first build, in about 30 days.
The after-hours retail hotline: balance, last transaction, card block and branch or ATM locating, with tiered verification and a human path throughout.
- Tiered identity verification you define
- Balance, last transaction and card status read from core banking
- Emergency card blocking with confirmation
- Branch and ATM locating
- Warm handover carrying verified identity and transcript
Scoping everything at once is how these projects fail. These come after the first build is answering real calls.
- Outbound collections and EMI reminder campaigns
- Loan application and sanction status
- Dispute and chargeback intake
- Cheque book and account opening workflows
If three of these are true, book the diagnostic.
Your hotline closes, or drops to a message-taking IVR, outside business hours.
A large share of answered calls are balance, status or location questions.
Your outbound floor cannot work the full reminder or collections list.
You have been told a cloud-only assistant cannot be approved for your data.
Agents spend the first minute of every call on identity verification.
Does this replace our contact centre platform?
No. It sits in front of it. Your PBX, your routing, your agent desktop and your reporting stay where they are. The engine answers the layer that does not need a person and hands the rest over with context.
How do you stop it stating a wrong balance?
It has no ability to produce a balance on its own. Account figures are template slots filled from your system of record during the call. If the lookup fails, the engine says it cannot confirm the figure and offers a handover.
Can it run entirely inside our data centre?
Yes. On-premise deployment is supported, and for most regulated buyers it is the reason the project is approvable at all. Scope and sizing come out of the diagnostic.
What about callers who switch between Bangla and English?
That is the normal case here, not an edge case. The engine is built for Bangla, English and romanised Bangla inside a single conversation, with particular attention to digit strings.
A diagnostic, then a written scope.
Forty-five minutes on your call mix and your systems. You get the findings in writing inside 72 hours, and they are yours whether or not we work together.