Adding AI to a banking app is mostly an integration problem, not a model problem. The model understands what the customer wants. Everything that touches money has to go through the bank’s existing systems, with clear rules about what the AI may and may not do.
Here’s how that played out in a real project: an AI chatbot and voice agent for Aman Bank in Libya, which serves around 750,000 customers.
The stack
Layer Technology
Mobile apps Native Android (Kotlin) and iOS (Swift)
Chatbot Python with a GPT-based language model
Voice Google Speech-to-Text, Text-to-Speech, NLP
ML TensorFlow
Data / real time MongoDB, WebSocket
Infra / delivery AWS, Jenkins, Git, Jira
How a request flows
The customer types or speaks in the app.
Voice input goes through speech-to-text.
The language model works out the intent (check balance, transfer, pay bill, request statement) and extracts details such as the amount and payee.
The backend calls the right banking API. The model never computes a balance or decides whether a transfer is allowed.
The response is phrased in plain language and, for voice users, turned back into speech.
WebSocket keeps the conversation and alerts real time, for example low balance or suspicious activity.
Rule 1: the LLM classifies, the core system answers
The safest pattern is to treat the language model as a router. Here’s an illustrative sketch of the idea (not the production code):
python
INTENT_HANDLERS = {
"check_balance": banking_api.get_balance,
"recent_transactions": banking_api.get_transactions,
"pay_bill": banking_api.pay_registered_biller,
}
def handle(message, customer):
intent, slots = llm.classify(message) # model only labels intent and extracts fields
if intent not in INTENT_HANDLERS:
return escalate_to_agent(customer, message)
if requires_confirmation(intent):
return ask_customer_to_confirm(intent, slots)
data = INTENT_HANDLERSintent # source of truth
return llm.phrase(intent, data) # model only phrases the result
The model never invents a balance, and anything outside known intents goes to a person.
Rule 2: confirm before money moves
Transfers and payments need an explicit confirmation step: amount, payee and account read back to the customer, plus the bank’s normal authentication. Voice makes this more important, because speech-to-text can mishear numbers.
Rule 3: voice needs its own testing
Speech recognition accuracy varies by language, dialect, accent and background noise. Test with real customer recordings, not clean samples. Keep spoken responses short. Long answers are hard to follow by ear.
Rule 4: design for low-end devices
Many customers used feature phones or had limited data. Native apps help, as does keeping heavy AI work server-side.
Results the bank reported
60% faster onboarding (with OCR-based digital KYC)
50% less workload for human support agents
40% faster average response time
Full case study: Aman Bank generative AI mobile banking platform. If you’re building something similar, this kind of work sits at the intersection of AI software development and native mobile engineering.
Top comments (0)