Key takeaways
- LLMs handle language. Personalisation that changes behaviour depends on enriched transaction data and decisioning, which LLMs do not provide on their own.
- Banks are mostly pointing AI inward: 85% of generative AI use cases at the 50 largest banks are restricted to internal functions (Evident, AI Outcomes Report, August 2025).
- The gap is in the data and decision layers. In a FICO poll of more than 30 senior APAC bank executives, only 43% said they use real-time data significantly or fully.
- A banking-specific AI banking platform should work alongside the LLM the bank already uses, not replace it.
- Personalised insights reach 15% to 30% engagement in Moneythor client reporting, against 1% to 2% for a traditional in-app banner.
Why this question is coming up now
Interest in AI in banking is shifting. In the US, searches for “AI banking” have risen by more than half in the past year, while searches for “generative AI in banking” have fallen by over 40%. Buyers are no longer asking what a model can do. They are asking what it takes to put AI in front of customers, safely and at scale.
That shift matters because the first wave of bank AI went inside the building. According to Evident’s 2025 AI Outcomes Report on the 50 largest banks, 85% of generative AI use cases are restricted to internal functions. Copilots for staff, document search and code assistants were the sensible place to start. Personalising the customer relationship is harder, because it needs more than language.
What a general-purpose LLM does well
Large language models are excellent at the communication layer of banking. They can draft a nudge in the right tone, summarise a month of spending in plain language, translate a message into Bahasa or Arabic, and hold a conversation about a customer’s account.
For many banks, that alone is valuable. Generative AI in banking is already shortening the time it takes marketing and product teams to produce and test content.
What an LLM does not do, on its own, is know which of a customer’s transactions this month was salary, which was a refund, and which recurring payment has just gone up. It does not decide whether this customer should see a savings prompt, a credit offer or nothing at all. And it does not check that decision against the bank’s product eligibility, consent and risk appetite before the message is sent.
Where personalisation actually breaks: data, decisions and control
Banks do not lack data. They lack the ability to act on it while the moment is still live. Three layers sit between raw data and a useful customer interaction, and none of them is the language model.
1. Financial data enrichment. Raw transaction descriptions rarely identify the merchant cleanly. A credit could be salary, a transfer or a refund. Transaction enrichment and categorisation turn this into context an AI can reason about. As open finance develops in markets such as the UAE, banks will receive more account data from other institutions, which makes reliable enrichment more important, not less. Deloitte’s 2026 Banking and Capital Markets Outlook, citing its own data and analytics survey, reports that more than 90% of data users in banks say the data they need is often unavailable or takes too long to retrieve.
2. Decisioning and next best action. Once the signal is clear, someone has to decide what to do. A falling balance could mean ordinary spending before payday, a planned purchase, financial pressure or a customer quietly moving to another bank. Next best action logic weighs the bank’s products, the customer’s situation and the commercial priority, and picks one action or deliberately picks none. That judgement is the bank’s competitive advantage, and it needs to be configurable by the bank, not improvised by a model.
3. Bank controls. Before anything reaches the customer, the decision has to pass eligibility, consent and risk checks, and leave a record of why it was made. Regulators expect this. Singapore’s new Guidelines on AI Risk Management, issued by MAS on 7 October 2026, make financial institutions accountable for AI used in their services even when a third party provides it.
An LLM can sit on top of all three layers to write the message. It cannot replace them.
LLM vs AI banking platform: how the roles divide
| General-purpose LLM | Generic customer engagement platform | Banking-specific AI banking platform | |
|---|---|---|---|
| Writes and adapts the message | Strong | Good, with templates | Uses the bank’s chosen LLM |
| Understands transaction data | No, unless enriched data is supplied | Limited: built on clicks, events and demographics | Yes: enrichment and categorisation built for financial data |
| Decides the next best action | No | Rules and campaigns, not financial signals | Real-time decisioning on financial events |
| Applies bank eligibility, consent and risk rules | No | Partly, outside the banking context | Built into the decision before delivery |
| Explains why a customer was selected | Hard to evidence | Campaign-level | Decision-level logic the bank defines |
| Deployment | Usually vendor cloud | Usually SaaS | SaaS, private cloud or on-premise |
The rise in searches for “customer engagement platform” shows how many teams are looking at cross-industry tools for this job. They are strong at orchestrating messages. They were not designed to read a salary credit, forecast a cash-flow shortfall or respect a credit policy.
Hyper-personalisation in banking depends on the layer below the model
Hyper-personalisation in banking is usually described as one message for one customer at the right moment. The model can make every message unique. Only the data and decision layers can make it relevant. That is why results come from timing and context rather than from the sophistication of the model.
In Moneythor client reporting, personalised insights reach 15% to 30% engagement, against 1% to 2% for a traditional in-app banner. Fee-related insights reach 33%, and savings insights 26%. Trust Bank used personalised, behaviour-linked customer journeys to generate US$1 billion in deposits within seven months, with up to 2.3 times more card transactions through personalised engagement.
How banks combine the two
The practical answer is not “LLM or platform”. It is a division of labour:
- The bank keeps the customer proposition, product rules, risk thresholds and the choice of LLM.
- The AI banking platform provides financial data enrichment, real-time decisioning, next best action and orchestration across the bank’s digital channels.
- The LLM provides language: drafting, summarising, translating and conversing, using the context and decision the platform has already produced.
Moneythor works this way. Its AI covers predictive, generative, conversational and agentic capabilities, and its generative tools work with a wide range of LLMs. The platform turns transaction data into insights, decides the next best action, and delivers it through the bank’s existing digital channels, deployed as SaaS, in a private cloud or on-premise. Content Sidekick tailors messages to each customer’s context, and Code Sidekick lets business analysts create personalisation logic from simple text prompts instead of code.
Who this approach is for, and who it is not for
Best for: retail and digital banks that want customer-facing AI personalisation built on their own transaction data, with control over decisions and deployment, and without rebuilding their core or replacing their LLM.
Not for: banks looking only for an internal productivity copilot, or a stand-alone chatbot with no link to customer financial data. A general-purpose LLM on its own is usually enough for those.
The model is the easy part
AI has made language cheap. Understanding a customer’s finances, deciding what the bank should do about it, and doing so within the rules is still the hard part, and it is where the results come from. Banks that treat the LLM as one layer of an AI banking platform, rather than the whole answer, will get personalisation that customers notice and the business can measure.
Talk to Moneythor about adding a banking-specific AI layer to the stack you already run.
Frequently Asked Questions
For writing and conversation, yes. For deciding what each customer should see, no. An LLM does not enrich transaction data, choose the next best action or apply the bank’s eligibility and consent rules. Most banks pair their chosen LLM with an AI banking platform that supplies the financial context and the decision.
An AI banking platform is software that turns banking data into customer action. It enriches transactions, detects financial signals, decides the next best action and delivers personalised experiences through digital channels, under the bank’s controls. Moneythor is a banking personalisation and decisioning platform of this kind.
A customer engagement platform orchestrates messages across channels and is usually built for many industries, using clicks, events and demographics. An AI banking platform is built around financial behaviour: salary credits, spending patterns, balances and recurring payments, and it applies banking rules before acting.
Look for four things: enrichment built for transaction data, real-time next best action decisioning, compatibility with the LLM the bank already uses, and deployment options that meet the bank’s data residency requirements. Ask for named customer results tied to deposits, card spend or activation.
Hyper-personalisation in banking means each customer receives insights, offers and guidance based on their own financial behaviour, at the moment it is relevant. It depends on enriched transaction data and real-time decisioning more than on the language model that writes the message.