How are banks using generative AI to personalise customer communications?
What the model does, what it needs underneath it, and how to get personalised messages live without a long data science project. · 7 min read
Banks use generative AI to write and adapt customer messages at the level of the individual: the same alert or nudge, phrased differently for a student, a working adult and a retiree, at the moment it is relevant. The model handles the wording. Deciding which customer should receive a message, what it should say and whether it should be sent at all depends on enriched transaction data and decision rules the bank owns. Banks that get results from generative AI in banking put a decisioning layer between their data and the model.
Key takeaways
- Generative AI personalises the wording of a message. The trigger, the audience and the next best action come from transaction data and the bank’s own rules.
- Banks are early. In research by FinTech Futures and Moneythor, banks said AI personalises an average of 38% of customer interactions, while they estimate 42% of customers expect AI-driven experiences.
- Much of bank generative AI has faced inward. Evident’s AI Outcomes Report, published in August 2025, found 85% of generative AI use cases at the 50 largest banks were limited to internal functions.
- In Moneythor client reporting, personalised insights reach 15% to 30% engagement, against 1% to 2% for a traditional in-app banner.
- An in-house data science team can build models. Getting personalised campaigns live also needs enrichment, event processing, decisioning and delivery, which is where most of the time goes.
Where generative AI is being used in customer communications
Most banks started with generative AI for staff: copilots, document search, code assistants. Customer-facing use is newer, and it tends to fall into four kinds of message.
- Proactive alerts and insights: a notification that spending in a category has risen, a bill is due before salary lands, or a subscription price has changed. Generative AI adapts the tone and length to the customer.
- Savings and money management nudges: prompts to create a budget, move a surplus into a goal or review recurring payments, worded for the customer’s situation.
- Product and offer messages: a relevant card, savings or lending offer explained in terms of what the customer has just done, rather than a generic campaign line.
- Conversational replies: in-app assistants and chatbots that respond to customers in real time.
In each case the model is doing language work. The part that makes the message relevant happens before the model is called.
What generative AI needs underneath it to personalise well
A language model on its own cannot tell which credit this month was salary and which was a refund. It cannot see that entertainment spending is up 20% on last month, and it has no view on whether this customer should get a budgeting nudge, a credit offer or nothing. Three things have to be in place first.
Enriched transaction data. Raw transaction descriptions are messy. Financial data enrichment cleans them, identifies the merchant and assigns a category, so a signal such as “spending on entertainment up 20%” exists in the first place. Moneythor’s enrichment engine combines customer-specific learning, rule matching on merchant category codes and transaction descriptions, optional machine learning models and third-party classification services.
Real-time decisioning. Once a signal exists, something has to decide whether to act, for whom and with what. That is next best action logic: the bank’s products, eligibility rules, risk appetite and commercial priorities applied to one customer at one moment. It should be configured by the bank, rather than improvised by a model.
Delivery through the bank’s own channels. The message has to reach the customer in the app, by push, email or SMS, or through a chatbot, through the bank’s existing infrastructure.
The FinTech Futures and Moneythor research shows why this matters. 61% of banking professionals said they cannot track the customer journey accurately, and 55% said they cannot deliver personalised experiences at scale. Those are data and orchestration gaps, and a better model does not close them.
An example: one trigger, three customers
Here is how Moneythor’s Content Sidekick handles a single rule. The bank defines the trigger and the action. The trigger is a customer spending 20% more on entertainment than they did last month. The action is a notification about the increase that encourages the customer to create a budget for expenses.
The decision is the same for every eligible customer. Content Sidekick then tailors the message to each one. A 20-year-old student might read: “Hi! Noticed your entertainment spending this month? It’s gone over 20% higher than before. Let’s chat about managing it better!” A 65-year-old retiree gets a more measured version: “Hello! Your entertainment expenses are 20% above average this month. Let’s discuss ways to adjust and manage your budget effectively.”
The bank keeps control of who qualifies and what action is offered. Generative AI changes only how the message reads. That division of work is what makes generative AI in banking safe to put in front of customers.
How business teams change messages and logic without code
A common bottleneck in bank personalisation is the queue for engineering time. A marketing or product team wants to change a threshold, add a merchant or test a new message, and the change waits for a release.
Moneythor Code Sidekick lets business teams make those changes with simple text prompts instead of code. A user types an instruction such as “Add a parameter to set the minimum amount of the eligible transactions triggering the nudge”, and Code Sidekick adds the parameter to the configuration form and updates the logic of the nudge. Business teams adjust messages and logic in a no-code interface, and Code Sidekick works with a wide range of large language models. See how Moneythor AI supports banking teams.
In-house data science team or vendor AI platform: which is faster?
For personalised campaigns, the answer usually depends on how much of the work is modelling and how much is plumbing.
An in-house data science team is strong at building propensity models and analysing customer behaviour, and many Tier 1 banks already have one. What slows personalised campaigns down is the work around the models: enriching and categorising transactions, processing events in real time, encoding eligibility and consent rules, connecting to every channel, and maintaining all of it after launch. BCG estimates that more than 60% of bank technology spend already goes to run-the-bank activity, and that only 20% of banks have robust data-quality frameworks. Each new internal build competes for that same capacity.
A vendor AI platform built for banking supplies the enrichment, event processing, decisioning and delivery, so the bank’s data scientists can spend their time on the judgement that differentiates the bank. Moneythor can also hook into the bank’s own analytical and machine learning models, so existing data science investment is used rather than replaced.
| In-house data science team | Banking AI platform | Hybrid | |
|---|---|---|---|
| Analytical and predictive models | Strong | Predictive AI that identifies patterns and anticipates behaviour | Bank’s own models connected to the platform |
| Transaction enrichment and categorisation | Has to be built and maintained | Included | Included |
| Real-time event processing and next best action | Has to be built | Included, configured by the bank | Included, configured by the bank |
| Generative content per customer | Has to be built and governed | Included (Content Sidekick) | Included |
| Changes by business users | Usually needs engineering | No-code, in Moneythor Studio | No-code, in Moneythor Studio |
| Delivery to app, push, email, chatbot | Has to be integrated per channel | Through the bank’s existing channels | Through the bank’s existing channels |
Keeping generative AI under the bank’s control
Regulators expect banks to remain accountable for AI used with customers. In Singapore, the Monetary Authority of Singapore’s Guidelines on AI Risk Management, issued on 7 October 2026, make financial institutions accountable for AI used in their services even when a third party provides it.
In practice, that means banks putting generative AI into customer communications need:
- decisions made by bank-defined rules before any text reaches a customer
- content and logic that business teams can configure, test and version themselves
- deployment that fits the bank’s data residency and security requirements
Moneythor Studio gives business teams a back-office environment to design conditions, configure customer-facing content, test in a sandbox and manage versions. Moneythor can be deployed as SaaS, in a private cloud, on public cloud or on-premise, and holds ISO 27001 and SOC 2 Type 2 certifications. Our guide to agentic AI governance in banking covers the governance questions in more depth.
What results look like
Generative AI improves the wording, and relevance and timing drive the results. In Moneythor client reporting, personalised insights reach 15% to 30% engagement, against 1% to 2% for a traditional in-app banner.
Trust Bank in Singapore, which runs personalised customer engagement on Moneythor, passed S$1 billion in deposits in May 2023, eight months after its September 2022 launch. At Chiba Bank in Japan, 79% of users who opened a campaign message completed the full engagement journey.
Frequently Asked Questions
Banks use generative AI to write and adapt alerts, insights, savings nudges, offers and assistant replies for each customer. The trigger and the decision about what to send come from enriched transaction data and the bank’s own rules. The model adapts the wording to the customer’s profile and situation.
A general-purpose LLM can write the message. It cannot enrich transaction data, decide which customer should get which message, or apply the bank’s eligibility and consent rules. Banks pair the LLM with a banking AI platform that supplies that context and decision.
Most banks get there faster with a hybrid. The data science team keeps ownership of models and customer strategy, and a banking AI platform provides the enrichment, real-time decisioning and delivery that take most of the build time. Moneythor can connect to a bank’s existing models.
Keep the decision with bank-defined rules, generate text only after a customer qualifies, keep a record of why each message was sent, let business teams test and review content, and deploy in a way that meets data residency requirements.
