The banking industry is entering an important phase of artificial intelligence adoption. Large language models (LLMs) and generative AI have created enormous excitement, with banks experimenting with AI assistants, conversational banking, automated document generation, coding copilots, research tools, and customer-service applications.
These capabilities are powerful. But there is a risk of confusing what is technologically impressive with what creates the greatest economic value.
For banks, the most important AI investments are often not the newest LLM features. They are the less glamorous capabilities underneath them: financial data analytics, business intelligence (BI), machine learning, data governance, and decision infrastructure.
The reason is straightforward.
Banking is fundamentally a data-driven decision business.
Banks make money by pricing risk, allocating capital, managing liquidity, understanding customers, detecting fraud, controlling costs, and making better decisions than competitors. These activities depend heavily on high-quality data and analytical models.
LLMs can make those capabilities easier to access. They do not eliminate the need for them.
1. Banking Runs on Financial Data
Every major banking activity generates and depends on data. Consider the information required to manage a corporate lending portfolio:
- Customer financial statements
- Loan balances
- Interest income
- Funding costs
- Credit ratings
- Collateral
- Repayment history
- Industry exposure
- Probability of default
- Capital consumption
- Profitability
- Market conditions
The value is not created simply by storing this information. The value comes from turning data into decisions.
A bank needs to know which customers are becoming risky, which industries are deteriorating, which relationships are profitable, where capital is being consumed, and where pricing needs to change. That is the domain of financial analytics, BI, and machine learning. An LLM can explain the answers in natural language, but the analytical infrastructure must first determine what those answers actually are.
2. BI Directly Supports Management Decisions
Business intelligence remains one of the most important technologies in banking because executives and managers need reliable visibility into the business.
A CEO may ask:
Which businesses are driving revenue growth?
A CFO may ask:
Why did operating expenses increase this quarter?
A Chief Risk Officer may ask:
Which portfolios are showing deterioration?
A relationship manager may ask:
Which of my clients are becoming less profitable?
These are not primarily language-generation problems.
They are data and analytical problems.
A strong BI platform can provide standardized metrics, dashboards, drill-down analysis, historical trends, alerts, and management reporting. An LLM can make that information easier to access. For example, instead of navigating five dashboards, an executive could ask:
“Show me the three regions where corporate lending profitability declined most this quarter and explain the main drivers.”
That is a valuable LLM use case. But the LLM is only useful because the bank already has:
trusted data + governed metrics + analytical models + BI infrastructure.
Without those foundations, the chatbot becomes a sophisticated interface to unreliable information.
3. Machine Learning Solves High-Value Banking Problems
Machine learning has a different role from generative AI.
LLMs are particularly good at working with language, documents, and unstructured information. Machine learning is often better suited to prediction, classification, anomaly detection, and optimization. That makes ML highly relevant to core banking economics.
Examples include:
Credit Risk
ML models can estimate the probability that a borrower will default or become financially distressed.
Fraud Detection
Models can identify unusual transaction patterns that may indicate fraud.
Anti-Money Laundering
ML can help prioritize suspicious activities for investigation.
Customer Churn
Models can identify customers who are likely to leave or reduce their relationship with the bank.
Collections
Predictive models can help determine which customers are most likely to repay and which intervention is likely to be effective.
Cash-Flow Forecasting
ML can help predict future cash requirements for customers and businesses.
Next-Best Action
Models can identify relevant products or services based on customer behavior and financial needs.
These applications can directly influence revenue, credit losses, operating costs, capital allocation, and risk. That is why ML remains strategically important even in an LLM-driven AI environment.
4. Advanced LLM Features Do Not Automatically Create Business Value
The banking industry is currently seeing rapid development in LLM capabilities:
- Longer context windows
- Multimodal models
- Agentic AI
- AI coding assistants
- Autonomous workflows
- Advanced reasoning
- Voice interfaces
- AI research assistants
- Multi-agent systems
These technologies can be very useful, but a more advanced AI model does not always create more financial value.
For example, a bank might spend millions on an advanced AI assistant that helps employees find information faster. That can improve productivity.
But another bank might use a simple machine-learning model to improve credit decisions across a $100 billion loan portfolio. Even a small improvement in risk prediction could save the bank much more money.
The key point is :
AI value should be measured by business impact, not by how advanced the technology looks.
5. Data Quality Is a Bigger Bottleneck Than Model Capability
One of the biggest challenges in banking AI is not the model itself, but the quality and architecture of the data that supports it.
Large banks typically have data spread across core banking systems, operational databases, data warehouses, data lakes, and legacy applications. Traditionally, much of this data was designed for transaction processing, reporting, and regulatory requirements. Modern analytics, machine learning, and AI often require data to be more integrated, standardized, well-governed, and easily accessible.
For example, customer, transaction, credit-risk, and financial data may be stored in different systems using different formats and definitions. Before this data can be effectively used for ML or AI, it often needs to be integrated, cleaned, transformed, and governed.
A simple way to view the foundation is:
Data Sources → Data Warehouse/Lake → Data Integration & Quality → Analytics/ML → AI Applications
This is why data architecture is so important. A more advanced AI model cannot compensate for incomplete, inconsistent, or poorly structured data.
JPMorganChase is a good example. The bank has made significant investments in data and analytics infrastructure as part of its broader AI strategy. Its Corporate & Investment Bank has reported more than 175 AI use cases in production, including KYC, fraud detection, client recommendations, cash-flow prediction, and research.
The key lesson is clear:
Before banks invest in increasingly advanced AI models, they need a strong data foundation that supports traditional analytics, machine learning, and modern AI applications.
6. DBS Shows What AI at Scale Looks Like
DBS Bank provides another strong example of the importance of building the foundation.
The bank has spent years developing enterprise data, analytics, and AI capabilities. Its data platform and governance approach are designed to make information more reliable, accessible, and usable across the organization.
In 2025, DBS reported more than 2,000 AI models across more than 430 use cases, generating approximately SGD 1 billion in economic value from data analytics and AI/ML initiatives. The bank has also expanded its generative AI capabilities.
This illustrates an important sequence:
Data foundation → Analytics → Machine Learning → Governance → Generative AI
GenAI is not replacing the analytical foundation.It is becoming another layer on top of it.
7. The Most Valuable AI Architecture Is Not LLM-First
A useful way to think about the future banking architecture is:
Trusted Data
↓
Financial Analytics & BI
↓
Machine Learning & Predictive Models
↓
Business Decisions & Workflows
↓
LLM / GenAI Interface
Each layer has a different responsibility.
Data
Provides the factual foundation.
Analytics and BI
Explains what happened and what is happening.
Machine Learning
Predicts what is likely to happen and identifies patterns humans may miss.
Business Workflows
Turn analytical insights into decisions and actions.
LLMs
Allow employees and customers to interact with this intelligence naturally through language.
This is a much more powerful model than treating an LLM as the entire AI strategy.
8. LLMs Create More Value When Built on Quality Data
Banks should not slow down their adoption of generative AI. Instead, they should focus on connecting LLMs with trusted data, analytics, and machine-learning models.
For example, a corporate banking executive could ask:
“Which sectors are showing the fastest deterioration in credit quality?”
A well-designed AI system could then:
- Retrieve trusted portfolio and credit data.
- Analyze changes in exposure and delinquency.
- Run relevant credit-risk models.
- Identify sectors where risk is increasing.
- Compare current results with historical trends.
- Present the findings in simple language through the LLM.
- Provide links to the underlying reports and data.
In this approach, the LLM is the interface, while the data, analytics, and ML models provide the intelligence.
This is where generative AI becomes truly valuable in banking: it does not replace financial analytics or machine learning; it makes trusted financial intelligence easier to access, understand, and use.
9. The Real Competitive Advantage Is the Data and Decision Infrastructure
There is another reason banks should not focus exclusively on increasingly sophisticated LLM features.
Foundation models are becoming increasingly accessible. A bank can obtain powerful models from major technology providers without developing the underlying model itself.
What is much harder to replicate is the bank’s:
- Proprietary financial data
- Customer history
- Transaction data
- Risk models
- Analytical capabilities
- Data governance
- Domain expertise
- Regulatory knowledge
- Business processes
- Decision infrastructure
This creates an important strategic distinction.
The LLM may be a powerful technology, but the bank’s data and decision infrastructure can become the competitive moat.
Two banks may have access to similar foundation models.
But the bank with better data, better analytics, better ML models, and better integration into business processes can potentially create much greater value from those models.
10. Measure AI by Financial Outcomes, Not Model Sophistication
Banking executives should therefore evaluate AI initiatives using business metrics.
Instead of asking:
How advanced is the model?
Ask:
- Did credit losses decline?
- Did fraud losses decline?
- Did revenue increase?
- Did customer retention improve?
- Did operating costs decrease?
- Did relationship managers become more productive?
- Did risk identification improve?
- Did decision-making become faster?
- Did regulatory and operational controls improve?
This changes the conversation from AI experimentation to measurable business transformation.
A simple machine-learning model that reduces fraud losses by millions of dollars can be more valuable than an advanced generative AI application that produces impressive demonstrations but has limited impact on the bank’s economics.
Conclusion
The future of banking AI is not a competition between LLMs and traditional analytics. The real opportunity is to combine them.
Financial data analytics tells the bank what is happening.
Business intelligence makes that information visible and actionable.
Machine learning predicts what may happen next.
Data governance ensures that the information and models can be trusted.
Business processes turn intelligence into decisions.
And LLMs provide a powerful natural-language interface to the entire system.
The strategic mistake would be to treat the LLM as the foundation.
The stronger approach is to build the foundation first and then use generative AI to unlock it.
The winning banking architecture is therefore:
Trusted Data → Analytics & BI → Machine Learning → Decision Intelligence → Generative AI
The question for banks is no longer simply:
“How can we deploy the most advanced AI model?”
It is:
“How can we build the most intelligent, trusted, and measurable decision system—and use AI to make it accessible to everyone?”
That is where the long-term value of AI in banking is likely to be created.
Read More:
14 best LLMs for data analysis: Model overview, comparison, and use cases for business intelligence
AI Agents in Data Science: How Autonomous Systems Are Transforming Analytics in 2026
