For banks, the value of data is no longer simply in knowing more about customers. It lies in turning that information into better and faster decisions, while ensuring that personalisation, automation and AI do not override customer choice.
At Axis Bank, this balance is shaping how data and AI are being applied across decision-making. Prasad Lad, Group Head, Business Intelligence Unit, Axis Bank, says the starting point is to recognise that not every banking decision requires customer-level data.
A significant amount of business intelligence is generated at an aggregated level. Branch and employee performance, marketing reviews and portfolio analysis, for instance, can be conducted at cohort or organisational levels without directly using personally identifiable information.
The responsibility becomes more acute when data is used to make an individual decision — whether to offer a product, run a marketing campaign or make a credit decision.
“A critical thing around this is two aspects: are you taking the customer’s consent? And then, how is your data being governed so that the consent is executed in reality?” Lad says.
That distinction is becoming increasingly important as regulatory expectations around data governance and consent become more formalised. For banks, responsible personalisation is therefore not simply about collecting consent; it is about ensuring that the consent remains connected to how the data is subsequently used.
Consent cannot be predicted
This principle becomes particularly important as AI takes on more decision-making.
Lad distinguishes between machine learning and generative AI. Machine learning models used for decisions such as credit eligibility, fraud detection or product propensity are generally deterministic in how they are deployed. They undergo governance, testing and checks around accuracy and data quality before being put into production.
Generative AI is different because its outputs are probabilistic. That means a human-in-the-loop mechanism remains important until the bank has sufficient observability and confidence in the output.
Models are back-tested against historical data and monitored after deployment to identify drift. For GenAI use cases, Lad says human oversight remains necessary until the organisation reaches a level of confidence where the output can be treated more mechanically.
The distinction can be seen in practical applications such as credit memos. A foundational model can summarise information that previously required a person to consult multiple sources and perform several analyses. But the final credit decision remains with a human.
“You have simplified that human’s work,” Lad explains, rather than eliminating the human from the decision.
That approach also reflects the regulated nature of banking. Where the bank develops a model for a specific use case, it needs to establish that the model has been adequately tested. When it uses a foundation model supplied as an enterprise service, the bank can rely on the provider’s model governance for the underlying model while still governing how that model is applied to its own use case.
From historical patterns to emerging needs
Data-driven banking is also becoming more responsive. Models that once refreshed annually or monthly can increasingly be refreshed weekly, allowing banks to incorporate more recent customer behaviour.
But historical data has limits.
A model can identify patterns associated with a market decline, for example, but may not fully understand an unusual external event that causes customers to behave differently from historical patterns. Lad points out that models predict based on historical behaviour and comparable cohorts; they do not automatically understand every new event that changes the context.
Even an individual customer can deviate from the behaviour of the cohort to which the model assigns them.
That makes predictive banking inherently probabilistic. The question, then, is where the bank draws the line between useful prediction and intrusive intervention.
For Lad, that line is determined by the customer.
“If the customer has defined the choice, that choice must govern how the bank communicates with them,” he says. A customer may consent to receiving proactive offers through a particular channel and later withdraw that preference. The system must then reflect the changed choice.
This creates an important separation: analytics can remain predictive, but consent cannot.
“Customer choice cannot be predictive. The customer choice has to be deterministic,” Lad asserts.
In other words, a model cannot infer that a customer who previously rejected marketing might now be receptive simply because similar customers changed their preferences. Consent has to be explicitly recorded and respected.
Banking intelligence is already becoming a shared layer
The future of intelligent banking may therefore not be about replacing existing banking applications with a single AI system. Instead, intelligence can increasingly operate as a layer across applications.
Lad points to Axis Bank’s use of a common business rule engine across different applications. Loan and card systems may remain separate, but decisioning logic such as allowable limits can be shared. The same underlying intelligence can also feed other engines, including fraud detection.
What changes with more advanced technology is the scale and granularity of that intelligence.
Today, Axis Bank creates thousands of customer features that can be used across decision-making. Future technology could enable vastly more features to be considered, allowing decisions to become increasingly granular.
But Lad does not expect technology alone to standardise the decision philosophy across banks. Each institution will continue to make its own choices about how much customer-level information it uses and how much decision-making is shared across products and services.
Nor does he expect banking to become a completely autonomous environment.
“The nature of human supervision and human in the loop could change,” he says, with humans potentially moving one level higher—from examining individual decisions to reviewing more summarised recommendations and policies.
Three priorities for intelligent banking
Lad identifies three broad priorities for the next phase.
The first is strengthening data security, privacy and governance as new technology stacks emerge. The objective is zero tolerance for data privacy or data violations, without allowing controls to undermine the speed of innovation.
The second is improving the quality, granularity and speed of decision-making. This means moving beyond today’s feature sets and refresh cycles while keeping the economics of AI and cloud compute under control.
The third is making structured data more accessible to GenAI systems. Generative AI is particularly effective with unstructured information, but Lad argues that making structured, deterministic banking data available can significantly improve the effectiveness of use cases such as voice bots and AI-driven customer interactions.
This creates a broader definition of intelligent banking. It is not simply about deploying more AI models. It is about connecting data governance, decision engines, customer preferences, human oversight and AI capabilities into a system that can make better decisions faster.
For Lad, the measure is therefore not how autonomous the bank becomes, but how effectively technology sharpens decision-making while preserving the principles that govern those decisions.
Intelligence in banking is ultimately about better decisions at greater speed and granularity, without making customer choices or responsible data use a variable.