AI will expose banks that have not fixed the foundations

Written by Russell Pereira

Jun 23, 2026
10 min

Technology leaders from Hong Leong Bank, Maybank, State Bank of India and Ant Digital Technologies argue that artificial intelligence is forcing banks to confront the operating foundations that will determine whether automation becomes a source of advantage or exposure.

William Streitberg, chief information and technology officer at Hong Leong Bank, gave the bluntest warning: "If you don't get into AI, your company is going to be consumed by another one. It's that simple." The urgency was real, but the message was not that banks should race to deploy artificial intelligence (AI) models as quickly as possible. AI is moving from a support tool into an execution layer, changing the risk profile for banks because it can now influence decisions, trigger workflows, generate code and act across connected systems.
That capability creates direct exposure when a bank has not defined what AI is permitted to do, what data it may consume, what systems it may reach and who remains accountable when it acts. Hong Leong Bank’s risk appetite approach, Maybank’s data governance work, State Bank of India’s push towards real-time infrastructure and Ant Digital Technologies’ AI maturity framework all point to the same institutional test. Banks must govern authority, information, infrastructure, security and human judgement as one operating system before they can safely give AI greater autonomy. AI does not create that institutional discipline by itself, and banks that have not defined the foundations for autonomy will expose those weaknesses as they scale it.

Banks must decide what AI is allowed to do before it acts
Streitberg said Hong Leong Bank starts every AI initiative with a risk appetite statement because the bank must define what AI may and may not do before teams test its capabilities. Older models usually support analysis, recommendations or workflow optimisation, while agentic AI can initiate actions if a bank gives it system access and permissions. A bank therefore needs to decide where AI can assist, where it can recommend and where it must not act without human approval.
"Sometimes you're so busy trying to work out whether you can do it, sometimes you need to actually ask yourself should you do it before the dinosaurs get loose," Streitberg observed. He cited a small company that gave an AI tool broad permissions, after which the system sold the company’s shares because the organisation had not set the right guardrails. The example illustrates how AI can create losses even when it follows the permissions it has been given, turning risk appetite into an operating control rather than a governance formality.
Streitberg described the practical questions a risk appetite statement must resolve: "Are you willing to go and give a computer admin passwords? Are you willing to allow it to go and sell and trade shares?" The same logic applies to credit decisioning, customer remediation, treasury operations, software deployment, fraud intervention and access to core banking systems. Banks that grant AI authority without defining boundaries may not suffer from a rogue model in the dramatic sense, but from a compliant system that acts within poorly designed limits.

AI can only be as reliable as the information a bank gives it
Simon Lim, group chief data officer at Maybank, widened the argument from risk appetite to strategic intent, alignment and data governance. He argued that banks must first decide why they want to use AI and how each use case fits the group’s wider strategy. Without that alignment, business units can embed AI in different ways, using different definitions, controls and success measures that produce scattered experiments rather than bank-wide capability.
The data challenge then becomes unavoidable because AI can only answer from the information the bank gives it. Lim described data governance as the part of AI programmes that institutions often treat as plumbing, invisible until something breaks. Maybank encountered that problem when it loaded standard procedure instructions (SPI), the documents governing branch and operational conduct, into a retrieval-augmented generation (RAG) system that retrieves source documents for an AI model to answer user queries.
The system produced unexpected responses, but Lim’s team traced the weakness back to the source documents rather than the model itself. "When the indigestion happens, as in the hallucination or the problems that the AI is facing, you are able to trace that back to where actually the problem exists," he noted. Maybank found that some SPI documents were 30 years old, lacked the structure required for reliable retrieval and had never been prepared for machine consumption, showing how poor internal documentation can become poor automated advice when a bank feeds it into AI.
Regional banking adds another layer of complexity because the same operational concept can carry different terminology across markets. Maybank operates across Singapore, Indonesia and Malaysia, where teams may use different glossaries, process language and local documentation for similar banking activities. Human staff may understand those differences through experience, but AI systems can treat them as separate or conflicting inputs unless the bank standardises language and maps lineage across jurisdictions.
Lim also placed people at the centre of the data problem. "There's not enough of those people who understand what needs to be done to sit across from a group of sophisticated AI specialists who is trying to drive this forward," he said. The point connects data governance to institutional judgement, not just model performance. Reliable data serves AI fully only when the underlying systems can supply it in real time, making batch-based infrastructure inseparable from the question of whether banks can let AI act when decisions matter.

Real-time banking decides whether AI can act when it matters
Balaji Rajagopalan, chief technology officer of State Bank of India, linked AI readiness to the infrastructure that supports daily banking operations. Many institutions still rely on end-of-day batch processing, where core systems close, reconcile and update through overnight jobs. That operating model may have supported traditional banking, but it constrains AI systems that need live data for fraud monitoring, customer service, credit decisions, payments and operational interventions.
"It is, I would say, it's a survival for banks in future," Rajagopalan observed of the shift away from batch-based operating models. Post-pandemic customer behaviour has made contactless and digital service availability a baseline expectation rather than a channel upgrade. A bank that closes critical systems at midnight for reconciliation cannot support AI-driven decisions during that window because the model lacks a complete live view of the customer, transaction or operational state.
Rajagopalan identified several requirements for banks that want to operate in that environment. They must modernise core systems for 24/7 availability, use application programming interfaces (APIs) across the complete product lifecycle and re-engineer customer journeys rather than digitise old paper-based processes. They must also build zero-trust architecture into every connected system, meaning each interaction requires authentication, authorisation and context rather than relying on trust granted earlier in the journey.
His argument makes infrastructure readiness a business condition for AI adoption because modernisation has to reach beyond channels into the processes, systems and teams that carry the full product lifecycle. Banks that treat infrastructure as a separate technology prerequisite risk separating AI ambition from the operating reality that must support it. Real-time capability therefore becomes part of the business model for AI, not a back-end upgrade that can follow later.

The next risk is how much autonomy banks are prepared to grant
William Yao, chief technology officer of Ant Digital Technologies, brought the issue back to the dual nature of AI as both productivity engine and risk multiplier. "The AI is the top line, and also the bottom line," Yao argued. "The top line is about efficiency. The bottom line is about security and risk." His point was that banks cannot separate the gains from the controls because the same technology that improves speed and productivity also expands the threat surface.
Yao gave two figures that made the risk tangible for bankers. Deepfakes now account for more than 5% of new know your customer (KYC) account applications in Ant Digital Technologies’ experience, he said. Large language models (LLMs) also now generate nearly 90% of the code written by Ant Digital Technologies’ engineering teams, changing what engineers do, what they must review and what skills the role now requires.
Those examples show why AI maturity must determine the control model. Yao described a framework that moves from AI as a query tool, to co-pilot, to system designer, and finally to AI-native operations in which AI carries out standard procedures and code generation without human initiation. A wrong chatbot answer may damage service quality, but a system that misreads a specification, generates faulty code or acts on an operating procedure without oversight can create production, security and compliance failures.
Deepfake applications also show that AI does not only live inside the bank, but also arrives from outside through fraud attempts that attack onboarding and identity controls. That makes maturity assessment a practical way for banks to connect AI capability with business criticality, customer harm, cyber exposure and regulatory accountability. The greater the autonomy, the more explicit the control model must become.

Human judgement remains the control that no tool has replaced
Banks still lack a single tool that can wrap deterministic safeguards around probabilistic AI. In practical terms, that means institutions need ways to ensure that AI systems produce outputs that remain consistent, explainable and within defined limits even as models change or encounter new data. Streitberg said no single tool has solved that problem for Hong Leong Bank.
"I don't think there's a tool per se. If there is, I'd be happy to know what it is. But I think it's a process," Streitberg acknowledged. He described a model in which two AI systems trained on the same or similar data sets produce outputs that can be compared, with a third model checking whether the first two diverge. When the outputs differ, the matter passes to a specialist who can judge the business and risk implications before the bank allows the action to proceed.
That gap leaves banks with a human capability constraint as much as a technology constraint. Lim’s warning about skills applies directly to this oversight problem because institutions need people who understand the business process, the data lineage, the model output and the risk consequence well enough to intervene. Human-in-the-loop control only works when the human has the expertise, authority and time to challenge the system before automated actions become customer, financial, operational or regulatory incidents.

The next divide will be between banks that can govern AI and those that cannot
Institutions that move from pilots to scale will need to treat AI readiness as a board-level operating discipline that links strategy, autonomy, data lineage, infrastructure resilience, identity controls and accountability into one governance model. That shift will require banks to change the order in which they pursue AI, moving foundations ahead of autonomy rather than attempting to retrofit controls after deployment. It will also require them to invest as much in the people who can challenge AI decisions as in the systems that generate them.
The next divide in banking will not be between institutions that use AI and those that do not. It will be between banks that can govern AI as part of the institution and those whose ability to oversee it is overtaken by the autonomy they give it.