Banking is at an inflection point. AI has moved beyond experimentation to become a core enabler of customer engagement, operational efficiency and risk management, but the future will not be determined by technology alone. It will be determined by how effectively intelligence is embedded across processes, systems and people. The Heads of Technology Roundtable at The Asian Banker Summit 2026 convened senior leaders from HSBC, OCBC, State Bank of India, Maybank, Hong Leong Bank, UOB, Bank Rakyat, GX Bank, AEON Bank, Alliance Bank, Ambank, China Construction Bank and other financial institutions, alongside technology partners Temenos, Red Hat and Systems Limited. The discussion raised several key questions: How can institutions embed AI across processes and decision-making? How should strategies around core modernisation, data governance, and workforce enablement be sequenced to scale? And what practical challenges exist in translating AI pilots into measurable business impact? Participants expressed a mix of ambition and caution. Legacy constraints, fragmented data estates and uneven workforce readiness continue to shape the pace and scope of implementation, and the discussion surfaced both consensus and debate on how to sequence the work ahead. Governance design, not technology choice, is now the focus of AI strategy Banks agree AI must be enterprise-wide, but disagree on how to get there. The governance question of centralised control versus distributed ownership emerged as a primary fault line shaping implementation. Several participants preferred centralised AI hubs to maintain control, consistency and governance, while others advocated a hub-and-spoke model that gives business units ownership to ensure agility aligned with their operational priorities. Some warned that decentralised adoption could create inconsistent outcomes; others argued that centralisation slows innovation. A recurring theme was the challenge of sequencing: balancing quick wins in low-risk pilots with the longer-term effort of embedding AI into core banking operations and customer journeys. Piloting AI in internal operations is often faster, while customer-facing applications require tighter governance. One participant framed the investment logic as integrating AI with digital platforms and enterprise data, creating what they termed an “AI, Digital, Data” or “ADD” approach, with priorities spanning AI infrastructure, workforce upskilling and embedding AI directly into workflows rather than isolated applications. The group converged on the need to link AI deployment to key performance indicators (KPIs), such as efficiency gains, revenue generation or customer experience improvements, rather than treating it solely as a technology initiative. “AI is most effective when it is operationalised across the enterprise,” said Frankie Wai, business solution director, Temenos. Treating it as a project limits its impact. Every business unit needs to see AI as part of their workflows, and governance must be integrated from day one. Banks are modernising around their cores to support AI use cases Legacy architecture emerged as another strategic constraint. Banks cannot embed AI at the process level without first deciding whether to modernise around their existing core or replace it, and that choice is reshaping implementation timelines across the region. Batch-processing architectures, tightly coupled systems and limited real-time capabilities were described as barriers to embedding AI into daily operations. However, participants diverged on incremental modularisation, where AI-enabled services are layered on existing systems, and full core modernisation to support real-time decisioning. Participants cited customer onboarding, compliance and risk management as areas where the absence of a modernised core directly limits what AI can deliver. Operational disruption, integration complexity and migration costs were identified as factors shaping adoption timelines. Wai noted that “Decision-making is increasingly event-driven rather than step-by-step. Simply layering AI on top of legacy systems is insufficient, meaning workflows must be redesigned and intelligence embedded at the process level.” His recommendation was to not patch AI onto existing processes, but redesign the process first. The real barrier is decision latency, as sequential processes cannot carry real-time intelligence. Embedding AI across a redesigned process chain delivers far greater impact. He raised that flexible, decoupled architecture is a prerequisite, allowing AI models and services to scale across multiple lines of business while maintaining operational continuity. Banks highlighted that independently deployable business surfaces enable scaling AI on specific operations without affecting the entire core. This approach reduces decision latency and supports iterative, rapid deployment through continuous integration pipelines, complemented by human-in-the-loop oversight. Participants diverged on sequencing, as some viewed cloud modernisation as a prerequisite for AI adoption, while others argued that tying AI to multi-year cloud migrations creates unnecessary delays. The consensus: both initiatives can, and should, run in parallel to accelerate adoption. Data fragmentation is now a compliance problem, not a technical one Banks agree that data quality and governance are prerequisites for AI at scale, but fragmented estates, cross-border regulatory divergence and sovereign AI requirements make data architecture as much a compliance question as a technical one. Without clean, governed and accessible data, participants noted, AI models cannot deliver reliable outcomes. Approaches varied across the room. Multi-market institutions face particular complexity, where compliance, consent and data residency regulations differ across jurisdictions, directly influencing architecture and governance decisions. Hybrid cloud deployments were debated as a way to balance agility against regulatory obligation. Technical and process debt, legacy data silos and missing integration layers remain prerequisites that institutions must address before AI can scale effectively. “Decentralised adoption works best when teams understand the business impact,” said Arvind Swami, head of financial services, Red Hat Asia Pacific. “Champions in each unit accelerate AI uptake while ensuring compliance and operational effectiveness.” People and skills are the binding constraint on AI scale Workforce readiness, not technology, emerged as the binding constraint on AI adoption. Banks cited shortages of AI skillsets, difficulty attracting and retaining talent and the complexity of building capable teams internally, while remaining divided on whether centralised training or business-unit-led experimentation drives faster uptake. Practical approaches included home labs for generative AI (GenAI) experimentation, linking adoption to key performance indicators (KPIs) and recognition programmes, and fostering cross-functional teams of engineers, data scientists and business owners. Several participants preferred centralised centres of excellence; others relied on unit-level AI champions. One institution embedded an AI champion in every business unit, linking adoption to KPIs. A monthly forum allows champions to share successes, spreading initiatives laterally and turning business heads into active owners of the AI agenda. Adoption pace also differs between frontline operations and back-office teams, reflecting different operational priorities and readiness. “Operational efficiency is where the biggest gains are to be made,” said Waseem Yusaf, general manager of Asia Pacific (APAC) at Systems Limited. “But the people challenge is significant. Process and policy owners will become the main drivers of change, with technology serving as an enabler.” The gap between pilot and scale is a question of risk appetite Banks are beginning to show real operational results from AI deployment, but the gap between selective pilots and enterprise-wide scaling reflects divergent risk appetites rather than technology limitations. Several participants shared examples of GenAI and agentic AI workflows deployed across back-office, customer service and risk functions. They also reported reducing process times dramatically, illustrating tangible efficiency gains. Some institutions favoured lightweight deployment across selected business units, while others pursued enterprise-wide rollout with rigorous governance, reflecting differing risk appetites and operational priorities. Balancing efficiency and speed of deployment with governance and oversight emerged as an ongoing operational tension. Participants also highlighted the importance of measurable KPIs for throughput, accuracy and compliance, and recognition programmes that reinforce accountability and adoption as essential to sustaining momentum beyond the pilot stage. Governance and oversight are moving into the design layer Governance is not the brake on AI adoption, but how it is structured, who owns it and when it is embedded determines whether it enables or obstructs scale. Banks discussed top-down versus distributed oversight models, the role of chief AI officers and the need to embed compliance and assurance mechanisms into operational workflows from the outset. For high-volume operations, one participant proposed a multi-model assurance approach: two models assess the same data independently, a third arbitrates divergent outputs, and only unresolved cases reach a human reviewer. Human-in-the-loop thus becomes a targeted exception rather than a universal checkpoint. However, some participants cautioned that fully automating maker-checker risks dissolving accountability when AI monitors AI. The central tension that emerged was speed versus control, as unmonitored deployment risks compromising data integrity or regulatory compliance. Yet, overly centralised governance was also identified as a barrier to experimentation. Chief AI officer roles vary in structure and permanence across organisations, reflecting differences in organisational maturity. Some participants also noted the role is primarily advisory, surfacing AI-related risks to senior management, with the expectation that once embedded, AI governance becomes part of business-as-usual operations. Governance models must adapt to organisational maturity, regulatory exposure and the pace of AI adoption. Cloud strategy, token optimisation and sovereign AI considerations were cited as ongoing challenges requiring careful alignment with regulatory frameworks. Constraints to scaling AI are increasingly organisational The roundtable made clear that enterprise AI adoption is a continuous journey, not a one-off project, which requires intelligence to be embedded across core workflows, decision-making and customer interactions. One participant identified four liabilities to scaling AI: technical debt in legacy systems, data debt where data was designed for regulatory purposes rather than AI, process debt in workflows not built for intelligent decisioning and organisation-wide skill debt. Three of these four lie outside the traditional remit of IT. How banks sequence them will separate the institutions that scale AI from those that stall at the pilot stage. Data consolidation, hybrid cloud deployment and sovereign AI requirements add to the complexity of scaling AI, particularly for multi-market institutions navigating divergent regulatory regimes. Centralised versus decentralised AI teams, sequencing of deployment and balancing speed with oversight remain unresolved across the industry. Practical examples show measurable efficiency gains in targeted deployments, but scaling across units requires careful orchestration of core systems, data, workforce and governance. Building an AI-driven intelligent bank remains a long-term, iterative effort, and institutions are only beginning to close the gap between ambition and execution.
