Banks are shifting from experimental artificial intelligence (AI) applications towards more integrated agentic workflows, as institutions seek to extend automation beyond employee productivity into core banking processes. The transition is placing greater emphasis on reusable infrastructure, computing costs and the ability to maintain oversight of increasingly autonomous systems. At the Dubai AI Festival, a panel moderated by Akshay Khanna, Managing Partner at Avasant, brought together executives from Mashreq Bank, Zand, First Abu Dhabi Bank Misr (FABMISR), HSBC and Network International to examine the infrastructure and governance requirements shaping AI adoption. The discussion highlighted a shift from demonstrating individual AI applications towards establishing the operational controls and infrastructure needed to deploy increasingly autonomous systems. While institutions are developing the capabilities needed for greater autonomy, the ability to control transactions, preserve business continuity and demonstrate economic value will determine how widely these systems can be deployed. Managing authorisation and control of transactions initiated by AI agents As financial institutions develop AI agents capable of interacting directly with banking services, governance is extending beyond model performance towards establishing who authorises autonomous actions and how those actions are controlled. Haishan Wu, Head of AI at Zand, outlined the digital bank's development of proprietary agentic infrastructure. The bank has already deployed an internal agent platform to support employee productivity, while a second phase focuses on developing infrastructure that would enable AI agents to interact directly with banking services. This includes work on agent identification and transaction protocols. Wu explained that the bank is developing its own infrastructure and standards to support agentic payments, reflecting the need for financial systems capable of recognising and interacting with autonomous agents. However, identifying an agent does not necessarily establish its authority to undertake financial transactions. Banks would need to distinguish between authenticating an autonomous system and determining which activities it is permitted to perform on behalf of a customer. This distinction becomes particularly important when considered alongside the payment infrastructure perspective presented by Görkem Köseoğlu, Group Chief Customer and AI officer at Network International. Köseoğlu emphasised the importance of reliability in payment processing, particularly given the scale and economic significance of transaction infrastructure. “In payments, especially when it touches the core payments, you really need to make sure that you are meeting the reliability requirements first and then you can put AI on top of that,” he said. He described a differentiated approach to agent access, distinguishing between core transaction-processing systems and applications supporting customer interactions, pricing and other services. Köseoğlu explained that agents are subject to different access and read-and-write permissions depending on the systems involved. Together, the approaches described by Zand and Network International expose a critical distinction between an agent's identity and its authority to execute financial transactions. While AI agents may support customer interactions or retrieve information, initiating payments requires additional controls over permissions and transaction execution. The panel also discussed developments in agentic commerce. Network International described payment application programming interfaces (APIs) capable of supporting agentic payment flows across different protocols, alongside work on broader customer journeys covering product discovery and fulfilment. Zand's ongoing development of agentic payment infrastructure and Network International's work on interoperable payment APIs illustrate how financial institutions are preparing for more autonomous financial transactions. Köseoğlu noted that much of the demand observed by Network International remains concentrated in closed-loop environments, highlighting the importance of interoperability across payment systems while maintaining transaction reliability and security. As agentic commerce develops, banks and payment providers will need to establish clear frameworks for authorising AI-initiated transactions, managing system access and maintaining accountability across interconnected platforms. The challenge will be to support greater autonomy without compromising the controls underpinning financial transactions. Retaining critical banking expertise when AI systems fail As AI assumes greater responsibility for banking processes, financial institutions face a growing operational resilience challenge: preserving the human expertise required to maintain critical services when automated systems fail. The increasing delegation of complex tasks to AI raises questions about whether banks will retain the institutional knowledge and operational capabilities needed to function independently of these systems. Shaarawy Mohamed, Chief Technology and Information Officer at FABMISR, raised the possibility that banks could become increasingly dependent on AI systems as complex operational processes are automated. Considering a hypothetical future in which large parts of banking operations are managed autonomously by AI agents, Mohamed questioned whether institutions would retain the expertise needed to continue operating if those systems became unavailable. “Do we have the institutional knowledge? Do we have the people who are capable and trained to know how to operate a bank when something wrong is happening?” he asked. Mohamed suggested that conventional disaster recovery arrangements may eventually need to be complemented by human recovery testing. Traditional disaster recovery focuses on restoring systems and infrastructure following disruption, while human recovery testing would examine whether employees remain capable of performing essential banking activities without AI assistance. As automated systems take on increasingly complex tasks, institutions could face a situation where technology can be restored but employees no longer possess sufficient familiarity with the underlying processes. Maintaining operational continuity would therefore depend not only on system recovery but also on preserving institutional knowledge. Mohamed also discussed mechanisms for isolating individual agents or processes during disruptions rather than shutting down an entire AI-enabled environment. Yan Bechet, Chief Operating Officer UAE and Interim Head of International Wealth and Premier Banking UAE at HSBC, reinforced the importance of adapting established risk controls to the growing scale and impact of AI. “AI doesn’t bring new types of risks. The same typology of risk that existed before exists today. The scale, the impact is different,” he said. Bechet emphasised the need to adapt control environments, including how automated decisions are explained, recorded and monitored. He also highlighted the importance of retaining the ability to switch off AI models and continue processing outside an AI environment when necessary. These perspectives suggest that operational resilience will increasingly depend on the ability to control automated processes, isolate failures and maintain the human capabilities needed to sustain critical services. Human recovery testing remains a proposed extension of existing business continuity arrangements rather than an established practice demonstrated by the panel. Nevertheless, it raises an important consideration for banks as they expand automation: whether the operational efficiencies gained through AI can be sustained without weakening their ability to function independently of it. Scaling agentic AI without proportionate increases in costs Beyond transaction controls and operational resilience, scaling agentic AI presents a further challenge for financial institutions: expanding automation across banking functions while managing the associated infrastructure and operating costs. As banks move beyond individual applications, reusable technology components and more efficient use of computing resources are becoming increasingly important to sustainable deployment. Xi Liang, chief AI officer at Mashreq Bank, described how the bank progressed from foundational applications, including knowledge extraction and employee productivity, towards a federated operating model combining centralised AI strategy, governance and infrastructure with implementation across business units. The approach focuses on developing reusable technology components that can support multiple applications while limiting duplication of development and computing resources. Describing the bank's objective, Liang said it aims to “build faster, reuse as much as we can, optimize the token across the different use cases.” She explained that Mashreq is working towards a model in which the benefits generated by additional AI applications can grow without proportionate increases in costs. This reflects a shift from delivering isolated productivity improvements towards building shared infrastructure capable of supporting wider deployment more economically. Mashreq is also extending AI into document-intensive processes, including trade finance, where intelligent document extraction and processing can reduce manual work. However, human intervention remains necessary when outputs fall below defined confidence thresholds, with exceptions and edge cases continuing to inform system improvement. The experience illustrates that scaling AI requires more than introducing additional applications. Banks must determine which capabilities can be reused across functions, where human oversight remains necessary and whether automation delivers sustainable operational benefits. Infrastructure choices also reflect regulatory and operational constraints. Mohamed explained that Egyptian regulatory restrictions on cloud use for banking data have encouraged FABMISR to develop on-premises AI capabilities. This has required careful selection of graphics processing units (GPUs), language models and use cases, alongside greater attention to computing costs and performance testing. This has reinforced the importance of resource optimisation and selective implementation, rather than applying generative AI indiscriminately across banking processes. Scaling agentic AI requires banks to align infrastructure investment with operational priorities while making efficient use of available computing resources. Mashreq's focus on reusable technology components and FABMISR's selective approach to AI deployment reflect different strategies for managing these requirements. As agentic AI becomes more integrated into banking operations, the challenge will be to translate wider deployment into measurable efficiency gains without proportionate increases in costs. This places greater emphasis on the operational value of AI applications rather than the number of use cases introduced. Operating-model readiness will shape AI adoption The discussion highlighted a transition in banking AI strategies from demonstrating individual applications towards establishing the infrastructure, governance and operational capabilities required for wider deployment. Zand's work on agentic payment infrastructure and Network International's differentiated access controls underline the importance of defining the authority granted to autonomous agents before extending their role into financial transactions. The discussion around human recovery testing highlighted a broader challenge for financial institutions: preserving operational resilience as AI assumes greater responsibility for critical banking processes. HSBC's Yan Bechet reinforced the importance of adapting existing risk management frameworks to the growing scale and impact of AI, ensuring that institutions maintain effective oversight, accountability and business continuity as automation expands. Meanwhile, Mashreq's emphasis on reusable infrastructure demonstrates that the economics of scaling will remain important as AI becomes more integrated into banking operations. For financial institutions, the next stage of adoption will depend on whether agentic AI can deliver improvements in processing efficiency and customer experience while maintaining transaction integrity, operational continuity and accountability for autonomous actions. Competitive differentiation is therefore likely to emerge not from the number of AI applications deployed, but from institutions' ability to integrate autonomous capabilities into scalable, cost-efficient and well-governed banking operations.
