AI is expanding the work financial institutions can undertake, from investigating alerts to supporting treasury operations. Capturing its commercial value requires firms to connect that capacity to better services, additional revenue or lower expenditure, while accounting for the cost of delivery. Across AI discussions on the first day of Insights 2040 in Singapore on 16 November, applications ranged from customer-service resolution and software development to financial crime investigations. The commercial opportunities extended to making smaller customers economical to serve and shortening processes that delay revenue. As agents execute actions and interact across organisations, institutions must also decide how much authority to delegate and how to maintain oversight. What becomes economical to deliver? AI changes service economics by expanding operating capacity and lowering the cost of serving customers whose needs previously required substantial human support. In financial crime investigations, staffing constraints influence which alerts receive attention. Agents that gather information and analyse more cases could allow institutions to widen coverage without increasing headcount at the same rate. Greater investigative capacity could improve detection even where expenditure remains unchanged. The value of that capacity depends on reliable execution. Constraining open-ended reasoning and routing critical checks through defined tools can make investigative workflows more consistent. Institutions must also assess those gains against an evolving threat, as criminals use AI for impersonation and fraudulent applications. For smaller businesses, lower delivery costs could make treasury and cross-border financial support more accessible. Providers could offer analysis and operational assistance that would otherwise require customers to maintain specialist finance teams. AI-assisted advice and customer support offer a similar opportunity to make smaller accounts economical to serve, while retaining human support for complex or sensitive cases, including customers in vulnerable circumstances. The cost of that support remains part of service economics. The customer benefit depends on how those services are designed and priced. Lower costs may support cheaper services or broader access, but providers could also retain the savings as higher margins. In investment services, systems rewarded for generating sales or trading activity risk increasing revenue without improving customers’ financial positions. Where do productivity gains become business returns? General-purpose AI tools help employees prepare documents and analyse information, but the larger commercial opportunity lies in changing how work moves through an institution. Faster completion of individual tasks delivers limited value when handoffs and approvals continue to delay the overall process. In financing, faster preparation of documentation becomes more valuable when it helps a transaction move through review and execution sooner, potentially bringing revenue forward. Assessing the return requires examining the whole process, including work passed between relationship teams, operations, legal and compliance. Successful demonstrations can stall when production funding, operational ownership and measures of success remain unresolved. Technology teams may prioritise model performance, business teams seek commercial benefits and operations teams require reliability. Without agreement on what qualifies an application for production, a promising demonstration can remain an experiment. The model may account for only a small part of the work required for production. Connectors, application programming interfaces and changes to surrounding systems and workflows can require substantial investment. Evaluation, maintenance and model usage add continuing expenditure. Upgrades require further work because changes can affect output quality, latency and user experience. Setting requirements early and reusing controls helps reduce repeated assessments and late redesign. Institutions also need to establish where those costs fall. As models and infrastructure are shared across applications, expenditure must be allocated to particular use cases and users, then compared with the benefits generated. Aggregate adoption figures offer limited insight into whether an application earns its running costs. What should institutions build, buy and control? Build-versus-buy decisions begin with identifying which capabilities distinguish an institution’s services and which can be sourced more effectively from providers. Customers experience these choices through the accuracy, timeliness and usefulness of the service. Those outcomes need to guide sourcing alongside development costs. The choice applies at different layers: banks can use external infrastructure and foundation models while developing their own knowledge resources, applications and decision processes. Lower software development costs could make more applications economical to build, increasing competition for technology vendors. Continuing maintenance and upgrades, however, can favour buying capabilities outside an institution’s strengths. Domain expertise, customer relationships and dependable implementation can sustain demand for providers even as institutions gain more options to build. Greater choice could nevertheless put pressure on vendors’ pricing power. Model improvements must justify operating costs and the work required to evaluate and integrate a replacement. Switching to a fallback during an outage also depends on prior testing that it delivers acceptable performance for the application. Across these choices, institutions need to retain control over data access, permitted actions and deployment decisions. They also need to understand which providers and components a service depends on, and whether those dependencies can be managed as requirements change. This requires sufficient assurance about external systems and the internal expertise to challenge providers and assess suitability. The Monetary Authority of Singapore’s (MAS) Guidelines on Artificial Intelligence Risk Management reinforce this responsibility. Institutions remain accountable for AI used in their services, including externally provided systems. Where assurance gaps arise, MAS expects compensating controls; risks that cannot be brought within an institution’s risk appetite may require limiting, suspending or replacing the service. How does agentic AI change decision rights? Agents introduce an organisational change when they execute actions instead of producing information for review. The extent of delegation matters more than the label attached to a system. Each agent needs a named business owner, a defined outcome and clear limits on which decisions it can make and when intervention is required. Authority can expand in stages, from assistance to recommendations and then execution under monitoring, as evidence demonstrates acceptable performance within defined limits. Human accountability, however, provides limited protection if the responsible person lacks the information, authority or time to act. Requiring confirmation at every step undermines automation, making the placement of checkpoints and enforceable restrictions central to the operating model. External agents extend this problem beyond institutions’ own deployment plans. Personal agents could increasingly interact with financial services on customers’ behalf. Providers need to establish how to accommodate them, particularly where agents can operate through interfaces designed for people. Customer agents could also change banking economics by making it easier to compare offers, move balances or optimise applications. Continual rate comparisons and automated transfers could accelerate deposit movements, affecting banks’ funding assumptions. Agents seeking advantageous credit or insurance terms could also challenge assumptions used in underwriting. Accepting those agents requires a receiving institution to establish who they represent and whether a requested action falls within a valid mandate. Different participants validate different elements of a transaction. A bank can verify payment authority and limits without knowing whether the underlying purchase met the customer’s instructions. Valid authorisation does not resolve every question about responsibility. Institutions also need clarity on how duties and losses are shared with providers and counterparties. Existing mandates and payment protections provide a starting point, but their application across an agentic transaction requires clear arrangements for disputes and redress. Practical frameworks are beginning to address these questions. MAS and industry published Safeguards for Agentic Finance at Runtime (SAFR) on 3 July 2026 under BuildFin.ai. The paper proposes governance checkpoints that verify and record agents’ proposed actions before execution, helping keep them within predefined mandates, policies and risk boundaries. Singapore’s Model AI Governance Framework for Agentic AI, launched in January and updated on 20 May 2026, provides complementary guidance on bounding agents’ powers and maintaining meaningful human accountability. The update includes case studies and practices addressing multi-agent systems, third-party agents and automation bias. How do different rules affect deployment? AI adoption takes place within existing financial obligations governing conduct, operational resilience and customer protection. AI-specific measures add requirements or clarify their application, creating different implementation tasks across jurisdictions. MAS issued its AI risk management guidelines on 7 October 2026, covering oversight, inventories, proportionate lifecycle controls and third-party assurance. Institutions should meet Sections 3–4 from 7 October 2027 and Sections 5–6 by 7 October 2028. Existing governance structures may be used where they provide adequate oversight and cross-functional coordination. The UK Financial Conduct Authority’s (FCA) good-practice publication of 19 October 2026 illustrates how existing obligations apply to firms’ AI use, including the Consumer Duty, governance and operational resilience requirements. Its examples do not create new requirements or prescribe a single governance model. The EU AI Act takes a cross-sector legislative approach. Annex III lists specified uses in assessing individuals’ creditworthiness or establishing credit scores, and in life and health insurance risk assessment and pricing, as high-risk. Financial fraud detection is excluded from the creditworthiness category; that exclusion does not itself provide an exemption from the Act as a whole. Following the AI Omnibus amendments, the rules for Annex III high-risk systems apply from 2 December 2027. At the international level, the Financial Stability Board published 12 sound practices on 15 October 2026, covering organisation-wide governance and risk management throughout the AI lifecycle. They provide a voluntary reference for institutions, including on human oversight and third-party dependencies. For cross-border firms, a technology stack may transfer across markets more readily than the service it supports, which remains subject to local licensing requirements, data localisation rules and supervisory expectations. Reusable evidence could reduce repeated work, but institutions still need to meet local obligations and establish the acceptability of each deployment. What can public-private cooperation deliver? Effective AI oversight requires policy expectations that engineers can implement and supervisors can test. Developing those requirements depends on practical experience of applications and failure modes, alongside an understanding of financial consequences. Institutions contribute operational problems, technical teams develop and test controls, and authorities clarify expectations. Supervisors also need enough practical understanding to challenge deployments whose risks are not apparent from general familiarity with AI tools. The Bank of England and FCA’s AI Consortium provides one mechanism for this engagement. It gathers stakeholder input on AI’s capabilities, uses, benefits and risks to inform the authorities’ approach to safe adoption in financial services. Cooperation can also help establish how institutions demonstrate that agents operate within their authorised limits. Common testing methods could help counterparties assess the evidence without repeating every evaluation, reducing the cost of deploying services across organisations. Changes to models, tools or permissions could invalidate earlier assessments and require renewed testing. Shared dependencies require examination beyond individual applications. Institutions using similar models and infrastructure could make correlated decisions or suffer simultaneous disruption despite sound internal controls. Joint testing needs to examine those interactions as well as the reliability of individual agents. Capturing AI’s commercial value requires institutions to redesign delivery, measure its full costs and establish clear limits on delegated authority. Shared assurance could make it easier to extend those arrangements across organisations and deliver services customers value.