AI agents expose the readiness gap in financial institutions

Written by Russell Pereira

May 20, 2026
9 min

Anouska Ladds of Mastercard and Johnson Poh of Singapore’s Infocomm Media Development Authority argue that AI agents will scale only when financial institutions can govern, trace and supervise their actions inside live operations.

Artificial intelligence (AI) agents are entering financial services at the point where accountability, trust and operational resilience become harder to manage than model performance alone. Banks, payment networks and enterprises are testing whether agents can execute tasks inside live workflows. Across Asia Pacific, that shift is emerging in payments, cybersecurity and enterprise workflows, where productivity gains must come with stronger controls.

Anouska Ladds, executive vice president of commercial and new payment flows for Asia Pacific at Mastercard, sees active agentic AI experimentation in Asia’s business payment workflows, where companies procure, approve, reconcile and pay. Johnson Poh, assistant chief executive of Singapore's Infocomm Media Development Authority (IMDA), works with small and medium-sized enterprises (SMEs) and larger organisations on the digital, data and AI foundations that determine whether those deployments can scale. The challenge sits in the systems, controls, data and people that determine whether an agent can act safely, be supervised properly and be held accountable.

Agents change how institutions make, supervise and audit decisions, especially when they move from contained pilots into workflows that affect customers, counterparties and regulated transactions. That shift raises the threshold for safe deployment because financial institutions must prove that agentic systems can act within defined authority, leave a traceable record and support clear accountability when something goes wrong.

Bounded autonomy shows where agents can work

Mastercard’s long-running use of AI in fraud prevention, predictive modelling and credit risk scoring shows where AI has already scaled in financial services: high-volume, high-stakes environments where decisions are data-rich, rule-governed and tightly controlled. That experience fits Poh’s broader idea of "bounded autonomy", where AI operates within defined guardrails, connects to enterprise systems and keeps humans involved at critical checkpoints.

The clearest production use cases remain repetitive, information-dense and process-driven, including software development, cybersecurity operations and enterprise workflow automation. Agents work best where tasks are well understood, systems are connected and humans supervise critical checkpoints. Banking and financial services face a higher control threshold because they operate in a regulated environment, where productivity gains must be matched by clear accountability and governed execution.

"The winners are not the ones chasing futuristic demos," Poh argued. "The winners moving forward will be the ones who are solving real operational bottlenecks with measurable return on investments."

Fragmented workflows create the real value

Ladds identified delegated authority as the next layer of agent deployment, where systems do not merely assess risk, flag anomalies or recommend decisions, but act within approved parameters. In commercial payments, this shift matters because businesses often manage procurement, supplier selection, approval, reconciliation and payment through fragmented systems that still rely on repeated human intervention. The value lies in giving agents controlled responsibility for the sequence, rather than using them to automate isolated tasks.

Across Asia Pacific, that fragmentation is visible in business payment workflows that sit at the intersection of payments infrastructure, ecosystem partners and the processes companies use every day. A company may receive a supplier invoice in one system, reconcile it in another and make the payment through a separate process, while staff still need to verify whether the supplier is compliant. "It's less about AI acting on one single task," Ladds noted, "but about connecting the fragmented workflows and ecosystems that happen today. That is where the real value can occur."

The practical stakes are especially clear for smaller businesses. SMEs often lack dedicated finance, security or operations functions, even though they face many of the same administrative demands as larger corporates. Agents operating within approved parameters could carry part of that load, while Mastercard's first agentic transactions in Australia and New Zealand offered an early example of controlled experimentation moving into commercial practice.

Control gets harder beyond the institution

Internal deployments are easier to control because they operate inside a known environment. The control problem becomes structurally harder when agents act across institutions, counterparties and payment networks, especially when authority is defined in advance and execution happens later in an environment the deploying organisation does not fully control. Agentic AI then becomes an ecosystem trust issue, not merely an internal efficiency tool.

Traditional commerce makes intent visible at the point of transaction. A card tap or checkout click gives the bank, merchant, network and customer a shared signal that authorisation has occurred. Agentic execution changes that model because a customer or business may instruct an agent to purchase a product when it returns to stock, use a virtual card and stay within a defined price limit, while no human is present at the moment of execution.

The authority must therefore be codified upfront and must be recognisable and verifiable to every participant in the network, not only the institution that set it. "If you can't make the agent's role visible and auditable end-to-end, confidence erodes," Ladds observed. "It's not enough for the agent to be well-governed internally. The ecosystem overall needs to be able to recognise it and act in real time."
For regulated activities such as lending, underwriting and claims decisions, institutions will still need human approval because regulators expect them to remain fully accountable for AI-driven outcomes. That accountability must extend from the original instruction and authority limits to execution, exception handling and dispute resolution.

Governance must be built into the system

Regulated industries must govern AI like any other critical risk system, with strong oversight, continuous monitoring and clearly defined accountability. Institutions need model validation for bias and accuracy, audit trails, explainability and traceability before they allow agents to influence high-risk decisions. Customer-facing deployments carry higher trust, compliance and reputational risk than internal productivity use cases, which makes embedded governance essential.

In agentic payments, those controls must operate across the ecosystem. Once an agent acts on behalf of a business, every participant needs to know who authorised the action, what limits applied, whether the instruction was valid and who remains accountable if something goes wrong. Privacy, protection and fraud controls also need to travel with the transaction because partners on the other end of the flow need the same assurance as the institution that deployed the agent.

The control challenge is not solved by retrofitting a human-in-the-loop into a system not designed for it. Poh named four risks that organisations routinely underestimate: accountability, meaningful human oversight, escalation controls and auditability. Human oversight loses value when staff approve outputs without understanding the context, assumptions, data quality or consequences of the decision.

"Governance has to become part of the architecture in itself," he said. "The identity, the permissions, the audit trails, the context boundaries all need to be codified into code and embedded naturally into a system environment." Poh argued that leadership fluency and cross-functional ownership are often missing because AI cannot sit only with innovation, technology, data or information technology teams. Operations, human resources, legal, risk, cybersecurity and business units all need to move in step

Data foundations carry the trust burden

The quality of AI output depends on the quality of the data environment beneath it. "AI is anchored upon the maturity of digital as well as data," Poh noted. Auditability, explainability and accountable execution all depend on clean, structured, accessible and well-governed data, while poorly designed processes cannot be repaired by placing agents on top of them.

The work covers data operations, storage architecture, compute management and governance frameworks. It is unglamorous and time-consuming, and it does not diminish as AI advances. "The barrier to developing quality systems and applications will be a lot lower," Poh observed, "but the requirements for data will remain." Pilots and sandboxes help institutions de-risk early implementation, but their value lies in preparing agents for governed deployment into live systems.

Reliable data gives institutions the basis for oversight, but people still need the capability and authority to act on what the systems reveal.

Workforce readiness becomes a control issue

Workforce readiness is often treated as an afterthought, even though it determines whether human oversight has real meaning. Poh warned of a desynchronisation between the speed at which organisations deploy AI and the speed at which employees develop the capability to operate in the new environment. That gap reflects whether job scopes, decision rights and escalation responsibilities have been redesigned around agent-enabled workflows, not merely whether staff have received training.

Organisations that deploy agents at scale will need staff who can exercise meaningful oversight, challenge system outputs, identify exceptions and escalate decisions appropriately. A human-in-the-loop model will not satisfy regulators, customers or internal risk teams if the human lacks the knowledge, authority or context to intervene. The workforce therefore becomes part of the control architecture, not a separate change management item.

"Speed is valuable, but not at the expense of trust," Ladds observed. That warning applies as much to workforce readiness as it does to payments infrastructure because institutions cannot move faster than the people expected to supervise the systems.

Scale depends on governed execution

Ladds described success as agents that file and verify expense reports without prompting, pay suppliers when defined criteria are met and complete procurement sequences without manual handoffs. Agents act inside approved parameters, use existing tools, connect fragmented workflows and leave a visible record that every relevant party can understand. That ordinary execution will only hold if the organisation beneath it is ready, with reliable data, embedded governance and employees capable of supervising exceptions rather than rubber-stamping outputs.

Scale will take time because legacy infrastructure remains complex and agents need both ecosystem trust and institutional readiness before they can operate safely. Payment networks, banks, suppliers and businesses need visible authority, traceable execution, stronger data environments, embedded governance and staff capable of exercising real oversight. The emerging competitive question across Asia Pacific is less about which institutions have the most capable agents, and more about which have done the organisational work to make governed execution routine inside financial infrastructure.