Agentic finance is outpacing the governance needed to trust it

Written by Russell Pereira

Jun 12, 2026
8 min

Agentic finance is shifting trust from human-approved transactions to autonomous chains of delegated action. Banks now need to prove who authorised an agent, how it behaved, whether it stayed within its mandate and who is liable when it fails.

The agentic model adds a more consequential layer to open finance because autonomous agents do not simply access data or recommend an action. They introduce delegated execution by systems that can pursue goals and act with varying degrees of autonomy, shifting the control question from consent to trust, scope and accountability.

The logic becomes material in institutional finance, where agents could support corporate finance workflows, supply chain financing, approvals and reconciliations across multiple counterparties. Banks and regulators must define how intent is captured, how an agent's mandate is set, how counterparties recognise an agent as legitimate, how actions are logged, how disputes are resolved and who remains liable when an autonomous chain produces a wrong or harmful outcome.

Visibility must follow every handoff

Many institutions approaching agentic artificial intelligence (AI) still treat each agent as a discrete, controllable unit. Camilla Bullock, chief executive officer of Emerging Payments Association Asia argued that this instinct misses the nature of the technology itself. "We have to stop talking about as if it's one agent," Bullock observed. "If we're looking at agentic AI, it's several steps in that." Her concern was precise: any agentic transaction can move through authorisation, discovery, decision and execution, while few organisations command full visibility across all four stages.

That gap becomes harder to manage when third parties supply some agents in the chain. "Very few organisations actually master the transparency across the chain," she noted, "because you might not own all those agents yourself." Banks need visibility at every handoff, from the moment an intent receives authorisation to the moment a transaction settles.

Kenneth Gay, chief fintech officer of the Monetary Authority of Singapore (MAS) reinforced this point from the regulatory side. He described his priority as "observability" rather than transparency alone, because production-scale deployments require regulators and technical teams to see how agents behave in operation. Gay said they need "very detailed logs exactly of how the agents are operating, doing well, not doing well." The distinction matters because transparency explains outcomes, while observability exposes the process that produced them.

AI rules meet autonomous action

Singapore's work on AI governance gives the region a starting point. Tek Yew Chia, adjunct professor and advisor at the Asian Institute of Digital Finance at the National University of Singapore, pointed to the Fairness, Ethics, Accountability and Transparency (FEAT) principles. Those principles remain relevant, but agentic finance stretches them because agents can initiate actions, interact with systems and execute instructions within delegated boundaries.

Mindforge, a Singapore-based industry initiative that helps financial institutions assess and govern generative AI risks, points towards a lifecycle approach to responsibility, from design and development through deployment and operation. That matters because failures may originate in training data, permissions, system design, third-party integration, monitoring or execution, while accountability cannot sit only at the final output if the transaction depends on a chain of agentic decisions.

Humans remain accountable

Responsibility becomes harder to assign once agents act across several owners, developers and operators. Chia stated the principle directly: "The accountability is human. The agent cannot be held accountable." He framed the problem as one of traceable responsibility across the full agent lifecycle. "The key thing is who is held accountable for the agent and how the agent behaves, how the agent is built, and how the agent is executed," Chia argued. "Is it the owner, is it the developer, is it the operator?"

Financial services already operate through layered responsibility. Agentic finance adds another layer that can obscure where judgement, permission or control failed. Liability remains equally unsettled because current payment frameworks rest on intent and individual authorisation. A human clicks, signs, enters a PIN or otherwise confirms a discrete action. Agentic AI disrupts that model because the initial authorisation may happen once, in advance, before cascading through interactions the originator never individually approved.

"Without the liability frameworks there is no trust in the system," Bullock argued. That warning goes to the centre of the issue: trust depends not only on whether agents can act, but on whether institutions can determine who bears responsibility when they act incorrectly.

Know your agent becomes essential

The identity problem brings agentic finance back to familiar banking controls. Banks already understand know your customer, know your business and device-based authentication. Autonomous agents extend that logic into know your agent, where counterparties must verify the agent as part of the control environment.

Chia compared the concept to know your device, where a platform recognises a customer's usual phone, address or behavioural pattern and prompts for verification when those signals do not align. That analogy translates agentic governance into operational banking language: institutions may need to verify whether an instruction came from a recognised agent, whether the agent usually acts for that customer or institution, and whether the action falls within mandate.

Bullock challenged the industry to go further by distinguishing between agents with very different levels of trustworthiness. "Would you like to deal with the agent that I built on the weekend, or would you like to deal with the agent from CBA in Australia or Amazon?" Bullock asked, referring to Commonwealth Bank of Australia (CBA), one of Australia's largest banks. Enterprise-grade agents and consumer-built agents carry fundamentally different risk profiles, yet current governance frameworks do not reliably differentiate between them at the point of transaction.

Safeguards are still provisional

Institutions have begun managing the risks they can measure while liability frameworks remain unresolved. Manish Sharma, head of business development at payments technology company Worldline, described the current state plainly: "What we are doing is a patch job till such time we understand how these agentic commerce will work." Worldline's practical safeguard uses capped virtual cards, including single-use and multi-use instruments with defined credit limits, to contain the downside of autonomous transactions. "Thou shall not lose more than this much," he summarised.

Spend limits and scoped authorisations can reduce losses when an agent operates outside its intended parameters, but they do not amount to a complete governance framework. Chia argued that high-risk financial actions require rules, not open-ended autonomy. "You don't use GenAI on a probabilistic manner to drive agentic commerce," he argued, using the common abbreviation for generative artificial intelligence. "You use it with very clear rules... it's not probabilistic, it's deterministic."

Regulation starts with use cases

Gay described MAS's regulatory approach as risk-proportionate and use-case-driven. Rather than apply the same level of controls to every transaction regardless of size or risk, MAS is looking at the substance of the use case, the underlying risks and the existing control environment in which agentic transactions take place.

MAS favours industry consortia as a mechanism for developing practical governance. The model brings participants together around a specific use case, examines the governance issues that arise and turns those lessons into best practices for the broader market. "We bring sometimes the regulatory colleagues along to observe what is happening, look at what the governance issues that arise, and we also work with industry to come up with best practices," Gay noted.

Standards must precede scale

Chia endorsed compliance by design as the operating standard for any institution building agentic systems today. "You don't build something at the end to say, does this comply?" he noted. "Every agent that you build has to have compliance built in." That principle can strengthen institutional controls, but cross-border agentic finance will need shared standards for agent identity, permission expiry, mandate scope, log retention, auditability, dispute handling and liability allocation.

Asia-Pacific has worked for years to reduce fragmentation in payments through ISO 20022 messaging, interoperable quick response (QR) payments and bilateral payment linkages. Those initiatives provide a useful precedent for the next generation of standards around trusted autonomy. Agentic finance now requires a similar effort around identity, mandates, logs, liability and dispute handling before fragmented architectures become entrenched.

Bullock used her closing remarks to widen the frame beyond any single institution or regulator. "We've been battling fragmentation for the last ten years," she observed, referencing the region's long-running effort to align payment standards across jurisdictions. "Just act while the window is open," Bullock urged.

The deeper risk is not that the region fails to adopt agentic finance, but that adoption consolidates around architectures whose governance gaps remain unresolved. If consolidation happens before common assurance mechanisms emerge, Asia-Pacific could inherit weaknesses in accountability, liability and observability even as interoperability improves.

The next architecture of trust

The way forward is a coordinated assurance architecture that banks, regulators, payment networks, technology providers and market infrastructures can recognise across use cases and borders. It should begin with standards for agent identity, delegated authority, mandate expiry, transaction logs, exception handling and liability allocation, then test them in retail and wholesale workflows.

Asia-Pacific has experience turning fragmented payment markets into more interoperable networks through standards, linkages and regulatory collaboration. Banks that participate early will help define how trusted agents operate, how responsibility is assigned and how cross-border autonomy becomes bankable. Those that wait may find that others have already set the next architecture of financial trust.