The next phase of AI in payments may have less to do with automating payment initiation than allowing agents to complete the business activities that require payment. That shifts the challenge from execution to measurable outcomes, controls, identity and liability.

Much of the discussion around agentic payments begins with the transaction: an artificial intelligence (AI) agent identifies a payment requirement, chooses how to pay and executes it without waiting for a person to initiate every step.

The more consequential proposition is what happens before the payment. If an agent can perform the underlying procurement, treasury, marketing or operational task, payment becomes one action inside an automated business process rather than the activity being automated.

That distinction ran through the Sibos 2026 session, “Hands-off, always on: unlocking the opportunity in agentic payments”, which brought together representatives from Ant International, Goldman Sachs, Standard Chartered and NPCI International Payments.

Jiangming Yang, Chief Innovation Officer of Ant International, expressed the proposition most directly. “I believe agentic is not just the automation of payments. It's more the automation of the task,” he said.

From automating transactions to automating work

Yang used the familiar example of booking a ride. The passenger chooses the journey and type of vehicle; payment happens within the service without becoming a separate activity.

Agentic commerce could extend that model into business operations. Yang said the payments industry has spent the past decade or two digitising the payment itself. The next step is potentially to automate marketing, procurement and other tasks involved in running a business, with payment embedded within them.

That changes how the economic value of agentic payments should be assessed. The benefit is not that an AI system can initiate a payment more quickly than a person. It is whether the agent can complete the underlying task more effectively.

Yang distinguished between an agent becoming a new sales force and becoming part of a company's workforce. He argued that the second opportunity could emerge sooner because many business-to-business (B2B) operational tasks have clear, measurable objectives.

In marketing, for example, performance can be measured through customer-acquisition costs. In procurement, it can be assessed against inventory, price and delivery time. If the outcome can be defined, organisations can determine whether the agent actually creates value.

Treasury provides an immediate test

Treasury offers another example because its inputs and outcomes are relatively quantifiable. Yang said Ant International moves more than $1 trillion cross-border annually and uses AI modelling to predict treasury demand across entities, accounts and currencies. He said the approach has reduced the company's working capital requirement by more than 60%, and that clients are adopting the same solution.

In digital marketing, he said a human team might test 10 or 20 combinations of audiences, campaigns, images and discounts in a week, while an agent could test 100 or 200 and potentially improve efficiency by 20% or 30%. Other panellists identified applications already emerging across the transaction-banking value chain.

Vanessa Lin, Global Head of Product for Transaction Banking at Goldman Sachs, pointed to bank-statement reconciliation, where an agent could compare statements with a company's general ledger and identify discrepancies.

She also described contextual fraud detection. Instead of relying only on static rules, an agent could consider several factors together: for example, an unusually large payment, a recently changed beneficiary account and an unfamiliar internet protocol address.

Banking-fee analysis was another potential application. These examples begin with an existing operational problem, not with a search for somewhere to deploy an AI agent.

Autonomy creates a new control problem

The more authority an agent receives, however, the more important it becomes to determine exactly what that authority means.

Lin said existing bank controls are largely built around human intent, static permissions and predefined rules. Agentic payments introduce another actor into that framework.

Banks may need to know not only their customer but also the agent acting for that customer: what permissions it has, what authority has been delegated to it and what controls govern its behaviour.

If an agent independently makes a $10 million payment, Lin said, the institution needs to understand why the decision was made, what controls were applied and how the sequence can be reconstructed.

That makes explainability and auditability operational requirements. Lin argued that identity, authorisation and liability would have to be addressed across an ecosystem that includes banks, treasury management systems, enterprise resource planning platforms and fintech companies.

She also questioned whether payments would ever become completely autonomous. Some transactions may continue to require a human in the loop, particularly when circumstances fall outside the boundaries established for the agent.

Scaling depends on data, governance and people

Sunday Domingo, Global Head of Digital and Data Products and Solutions in Standard Chartered's Transaction Services business, similarly argued that scaling agentic systems is not solely a technology problem.

Banks need accountable owners for AI systems and employees capable of evaluating their output. The quality of the underlying data is equally important.

Domingo pointed to the industry's migration to ISO 20022 as part of that foundation, arguing that organisations should assess whether their people, processes and data are ready before asking whether their agents are ready.

Ritesh Shukla, Chief Executive Officer of NPCI International Payments, put the issue at an ecosystem level. Autonomous payments require trust around consent, control, identity and authentication, alongside sophisticated fraud and risk management.

Technology therefore has to develop together with standards, governance and interoperability. An agentic transaction may involve banks, fintech companies, payment networks, merchants, corporate systems and customers. Automation can only extend as far as the trust framework surrounding it.

Banks may eventually have to design for agents

If agents become genuine participants in financial workflows, financial products themselves may also need to change.

Lin said transaction-banking products are currently designed around human interaction: screens, dashboards, clicks and visual interfaces. Agents do not need the same experience. They interact through application programming interfaces, structured data and rapid machine-to-machine exchanges.

She suggested that banks may eventually need to think not only about the client journey and user experience but about the “agent experience” and “agent journey”.

That is materially different from putting a conversational interface over an existing banking application. The underlying financial product would have to expose information and actions in forms that software can interpret while retaining the controls needed to keep the agent within its authority.

Human accountability will remain important, but some interaction with banks and financial infrastructure could increasingly be delegated to software.

The task becomes the unit of automation

Yang's longer-term proposition is that the familiar payment moment may eventually disappear.

Today, even digital payments are generally understood as identifiable events: someone confirms a transaction and money moves. If business tasks themselves become autonomous, Yang argued, attention will move to the end-to-end outcome.

“What you will care about is the end-to-end task,” he said. The payment remains necessary, but it becomes infrastructure within the process.

That future is still constrained by unresolved questions around identity, permissions, liability, interoperability and human oversight. The Sibos discussion did not suggest that fully autonomous payments are about to replace conventional corporate payment processes.

It did, however, provide a practical way of deciding where to begin. Lin suggested looking at an organisation's most difficult process to see whether AI could help solve it. Domingo said banks should start with the client's pain point, not with the agent. Yang recommended looking for operational tasks with clear, measurable outcomes.

The starting question for agentic payments is therefore not which payments can be handed to an AI agent. It is which business tasks can be automated safely and measurably, and what payment capability the agent will need to complete them.