Ant International Pushes AI Agents Into Payments and Treasury

You have probably watched a colleague wrestle with a foreign-exchange approval, chase down a payment status or reconcile treasury accounts across three time zones. The work is real, the stakes are high and the tools have been static for years. Ant International, the Singapore-based fintech arm of Ant Group, has now released a technology stack that hands those tasks to AI agents. The platform is designed to let software reason through payments, accounts, foreign exchange and treasury operations without a human in the loop for every decision.

Peng Yang, CEO of Ant International, introduced the payment foundation model at the company's global merchant event. The announcement signals a bet that agentic AI – systems that plan, execute and adapt across multi-step workflows – can move from experimental chatbots into the operational core of finance. Where earlier automation required rigid rules and brittle integrations, this stack is built to interpret context, negotiate exceptions and act on behalf of treasury teams.

What agentic AI means for treasury

Agentic AI differs from the conversational assistants that summarise emails or draft replies. An agent is given a goal, a set of tools and permission to decide how to reach the outcome. In treasury, that might mean checking balances, comparing exchange rates, initiating a transfer, logging the transaction and notifying the approver – all as a single delegated task. The agent does not wait for a human to click through five screens; it completes the workflow and reports back.

Ant International's stack provides the infrastructure for agents to interact with payment rails, account systems and foreign-exchange markets. The foundation model understands financial primitives: currencies, counterparties, settlement windows, regulatory flags. It can parse a request like "settle the supplier invoice in euros by end of day" and translate that into API calls, compliance checks and execution steps. The platform also exposes hooks for developers to add custom logic, so a corporate treasury team can encode its own approval thresholds or hedging rules.

This architecture matters because treasury operations are fragmented. A single cross-border payment might touch a bank API, a compliance screening service, a foreign-exchange broker and an enterprise resource planning system. Humans have been the glue, copying data between screens and translating business intent into system commands. Agentic AI can hold that context in memory, call the right endpoints in sequence and handle errors without escalating every edge case.

The operational case

The immediate use cases are tactical. An agent can monitor cash positions across subsidiaries and trigger sweeps when balances fall below a threshold. It can compare spot rates from multiple providers, execute a hedge and record the transaction in the general ledger. It can answer a CFO's question about liquidity in real time by querying live data rather than waiting for a weekly report. These are not speculative scenarios; they are the daily work of treasury teams that currently require manual coordination.

Ant International's platform also addresses a constraint that has limited earlier automation: the need for structured, predictable inputs. Foundation models trained on financial language can interpret ambiguous instructions, ask clarifying questions and adapt to incomplete information. A treasurer might say "pay the Milan office what we owe them" without specifying an amount or currency. The agent can look up the outstanding invoice, confirm the amount, check the euro balance and propose a payment – then wait for approval or execute autonomously, depending on the delegation rules.

The risk is that agents will act on flawed assumptions or outdated data. Ant International has not published details on how the platform handles errors, but the design implies a need for audit trails, rollback mechanisms and human override. Treasury is a domain where a misplaced decimal or a missed compliance check can trigger regulatory penalties or liquidity crises. Agentic systems must be transparent enough that a human can reconstruct the reasoning and intervene before a mistake compounds.

What this means for L&D

For learning and development leaders, the shift to agentic AI in finance creates a new capability gap. Treasury professionals will need to understand how to delegate tasks to agents, set guardrails and interpret the decisions those agents make. That is not a prompt-engineering skill; it is a blend of domain expertise, systems thinking and risk management. Training programmes that focus only on using AI as a co-pilot will miss the operational reality of agents that act independently.

The broader implication is that work is being redesigned around what agents can own. In treasury, that might mean humans focus on strategy, relationship management and exception handling, while agents handle execution, monitoring and reporting. In other functions – procurement, compliance, customer service – the division of labour will look different, but the pattern is the same. L&D teams will need to help employees identify which decisions to keep, which to delegate and how to govern the boundary.

Ant International's platform also raises a question about skill durability. If an agent can execute a foreign-exchange hedge, does a junior treasurer still need to learn the mechanics of forward contracts and settlement risk? The answer is yes, but the learning path changes. Understanding how a process works becomes the foundation for supervising an agent that runs the process. The skill is not obsolete; it is repositioned as the basis for oversight rather than execution.

Organisations that wait for agentic AI to mature before investing in capability building will find themselves behind. The technology is already in production at Ant International, and other financial institutions are building similar stacks. The competitive advantage will go to teams that can deploy agents safely, scale them across workflows and adapt as the technology improves. That requires a learning culture that treats AI literacy not as a one-off training event but as a continuous practice of experimentation, feedback and refinement.

Sources:

• finews.asia, 18 September 2026