Agentic payments, where AI agents initiate and manage transactions on behalf of a business, raise a question every finance and treasury team will need to answer: what does it take to trust AI with payments?
Payments have evolved from cash and manual transfers to digital transactions that enable increasingly faster cross-border transfers. Businesses now operate across different payment methods, currencies and financial networks. As payment infrastructure continues to evolve and AI agents become capable of executing payments, a new question emerges: who, or what initiates and approves these transactions?
The agentic finance era introduces a new possibility: systems that can interpret business objectives, make decisions and initiate payments on behalf of an organisation. Imagine an agent that identifies an upcoming supplier payment, checks it against the business’s requirements and initiates the transaction with limited human intervention.
The potential is significant, from reducing manual work to coordinating financial operations more efficiently. But allowing an AI agent to move money also raises harder questions: can it make the right decisions, within the right authority and with the right safeguards?
That trust goes beyond the ability to execute a payment. It requires reliable payment infrastructure, governance that operates alongside execution, and integration into real business workflows.
What changes when AI starts moving money?
Traditional payment processes are typically initiated by people or automated systems following predefined instructions and rules. AI agents introduce a different approach: they can interpret business objectives, assess available information and determine when and how to initiate payments within a defined scope of authority.
This changes where payment decisions are made and how they need to be controlled.
Early research points to what agents may be capable of. A BIS Working Paper on AI-assisted cash management in wholesale payment systems (Aldasoro and Desai, 2025) tested an AI agent in simulated scenarios. The agent prioritised urgent payments, maintained precautionary liquidity buffers and balanced settlement delays against liquidity costs. These findings suggest possible operational benefits, although they do not establish how agentic payments will perform across live business environments.
The International Monetary Fund (IMF), in its April 2026 note How Agentic AI Will Reshape Payments, highlights a key technical challenge: probabilistic decision-making by AI systems must be reconciled with the precise, rule-based requirements of payment infrastructure. The note also identifies risks including traceability, opacity, cybersecurity and legal uncertainty. These are not reasons to rule out agentic payments, but they underline why greater autonomy must be accompanied by appropriate safeguards.
For businesses, the question is not simply whether payment decisions can be delegated. It is whether the wider payment environment can support that execution reliably, within clearly defined boundaries and with appropriate accountability.
1. Agentic Rails: Can the money move where it needs to go?
An AI agent may identify the right payment, but that decision creates value only if the payment can be executed reliably.
Businesses operate across different payment methods, currencies, jurisdictions and financial service providers. Agentic payments must work within this environment, connecting an agent’s decision to the infrastructure that processes and settles the transaction.
This involves more than speed. Payment systems need to support reliable execution, appropriate controls and clear outcomes. Where transactions cross different systems or jurisdictions, interoperability and settlement arrangements also matter.
The IMF note suggests one way to frame this: keep an agent’s probabilistic reasoning upstream, and keep authorisation and settlement deterministic, so that payment finality and control are preserved even when an agent initiates the transaction. In other words, the agent may decide what to pay, but the payment infrastructure must establish whether the transaction is permitted and ensure that it is processed and completed reliably. This distinction allows AI agents to make decisions while ensuring that the actual movement of money remains controlled and traceable.
2. Built-in Rules: How do we govern what AI agents are allowed to do?
If an AI agent can act on behalf of a business, its identity and authority become essential parts of the payment process.
Businesses need to know which agent is acting, whom it represents and what it is permitted to do. Its mandate might specify which suppliers it can pay, the transactions it can initiate and the limits within which it can operate.
Authorisation, however, does not guarantee that every decision an agent makes is correct.
What happens when an authorised agent makes the wrong payment decision?
Consider a business that allows an AI agent to pay approved suppliers within set limits.
The process might look like this:
- The agent receives payment instructions. It identifies a payment that needs to be made and gathers the relevant information.
- The agent makes a mistake. It misinterprets conflicting instructions and selects a bank account belonging to a different supplier.
- The payment reaches the final checks. Can the system identify the mismatch before the money is transferred? Would the payment be flagged for human review?
- The business needs to respond. If a problem is detected, can someone understand why the agent made that decision and stop the payment in time? If the payment goes through, who is responsible for investigating what happened?
In this scenario, the agent was authorised to make payments, but that did not guarantee that the final decision was correct. These risks highlight the need for governance to operate alongside payment execution, rather than relying solely on checks before an agent is deployed or reviews after a transaction has taken place.
Regulators and industry are already working on this problem. In July 2026, the Monetary Authority of Singapore (MAS) published Safeguards for Agentic Finance at Runtime (SAFR). SAFR is an industry white paper developed with financial institutions and fintechs under MAS’s BuildFin.ai initiative. It proposes governance checkpoints that verify and record an AI agent’s proposed actions against applicable institutional controls before execution.
In practice, AI payment governance should include clearly defined permissions, transaction limits, policy-based checks, monitoring and escalation procedures. Unusual or higher-risk transactions should be subject to additional risk-based scrutiny before execution, while records of an agent’s decisions can help businesses investigate problems and establish accountability. The appropriate controls depend on the payment system, the risks involved and whether intervention remains possible before a transaction is completed.
The objective is not to require a person to approve every payment. Human agency means people establish the boundaries for autonomous action, retain the ability to intervene when necessary and ensure clear accountability.
3. Real-world Reach: What does it take to scale agentic payments?
A single successful agentic payment is only the starting point. For agentic payments to deliver value at scale, they must connect with the payment infrastructure, networks and business processes that organisations already rely on.
For agentic payments to become part of everyday business operations, they must work across established payment channels, connect with existing financial systems and support different business needs.
As adoption grows, this means handling higher transaction volumes, more counterparties and a wider range of payment scenarios, from supplier payments to treasury operations. Organisations need to consider how agentic capabilities fit into existing workflows, how exceptions are managed and how oversight is maintained as more payment decisions become automated.
Ultimately, scaling agentic payments goes beyond the ability to automate transactions. It is about extending autonomous payment capabilities into real-world financial operations while maintaining the reliability, visibility and safeguards businesses require.
What should businesses assess before trusting AI with payments?
Before delegating payment decisions to AI agents, businesses should consider four questions:
- Infrastructure: Can the payment environment support reliable execution across the relevant systems and payment methods?
- Authority: Is each agent’s identity, mandate and transaction limit clearly defined?
- Oversight: Can unusual or potentially incorrect decisions be detected, reviewed and challenged before execution where the system permits?
- Accountability: Are decision records, escalation paths and responsibilities clear when something goes wrong?
Agentic payments are changing how businesses initiate, coordinate and manage payments.
But the ability to automate transactions is only part of the equation. For businesses to trust AI with payments, trust cannot be added at the end of the process. It must be built into every layer, from how an agent interprets a task to how a payment is authorised, executed and reviewed. This requires clear authority, appropriate safeguards and accountability at every stage.
智能体支付（Agentic Payments）是指由 AI 智能体代表企业发起和管理交易。这一发展也带来了一个所有财务及资金管理团队都必须思考的问题：企业需要具备什么条件，才能放心地将支付交由 AI 处理？
支付方式已从现金和人工转账，逐步发展为支持更快速跨境转账的数字化交易。如今，企业需要在不同的支付方式、货币及金融网络之间开展业务。随着支付基础设施不断演进，AI 智能体也逐渐具备执行支付的能力，一个新的问题随之浮现：究竟由谁，或由什么来发起并批准这些交易？
智能体金融（Agentic Finance）时代带来了新的可能性：系统不仅能够理解业务目标、作出决策，还能代表企业发起支付。举例来说，AI 智能体可以识别即将到期的供应商款项，根据企业的付款要求进行核对，并在有限人工介入的情况下发起交易。
这类技术具有相当大的潜力，不仅能够减少人工操作，也有望让财务运营流程更加高效。然而，当企业允许 AI 智能体调动资金时，也必须面对更深层次的问题：它能否作出正确的决策？是否在获授权的范围内行事？又是否具备适当的安全保障？
因此，信任不能仅建立在执行支付的能力之上，还需要可靠的支付基础设施、贯穿执行过程的治理机制，以及与实际业务流程的有效整合。
当 AI 开始处理支付，会带来哪些改变？
传统支付流程通常由人员发起，或由遵循预设指令和规则的自动化系统执行。AI 智能体则带来了不同的运作模式：它能够理解业务目标、评估可用信息，并在既定授权范围内判断何时以及如何发起支付。
这改变了支付决策的产生方式，也意味着我们必须重新思考如何对这些决策进行管控。
早期研究已初步揭示 AI 智能体的潜在能力。国际清算银行（BIS）一份关于批发支付系统中 AI 辅助现金管理的工作论文（Aldasoro and Desai，2025），通过模拟场景测试了 AI 智能体的表现。研究发现，该智能体能够优先处理紧急付款、维持预防性流动性缓冲，并在结算延迟与流动性成本之间取得平衡。这些结果显示，AI 智能体可能为运营带来效益，但尚不足以证明智能体支付在各类实际业务环境中的表现。
国际货币基金组织（IMF）于2026年4月发布的研究报告《How Agentic AI Will Reshape Payments》则指出一项关键的技术挑战：AI 系统基于概率作出决策的能力，必须与支付基础设施精确、以规则为基础的运作要求相协调。报告也指出了可追溯性不足、运作不透明、网络安全及法律不确定性等风险。
这些挑战并非否定智能体支付发展的理由，而是进一步说明，AI 自主程度越高，越需要配套的安全保障机制。
对企业而言，关键不只是能否将支付决策交由 AI 处理，而是整个支付环境能否可靠地支持这些操作，同时确保授权界限明确，并建立清晰的问责机制。
1. 智能体支付通道：资金能否顺利流向所需目的地？
AI 智能体或许能够识别正确的付款需求，但只有在支付能够可靠执行的情况下，这项决策才能真正创造价值。
企业的业务往往涉及不同的支付方式、货币、司法管辖区及金融服务提供商。智能体支付必须能够在这样的环境中运作，将智能体作出的决策与实际处理及结算交易的基础设施连接起来。
这不仅关乎速度。支付系统还必须确保交易能够可靠执行、受到适当管控，并产生明确的处理结果。当交易涉及不同系统或司法管辖区时，系统之间的互操作性及结算安排同样重要。
IMF 的研究报告提出了一种值得参考的思路：应将 AI 智能体基于概率进行的推理保留在支付流程的前端，同时确保授权及结算环节按照确定性的规则执行。这样，即使交易由 AI 智能体发起，仍能确保支付的最终性及必要的管控。
换言之，AI 智能体可以判断需要支付什么款项，但支付基础设施必须负责确认交易是否获得许可，并确保交易能够可靠地完成处理与结算。
这种分工让 AI 智能体能够参与决策，同时确保资金的实际流转仍处于可控、可追溯的范围内。
2. 内置治理规则：如何规范 AI 智能体的行动权限？
当 AI 智能体能够代表企业采取行动时，其身份及授权范围便成为支付流程中不可或缺的一环。
企业必须清楚知道：正在执行操作的是哪个智能体、它代表谁行事，以及它获准执行哪些操作。其授权范围可以明确规定可向哪些供应商付款、能够发起哪些交易，以及可在什么限额内运作。
然而，获得授权并不代表智能体作出的每一个决定都是正确的。
当获授权的智能体作出错误的支付决定时，会发生什么？
假设一家企业允许 AI 智能体在指定限额内向获批准的供应商付款。
整个过程可能如下：
智能体接收付款指示。 它识别需要支付的款项，并收集相关资料。
智能体出现判断错误。 它误解了相互冲突的指示，选择了另一家供应商的银行账户。
交易进入最终审核环节。 系统能否在资金转出前识别账户信息不匹配的情况？是否会将这笔交易标记出来，交由人工审核？
企业必须及时应对。 如果系统发现问题，相关人员能否了解智能体作出该决定的原因，并及时阻止付款？如果交易已经完成，又由谁负责调查事件经过？
在这个情境中，智能体虽然获得了付款授权，但这并不能保证其最终决策一定正确。
这些风险说明，治理机制必须贯穿支付执行的整个过程，而不能仅依赖智能体部署前的检查，或交易完成后的事后审核。
监管机构与业界已经开始着手应对这一问题。2026年7月，新加坡金融管理局（MAS）发布了《Safeguards for Agentic Finance at Runtime》（SAFR）。这份行业白皮书由 MAS 在 BuildFin.ai 计划下，与金融机构及金融科技企业共同制定，提出在交易执行前设置治理检查点，以核实并记录 AI 智能体拟采取的行动是否符合相关机构的管控要求。
在实际应用中，AI 支付治理应涵盖明确的权限设定、交易限额、基于政策的检查机制、持续监控及风险升级处理流程。对于异常或风险较高的交易，应在执行前进行额外的风险审查。同时，保留智能体的决策记录，也有助于企业在出现问题时展开调查，并厘清责任归属。
具体需要采取哪些管控措施，则取决于支付系统本身、涉及的风险，以及交易完成前是否仍有介入和干预的可能。
治理的目标并不是要求每一笔付款都必须经过人工批准。真正的人本监督，是由人来设定自主行动的边界，在必要时保留介入能力，并确保责任归属清晰明确。
3. 拓展现实应用：智能体支付如何实现规模化？
单笔智能体支付的成功，只是一个起点。要让智能体支付在规模化应用中真正发挥价值，就必须与企业现有的支付基础设施、金融网络及业务流程有效衔接。
要让智能体支付成为日常业务运营的一部分，它必须能够通过现有支付渠道运作、连接既有金融系统，并满足不同的业务需求。
随着应用范围不断扩大，系统还需要应对更高的交易量、更多的交易对手方，以及更广泛的支付场景，从供应商付款到资金管理运营，都必须能够支持。
企业也需要考虑智能体能力如何融入现有工作流程、如何处理异常情况，以及随着越来越多支付决策实现自动化，如何持续维持有效监督。
归根结底，智能体支付的规模化发展并不只是实现交易自动化，而是将自主支付能力延伸至实际金融运营，同时维持企业所要求的可靠性、透明度及必要的安全保障。
企业在信任 AI 处理支付前，应评估哪些方面？
在将支付决策交由 AI 智能体处理之前，企业应先思考以下四个问题：
基础设施（Infrastructure）： 现有支付环境能否支持不同系统及支付方式之间的可靠交易执行？
授权（Authority）： 每个智能体的身份、授权范围及交易限额是否都有明确界定？
监督（Oversight）： 对于异常或可能存在错误的决策，系统能否在交易执行前进行识别、审核及质疑，并在条件允许时及时介入？
问责（Accountability）： 当问题发生时，是否具备完整的决策记录、明确的风险升级处理流程及清晰的责任分工？
智能体支付正在改变企业发起、协调及管理支付的方式。
然而，实现交易自动化只是其中一部分。企业若要真正信任 AI 处理支付，信任就不能在流程的最后阶段才被加入，而必须从一开始就融入每一个环节。
从智能体如何理解任务，到支付如何获得授权、执行及审核，每个阶段都必须具备明确的权限界定、适当的安全保障及清晰的问责机制。
Connect with us
Get in touch with our team at stablexkya@mce.sg to learn more about the StableX KYA Framework, or discuss partnership opportunities.