How Agentic AI Will Reshape Payments

Source: Sonja Davidovic and Hervé Tourpe; International Monetary Fund, IMF Note 2026/004 · 2026-04-24

Artificial intelligence (AI) is rapidly evolving from a tool that assists financial decision making into a technology capable of acting on behalf of economic actors. A new generation of agentic AI systems, software agents able to interpret objectives, plan multistep actions, and autonomously interact with digital services, has begun to emerge across e-commerce and financial markets. In payments, these systems are increasingly able not only to recommend transactions but also to initiate, coordinate, and manage financial operations under delegated authority. This development potentially represents a structural shift in the architecture of financial systems. Historically, payment infrastructures have been designed based on instructions initiated by individuals and processed by deterministic systems. Payment rails, from card networks to real-time gross settlement (RTGS) systems, rely on predictable rules, legal certainty, and clear accountability structures to ensure trust and financial stability. Agentic AI introduces a new element into this framework: probabilistic decision-making systems capable of initiating financial actions at machine speed.

Motivation and Scope The rapid emergence of agent-mediated payment use cases raises a set of design and policy questions that existing payment frameworks were not built to address. Rather than users explicitly initiating transactions, these settings rely on software agents operating under delegated mandates to anticipate payment needs, evaluate options, and coordinate execution across multiple instruments and rails. They may even be authorized, under certain conditions, to make payment decisions.

This evolution can be characterized as a shift from explicitly human-initiated transactions (“click-to-pay”) toward agent-mediated decision processes (“decide-to-pay”), in which execution increasingly occurs at machine speed and across multiple layers of the payment value chain, subject to predefined objectives, constraints, and governance arrangements.

A growing set of industry actors, including payment networks, technology platforms (for example, open-source software such as Ethereum), and AI model providers are in a race to experiment with these capabilities. Although current implementations remain largely focused on improving how people find and compare products, experimentation is now rapidly expanding across a broad landscape of payment‐related use cases, from fraud detection and compliance monitoring to treasury optimization and cross‐border payment orchestration, reflecting the widening scope of agentic AI pilots across the ecosystem (Boston Consulting Group 2025). The current velocity of innovation means developments that once required years can now materialize within months, indicating that more rapid and substantive change could be imminent.

In the context of this, a central architectural challenge emerges. Core payment infrastructures are built on deterministic logic, requiring predictability, auditability, and legal enforceability at each step of the transaction lifecycle. Agentic AI systems, by contrast, rely on probabilistic reasoning and adaptive decision making, which can yield different outcomes under similar conditions. This paper analyzes how and where these fundamentally different properties can be reconciled, and a three-layer model is proposed to assess the appropriate scope for integrating agentic AI into payment workflows.

What Is Agentic AI? Agentic AI refers to autonomous systems that sense their environment, set goals, and perform multistep tasks with little human input. For example, in the commercial sector, such systems may track household gas tank levels, identify the most cost-effective propane supplier, and automatically arrange for a refill. Within warehouse operations, these technologies can detect incoming deliveries, retrieve corresponding billing information, and initiate payment instructions in accordance with predefined protocols. Furthermore, the Bank for International Settlements (BIS 2025) notes that AI agents can independently manage liquidity and prioritize payments within real-time gross settlement systems, effectively mirroring established prudential cash management practices.

Unlike traditional AI models that provide static predictions (see Table 1) or require continuous human involvement, agentic AI systems integrate advanced capabilities such as planning, dynamic adaptation, and tool orchestration, thereby enabling these systems to operate as self-directed agents within complex ecosystems (FinRegLab 2025). Technology protocols and standards are evolving very rapidly to enable AI agents to interact with any payment actor and data source (Table 2).

AI agents are therefore very powerful, but constrained by their probabilistic nature, restricting their use in certain tasks in the payment chain. By contrast, payment infrastructures such as RTGS systems, card networks, instant payment platforms, and distributed ledgers operate deterministically:5 transactions follow preset rules, outcomes are binary, and legal finality is guaranteed. This difference creates a structural challenge. The main risk with agentic payments does not come from using probability-based reasoning itself, but from letting adaptive systems make irreversible payments without proper controls, checks, or accountability. Therefore, the central issue is not whether AI should be used in payments - as AI has already been used for more than four decades (see Table 1) - but how to keep uncertain, probabilistic decision making separate from automatic payment execution. The next section introduces a three layer-model to clarify these roles and architectures.

AI-Payment Layered Model Traditional payment systems tie authorization directly to explicit instructions from a user or institution, who starts a payment that is then checked against established rules. By contrast, agentic payment models rely on objectives and constraints to express intent. For instance, a buyer might instruct an agent to “buy the latest book on topic x when it becomes available on any platform, at the best price.” Here, the agent decides when and where to search for the book and compares prices from various merchants before suggesting optimal choices. Despite these advances, humans typically remain involved in reviewing and finalizing purchases, as the mix of probabilistic decision making and deterministic execution can complicate accountability if issues arise. the most impactful standards is Google’s UCP (Google 2026), which provides a shared grammar for discovery, comparison, and the ability for agents to create and manage post‐purchase logic.

Industry pilots already leverage these new standards. For example, Visa’s Intelligent Commerce and Mastercard’s Agent Pay test agent‐initiated shopping and payment flows, where agents autonomously construct purchase intent under predefined limits. Thus, Layer 1 serves as the locus for adaptive orchestration and delegated initiation, producing structured intent that must still pass through the deterministic authorization and settlement layers, which will be described next.

Evolution of E-Commerce AI agents are changing e-commerce by replacing simple automation with autonomous reasoning. With reduced human involvement, transactions can accelerate and consumption trends may shift. Agentic AI could enhance financial inclusion by reducing search costs and information asymmetries, allowing users to compare financial products, fees, and terms more effectively (Cao 2026). As agents prioritize low-latency payment solutions and remove inefficiencies, they could directly increase money velocity.

Several companies have launched agentic commerce initiatives to test potential efficiency gains for both operations and consumers (Box 1). In e-commerce, agents can reduce transactional friction by autonomously managing the purchase process, from product search and price comparison to discount application, availability verification, and execution. For merchants, this automation can accelerate purchase cycles, with evidence suggesting that LLM-based agents significantly shorten consumer interaction times in digital marketplaces (Yan and others 2025). Consumers benefit from personalized decision support that incorporates preferences, constraints, and real-time price signals. Agents can leverage the richer context provided in the user’s prompt to anticipate demand patterns and transmit structured feedback to merchants (Beard 2025). Besides lowering cognitive burdens for users, this personalization can increase customer lifetime value through more relevant product matching and streamlined purchasing experiences.

Agentic systems may also extend beyond checkout to post-purchase functions, including delivery coordination and returns management. Gartner (2025) estimates that AI agents can autonomously resolve up to 80 percent of common customer service issues by 2029. For merchants, automated returns processing and adaptive pricing mechanisms can reduce service costs, improve dispute resolution, and enable faster responses to shifts in demand and inventory conditions (McKinsey 2025b). Reflecting these expected gains, surveys indicate that firms are prioritizing AI investments in customer service, marketing, and product development (Microsoft 2025).

More broadly, agentic AI commerce services can spur modernization of e-commerce platforms as they enable integration across AI systems. For instance, PayPal recently announced strategic partnerships allowing merchants to seamlessly enable product discovery on AI platforms (PayPal 2025). There could also be important workforce consideration as agents could replace or augment operational roles (including merchandising, supply chain management, and product optimization), thus reducing operating costs and boosting productivity (Brynjolfsson, Li, and Raymond 2023).

The convenience of agentic AI for e-commerce introduces risks to consumer autonomy. Agentic systems may misinterpret user intent, optimize provider incentives rather than user welfare, drift from original objectives over time, or engage in subtle behavioral nudging at scale. Communication challenges between agents and end users might also arise because consumers do not always have a clear sense of their own short- and long-term financial goals or risk appetites (Aldasoro and others 2024). In multi-agent environments, there could be a misalignment in objectives between different elements or layers of an agentic system, or divergence between individual agent goals (Carichon 2025). Algorithmic herding occurs if dominant models identify identical market signals, while simultaneous execution could trigger flash crashes that bypass traditional circuit breakers (Ogbuonyalu and others 2024). Algorithmic herding can impair the functioning of payment systems by synchronizing liquidity demand, amplifying procyclical behavior, and creating congestion across payment rails, thereby undermining the predictability and resilience of core financial market infrastructures (BIS 2024). Too many economic agents being able to autonomously trigger and process financial operations creates a significant risk if malicious actors can influence or control agents through prompt injection attacks (Hogan Lovells 2025).

Ongoing access to financial data, ability to engage in behavioral nudging, and capacity for autonomous action can potentially increase the scale, subtlety, and personalization of activities that run counter to consumers’ interest (FinRegLab 2025). These dynamics have already arisen in the context of some AI-driven personal finance apps, which have been criticized for recommending cash advance offers and promoting borrowing products when a user’s account balance dips below particular thresholds, even in situations where those products would worsen consumers’ long-term financial outcomes (Aldasoro and others 2024).