Agentic Commerce and Agentic AI: Making Payments Invisible in Autonomous Shopping

The toughest snag in AI-powered shopping remains checkout: people still have to handle a card.

To make autonomous, agent-driven buying feel effortless—from craving a burger to delivery—payment plumbing still has roadblocks that companies are working to clear.

In broad terms, agentic commerce is a model where an AI agent can shop and transact on a user’s behalf. Instead of a person searching, comparing, and clicking “buy,” the user sets goals and constraints—like budget, timing, brands to avoid, delivery preferences, and approval limits—and the agent carries out the purchase within those guardrails.

In practice, the flow is straightforward: a user expresses intent (or pre-authorizes routine needs), an AI agent translates that intent into a plan, the agent evaluates options and selects a merchant or service, and then a payment system executes the transaction using the user’s chosen funding source and risk controls. The user’s role shifts from doing each step to setting preferences, approving exceptions, and reviewing outcomes; the payment layer’s role is to verify identity, enforce limits, manage authorization, and support disputes and fraud monitoring at machine speed.

The upside is largely about reducing friction: more convenience, faster decision-making, fewer abandoned carts, and more personalized outcomes that reflect a user’s history and preferences. Done well, it can also improve efficiency for businesses by automating routine service tasks and turning “I need this” into a completed order without the usual handoffs.

The risks are equally clear. Security and privacy become central because agents may have access to sensitive data and delegated payment privileges. Users can lose a sense of control if settings are unclear, and businesses have to contend with new fraud patterns, identity spoofing, merchant impersonation, and errors that occur at scale when automation is wrong. Even when fraud is contained, there are practical challenges around consent, refunds, chargebacks, customer support, and explaining “why the agent did that” in a way customers accept.

For businesses, preparing typically means tightening the basics and then designing for delegation: ensure product data and policies are consistent, make checkout and post-purchase actions accessible via secure APIs, build strong authentication and permissioning, and implement clear limits and confirmations for higher-risk transactions. Operationally, teams need compliance review, logging and audit trails, and customer education that spells out what an agent can do, what it cannot do, and how to revoke access quickly.

Real-world use cases extend beyond retail carts. Autonomous shopping for routine replenishment is one; automated bill payments and subscription management are another. Travel booking—where an agent can compare options, apply preferences, and handle changes—fits the model, as do B2B workflows like reordering supplies, reconciling invoices, and triggering payments once delivery conditions are met.

Compared with traditional e-commerce, the biggest change is who “drives” the session. Traditional flows assume a human is browsing and explicitly clicking through each step; agentic commerce assumes software can discover, decide, and execute under delegated authority. That shifts the experience from page-level persuasion and manual checkout toward machine-readable offers, policy-based approvals, and payments that happen in the background when conditions are satisfied.

The enabling technologies are a mix of AI and payments infrastructure: AI agents built on large language models, machine learning for risk scoring and anomaly detection, payment APIs and tokenization to reduce exposure of sensitive credentials, and security controls such as authentication, authorization, and policy enforcement. On top of that, agent frameworks, observability, and audit logging matter because businesses need to monitor what agents are doing and why.

The Agentic Commerce Protocol (ACP) is often discussed as the connective tissue that helps agents and commerce systems communicate in a consistent way. In essence, it is a set of conventions for how an agent discovers offers, requests a quote, confirms terms, and triggers payment and post-purchase actions—so workflows can be more interoperable across merchants, platforms, and payment providers without every integration becoming bespoke.

Looking ahead, the direction is toward more delegated purchasing with tighter controls: more granular permissions, better identity and risk tooling, and clearer customer experiences for approvals and exceptions. As these systems mature, the impact is likely to spread from consumer shopping into customer service, travel, healthcare administration, and enterprise procurement—anywhere routine decisions and transactions can be safely automated without eroding trust.

John Kain of Amazon Web Services says the aim is straightforward: build an AI agent that can decide and execute transactions on a customer’s behalf.

Advancing that agenda, AWS and Visa announced in December a collaboration to create tools that enable network-agnostic agent workflows for the payments layer of AI commerce.

As the companies put it, the objective is to provide a secure, scalable foundation for the next generation of intelligent commerce solutions.

AWS is Amazon’s fastest-growing and most profitable division. Operating income rose 14% to $45.6 billion, with sales up 20% to $128.7 billion, the company noted in its annual report last month. New York–based Kain, a former JPMorgan Chase executive, discussed the work in a Thursday interview.

Editor’s note: This interview was condensed and clarified.

Where Does AWS Concentrate Its Work in Agentic Commerce?

John Kain: We prioritize the payments execution tier. While shopping journeys are important, the essential layer is processing secure transactions at scale so agent-driven buying actually functions. We’re focused on making it easier for customers to build applications that integrate agents with existing payment rails.

What Does AWS Provide for Autonomous Shopping: Platform, Software, or Services?

It’s a platform-plus-partners model. Customers can choose from models by OpenAI, Anthropic, AWS, and other providers, tailoring the fit to each use case. The environment keeps prompts, inputs, outputs, and any training confined to the customer’s own context. With emerging agent frameworks, teams can assemble agents using generative AI while adding governance, controls, and policy guardrails to operate reliably at scale.

Beyond Retail: Which Long-Term Use Cases Will Dominate?

Most services people rely on will see major user-experience gains. One of the largest early areas is customer support: conversational systems that infer intent and route users to the right resource—often an autonomous agent rather than a human queue.

How Are AWS Customers Countering Fraud?

Traditional machine learning remains a mainstay. Mastercard and Visa run ML solutions on AWS that can detect suspicious activity in mere milliseconds. As payments infrastructure shifts toward instant rails, more volume will flow there; because settlement is final, fraud prevention and detection become central both to rail selection and to live monitoring. That challenge is largely about data and AI.

Is Fraud Slowing Adoption of Real-Time Payments?

No. Even before deploying instant-payment rails, firms were investing significant research and effort in this area.

With Similar AI Tools, Is the Fight With Criminals Even?

Organizations with large datasets and long-standing customer relationships have a structural advantage. AI performs best with high-quality data: deep knowledge of typical behavior raises the odds of spotting anomalies. That history puts established firms ahead of newcomers and fraudsters.