Agentic commerce: what it is, how it works & what it demands from your payments infrastructure

Agentic commerce is a model where AI agents can search, decide, and complete purchases on behalf of a user based on predefined rules, preferences, and budgets. Instead of guiding the shopping process, AI executes it end to end, including discovery, payment, and post-purchase actions.

Here’s how agentic commerce compares to other commerce models:

There are a few things that make agentic commerce possible, reasoning chief among them. Today’s AI models are sophisticated enough to parse complex instructions written in plain English, interpret what the user actually wants, and figure out what next steps they need to execute. For example, if a user were to ask for a winter coat that’s under $250 and ships before next Friday, the agent would work through their conditions one at a time, including price, shipping, and quality signals such as reviews. 

AI agents are also able to handle when things go wrong. If, for example, one of the coats the agent shortlisted was back-ordered, it would choose another from another on its list, without the user ever having to intervene. AI agents designed to support agentic commerce use open standards such as the Model Context Protocol (MCP)¹, Google’s A2A Protocol², and the Agent Payments Protocol (AP2)³ to access systems, such as merchant’s catalogs, inventory systems, and the payment rails that finalize purchases.

What are different use cases for agentic commerce?

Use cases for agentic commerce are split between consumers and businesses and include personal shopping, subscription management, autonomous procurement, timing supplier payments, and managing foreign exchange (FX) risk exposure.

For consumers, personal shopping agents can handle anything from keeping your fridge stocked according to your household’s dietary needs and favorite brands to booking flights, hotels, and even airport transfers for upcoming trips. Managing subscription services is another good consumer use case: An AI agent might notice you haven’t used a streaming service in months and suggest, or even automatically make, changes such as downgrading your plan or canceling it, depending on the rules you gave it.

On the business side, autonomous procurement is the leading use case for agentic commerce. Agents can pull qualified suppliers from procurement systems, send out requests for quotes, and score responses based on criteria set by a company’s procurement team. In this example, the winning bid would automatically trigger an order and, if the deal falls within the limits leadership has approved, the agent would authorize the payment. 

Treasury teams can use agentic commerce in a similar way to time supplier payments, balancing cash position against FX risk exposure and any early-payment discounts on offer. The most ambitious business use case for agentic commerce is contract negotiation; some companies are already testing setups in which the buyer’s AI negotiates terms directly with the seller’s AI before either side’s teams sign off.

Agentic commerce works by handing the buying process over to an AI agent that takes it from a user’s initial brief through to delivery and any follow-up needed

1. Intent capture: The user tells the agent what to buy. Their instructions can be either specific or open-ended. If the instructions are open-ended, the agent asks follow-up questions to fill in missing details.
2. Discovery: The agent searches across merchant catalogs, marketplaces, review sites, and product feeds and presents the user with a curated shortlist.
3. Evaluation: The agent evaluates options on the shortlist against the user’s rules and makes a recommendation. The user can further refine the search at this stage by introducing additional parameters.
4. Decision and authorization: The agent either checks in with the user before buying or proceeds on its own within pre-approved limits. Standing authorization covers routine purchases such as groceries, while bigger or unusual purchases require explicit approval. 
5. Purchase: With the user’s permission, the agent initiates an agentic payment. The agent may do so right away, or wait until certain conditions are met, such as a price drop. Although the agent technically makes the purchase, the user is still responsible for it. purchase.
6. Post-purchase: The agent tracks the order through delivery. If something goes wrong, the agent steps in, either following the user’s preset rules or consulting them for a quick decision.

This entire process is powered by a few key technologies:

• Reasoning models let the agent think through open-ended requests. They’re the same kind of large language models (LLMs) behind tools such as ChatGPT, applied to a specific job: turning a goal into a plan, and adjusting that plan as new information comes in.
• Tool use refers to an AI model’s ability to call external software functions to complete a task. When an agent needs to search a product’s price, search a catalog, or submit an order, it calls a tool that handles that specific task and returns a result. 
• Open standards such as MCP, A2A, and AP2 keep tool use from requiring a custom integration for every connection between an agent and an outside system.

Agentic commerce refers to the full buying journey executed by an AI agent, while agentic payments refer specifically to the step in that process where the agent makes a purchase using pre-authorized credentials. Agentic commerce creates demand, while agentic payments capture revenue.

Agentic commerce is a broad category that encompasses everything an AI agent does when acting as a buyer, including:

• Gathering information from a user
• Searching for products
• Comparing products against the user’s criteria
• Deciding which product to purchase
• Completing the transaction
• Tracking delivery
• Making returns

The agent’s job spans the entire buying process, from start to finish. Agentic payments come into play at a specific part of the process: completing the transaction. 

When the agent is ready to pay, it submits the payment using a mandate pre-authorized by the user. This mandate is a record that specifies what the agent can buy, how much it can spend, and any other conditions the user has set. The agent presents tokenized credentials, such as a tokenized card, rather than the user’s underlying payment details, which keeps the user’s actual card information out of the agent’s hands. From there, the payment moves through payments infrastructure the same way as any other transaction does; it’s authorized, routed, screened for fraud, and settled with the merchant. 

The relationship between agentic commerce and agentic payments is one of scope. An AI agent compiling a list of items within a user’s budget is engaging in agentic commerce, as is an agent scanning reviews to determine which to purchase. Only the moment when the agent draws on its mandate to move money qualifies as an agentic payment.

Both agentic commerce and agentic payments raise important considerations for merchants. For agentic commerce, merchants must figure out how to make their products show up when an agent is searching, and how to present information in a way an agent can read and act on. For agentic payments, merchants must use payments infrastructure capable of handling transactions with no human at the keyboard, particularly for fraud and authentication.

Why is payments processing important for agentic commerce?

Payments processing is important for agentic commerce because the payments infrastructure determines whether agent-initiated transactions succeed, fail, or are declined, which makes it a critical control point. In an agentic world, authentication has to verify a delegated permission rather than a person, fraud systems have to model agent behavior, and routing has to happen at machine speed across multiple rails. Without payment processing built for these conditions, an agent can pick the right product, build the right cart, and submit the right transaction, only to have authorization fail.

How do agentic payments change the role of checkout?

Agentic payments makes checkout a background process where payment decisions may happen instantly or conditionally, such as when a price drops or an item comes back in stock. This requires payment systems that can make fast, deterministic decisions, and support machine‑to‑machine interactions, rather than relying solely on human confirmation. With agentic payments, checkout becomes invisible infrastructure that must operate reliably without user input.

Why is payments orchestration important for agentic commerce?

Payments orchestration helps manage the complexity that comes with agent-driven transactions by coordinating routing, fraud checks, authorization logic, and settlement from a centralized layer. This enables merchants to support agentic payments while maintaining control, consistency, and performance across channels. In an agentic world, orchestration is how merchants maintain control.

Payment systems need to be able to verify pre-authorized mandates from users, and everything downstream of that verification has to adapt to a buyer that isn’t a person. Enterprises preparing for agentic commerce will need revisit their payment stack in the following ways:

• Fraud detection: Fraud models watch for behavioral signals that might suggest a transaction is illegitimate, such as unusually fast checkout and billing and shipping mismatches. Most of these signals are trained on human behavior, which means AI agents are more likely to trip them when they go to make purchases. 
• This introduces two types of risk: legitimate transactions being blocked as fraud, and bad actors using AI agents (or pretending to) to push transactions through systems that have yet to adapt to agentic commerce. In this new world, where AI-driven shopping is likely to become the norm, fraud detection needs to use new models trained on what legitimate agent behavior looks like, and on the patterns that signal an agent is being misused or impersonated.
• Routing: Agents can transact at higher volumes than humans and have access to more payment options. Each rail has different cost, speed, and reliability characteristics, and an agent making a lot of purchases needs to be able to pick the right one for each transaction. Static, rule-based routing, which is set up to send everything through one or two preferred rails, won’t hold up to the volume or the variety AI agents will produce. 
• Compliance: Know Your Customer (KYC), anti-money laundering (AML), and consumer protection obligations don’t go away just because the buyer is an AI. With that said, these frameworks were written with a human counterpart in mind. Adapting them to agentic commerce transactions raises questions without settled answers (yet). Questions such as how to apply identity verification when the entity initiating the payment is an agent acting for an already-verified user, or how to handle disclosure requirements when the buyer never sees a checkout screen.
• Payments infrastructure: needs to capture the right information about both the agent and the user it represents, apply the relevant checks against both, and stay flexible as regulatory guidance catches up to the new transaction patterns.
• Auditability: If a user disputes a charge, or a regulator asks how a transaction was approved, enterprises need to be prepared to present an audit trail that shows what the user authorized, what the agent did with that authorization, and how the infrastructure responded at each step. That requires a tamper-resistant record across all three layers — and the requirement gets stricter as agentic transaction volumes grow and edge cases become more common.
• Multi-rail flexibility: Agents will pick whichever rail fits each transaction best, but they can only do that if those rails are available to them. Payments infrastructure needs to expose cards, real-time payments, account-to-account, stablecoins, and other options through a single integration, so an agent isn’t limited to whichever rail the merchant happens to support best.

The biggest risks in agentic commerce include agent identity fraud, unclear accountability, evolving regulatory requirements, and challenges in maintaining user trust. Let’s look at these in detail:

• Agent identity fraud: How can merchants spot the difference between transactions from legitimate agents and fraudulent ones? Bad actors can deploy their own AI agents — or impersonate someone else’s — to push fraudulent transactions through at speeds that would overwhelm a human review queue. Verifying agent identity is a new problem for fraud teams, and the tools to do it well are still in the works.
• Accountability: If an agent buys the wrong thing, or buys something the user didn’t actually want, who’s responsible? Is it the user, who set up the permission; the agent’s developer, whose model made the decision; or the merchant, who accepted the transaction? Existing consumer protection laws assume a human made the purchase and don’t map cleanly to a world where AI is acting on behalf of shoppers. Until courts and regulators settle this question, enterprises taking agentic transactions will operate without a clear liability framework.
• Consumer trust: Agentic commerce only works if users are willing to delegate purchasing decisions to an AI. That’s a higher bar than it might seem. Users expect clear visibility into what their agents are doing, easy ways to set and adjust spending limits, and confidence that they can override or revoke agents’ permissions at any time. Designing for that kind of trust — and earning it consistently — is its own challenge, separate from the technical work of making AI agents function.
• Regulatory uncertainty: Most payment regulations were written with human-initiated transactions in mind. Strong customer authentication rules under PSD2, for example, assume a person is present to authenticate; KYC rules assume the buyer is the person being verified. Regulators are aware of agentic commerce and have begun exploring how existing frameworks should apply, but settled guidance is some way off. Enterprises moving early will need to make their best interpretation of current rules and stay ready to adjust as new guidance arrives.