|
flevyblog
The Flevy Blog covers Business Strategies, Business Theories, & Business Stories.
|
Editor's Note: Take a look at our featured best practice, 6 Core Elements of Agentic AI (36-slide PowerPoint presentation). Agentic AI represents a shift toward autonomous, intelligent systems that can make decisions and take actions with minimal human intervention. Evolving from traditional machine learning, this technology enhances operations by automating complex workflows, optimizing decision-making, and enabling [read more]
* * * *
Conventional checkout assumes a person at the screen. They pick the product, choose how to pay, and confirm. Fraud tools, authentication steps, and even error messages are designed around that moment.
AI agents break the assumption. An agent can be asked to find a product within a budget, compare sellers, and in some cases complete the purchase. For now, this is a small and mostly experimental slice of online payments; conventional checkout is still how nearly all transactions happen. But card networks, large platforms, and payment providers have started building for it, which means payment teams can’t treat AI-initiated transactions as a purely theoretical problem.
Six requirements stand out, and each one touches a different part of the stack.
Several industry efforts address parts of this. Visa’s Trusted Agent Protocol deals with agent identity, Google’s Agent Payments Protocol (AP2) with signed records of user intent, and OpenAI and Stripe’s Agentic Commerce Protocol with agent-led checkout. These are young, they overlap, and none can yet be called the standard.
The least defensible design is giving an agent a customer’s raw card number. A reusable PAN in an agent’s memory, logs, or third-party tools multiplies the places it can be exposed, and it can be used for anything, anywhere.
Tokenized approaches narrow that. A token can be limited to one merchant, a maximum amount, or a time window, and it can be revoked without canceling the customer’s card. Mastercard’s Agent Pay, for example, uses tokens that are tied to a particular agent and to rules the cardholder sets. OpenAI and Stripe’s protocol describes a single-use token tied to a specific merchant and amount.
For a payment team, the practical question is whether the stack can accept and process these credentials alongside ordinary card payments. That means checking how network tokens and cryptograms are handled today, who provisions them, and whether your systems can distinguish a scoped agent credential from a standard card-on-file payment.
No single provider or network will cover every agent channel. Different assistants, platforms, and card schemes are backing different approaches, and which ones gain traction is not yet clear. A merchant wired to one provider gets whatever that provider decides to support, on its timeline.
The same applies to payment methods. Agent-initiated purchases may arrive over card tokens, wallets, bank-based payments, or, in some experiments, stablecoins. A stack that handles only a narrow set of methods narrows what an agent can buy from you.
This is one reason infrastructure that sits across providers deserves attention. Teams that already use payment orchestration providers to connect several PSPs and acquirers have routing, credential handling, and reporting separated from any one vendor, which makes adding a new channel less of a rebuild. Support for agent-specific flows varies between vendors, so it should be verified rather than assumed.
Fraud systems often treat unusual automation as a warning sign: very fast sessions, no mouse movement, no history on the device. Those are reasonable signals when the goal is to block bots testing stolen cards. They are also a fair description of a legitimate agent buying something its owner requested.
Blocking all of it would turn away real customers. Allowing all of it would invite abuse. The workable approach is to evaluate agent traffic on its own terms:
This is where context supplied by networks and protocols may help, for instance, a verified agent identity or a signed mandate. Risk teams should plan to treat agent-originated payments as a distinct segment, with their own thresholds and reporting, so they can learn what normal looks like as volumes grow. Today those volumes are too small for confident benchmarks, so assumptions should be tested rather than copied from human checkout.
Nobody can say which standards will last. The sensible response is to build for change instead of picking a winner.
AI-initiated transactions are early, and most payment teams don’t need a rebuild this quarter. They do need to know where their stack is rigid: credentials that can’t be scoped, risk rules that can’t be segmented, integrations that depend on one provider. Fixing those now is useful whether agent payments grow quickly, slowly, or take a form nobody has described yet.
You can download in-depth presentations on Agentic AI and 100s of management topics from the FlevyPro Library. FlevyPro is trusted and utilized by 1000s of management consultants and corporate executives.
For even more best practices available on Flevy, have a look at our top 100 lists:
These best practices are of the same as those leveraged by top-tier management consulting firms, like McKinsey, BCG, Bain, and Accenture. Improve the growth and efficiency of your organization by utilizing these best practice frameworks, templates, and tools. Most were developed by seasoned executives and consultants with over 20+ years of experience.
