From Recommendation to Purchase

Agentic AI has a range of merchant-driven applications, from improved recommendation systems¹ to customer support.² However, it is also an external source of purchase intent; Avenue Z’s 2025 Beauty Index found that Perplexity triggered a shopping result on 92% of consumer prompts.³

Now, AI agents have moved from making recommendations to completing purchases. With Visa’s Intelligent Commerce program, a customer can authorize an agent to make purchases within pre-set guardrails using a one-time tokenized credential.⁴ Data can even be used to train the agent on the customer’s spending patterns and drive personalization.⁵ Mastercard similarly launched their Agent Pay solution this year,⁶ and providers such as Stripe and PayPal have developed agent toolkits that can be integrated into LLM providers that support function calling to make payments.^7-8

Merchants may not have seen agentic commerce before, but they have embraced many seismic shifts in the way customers pay. For those preparing for agentic shopping, there are some lessons we’ve learned before.

1. Classifications Matter

When it comes to shifts in commerce that require entirely new designations, this isn’t merchants’ first rodeo. They know that categories defined by the payments ecosystem have a tendency to stick.

Take Card Present and Card Not Present transactions. Today, customers move fluidly between online and in-store payment experiences: paying on their phone at a restaurant table, or returning online purchases in-store, or purchasing a parking ticket as they sit in their cars. Merchants know their online customers better than ever, with hundreds of data points such as IP address, device ID, and shipping address outpacing the often-more-limited information they can access in-store. Despite that, the overarching Card Not Present classification remains, and these payments often experience a lower chance of approval, higher fees, and greater authentication requirements due to the risk associated with their original classification.

Similar battlegrounds can be expected with Agentic AI. For example, if a customer is kicking off a purchase process with their AI prompt, is that mandate commutive to the AI agent, making the payment a form of Customer-Initiated Transaction (CIT)? Or do we need a designation for Agent-Initiated Transactions (AITs), especially for use cases where the agent makes a purchase without the consumer present to confirm? The industry so far has largely leaned towards the former, but if the past is anything to go by, these early decisions are meaningful, and it is crucial that merchants are at the table early to ensure classifications fit the reality on the ground.

2. Produce, or Partner?

Whether it’s search engines, marketplaces, or commerce platforms, the idea of selling via – or operating as – an intermediary is nothing new to merchants. In recent years traditionally-payment method providers such as Afterpay⁹ and Klarna¹⁰ have developed the functionality to act as their own hub for purchases – prompting the question of which comes first: the product or the payment flow?

Perplexity’s free Merchant Program allows affiliated merchants an increased chance of surfacing ‘recommended products’ by sharing product specs with the platform. At the same time, merchants are developing tools such as Amazon’s Buy For Me which is currently in testing,¹¹ and Google’s AI-powered shopping.¹² Merchants are also leading the charge in standard-setting; Google’s Agent Payments Protocol aims to establish a payment-agnostic framework for agent-led payments.¹³ Merchants embraced a ‘platform-first’ world once, and their ability to leverage agentic AI now is key to ensuring customers remain at the center of its growth.

3. The Intermediation Question

Linked to the age-old ‘build or buy’ question is the fear of intermediation in an agentic future. If a customer never has to access the merchant’s website to make a purchase, will retailers become commoditized? What role does brand loyalty play when agents can propose close substitutes? In reality, if merchants are part of the conversation, the world of agentic commerce doesn’t need to be zero-sum. Agents can link commerce through knowledge of customer behavior, identifying complementary goods (such as a smoothie after a workout class, or a suit after an event booking) to expand commerce – not shrink it.

4. The Rational Consumer

Agents can make choices and recommendations more quickly, more rationally, and more decisively than a customer can. For merchants used to ‘impulse’ buys and a well of unused loyalty points, interacting with a purely computational agent poses new questions. We’ve seen a flavor of this before, with discount-finding browser extensions or price tracking tools making customers more efficient. Agentic price comparison tools such as Google’s shopping assistant, when combined with actual purchasing commands, could make pricing arbitrage even more impactful.

The payments industry will likely grapple with the same shift. According to a recent study, 1 in 4 consumers did not redeem their credit card rewards in the past year.¹⁴ Imperfect actions like these could become a thing of the past, driving competition between card products in a similar vein to Open Banking-style aggregation tools.¹⁵

5. Teaching an Old Bot New Tricks

Despite some similarities with previous payments developments, there is one thing keeping merchants – and their fraud teams – up at night when it comes to agentic AI. Whether it’s detecting mouse movements on a page or clicking all the boxes which contain traffic lights, the fraud prevention industry has spent years building systems that detect and block malicious ‘bots’ from making purchases. The problem? AI agents behave just like them. Now, the software blocking bots could also be blocking their benevolent cousins from making good purchases, and payment and fraud teams must become closer than ever as they overhaul their entire approach to ensuring good customers can pay.

Balancing the Present and Future

Amara’s Law states that people tend to overestimate the effect of a new technology in the short run and underestimate its effect in the long run. For payments leaders, merely untangling the immediate is a challenge when it comes to agentic commerce. Are agents using guest or member checkouts? Who has liability for a chargeback? How do I prepare for a rise in network tokenized transactions? We recently tested a digital gift card purchase using an AI agent, which ended in a physical card being shipped, sales tax being applied, and a mess for the merchant.

Throughout all of this change, it’s important to remember that merchants have always been the intermediary between customers and the payments industry. Agentic commerce has significant potential to enhance their loyalty and personalization strategies and to connect commerce across brands. However, merchants are also the first to see challenges on the ground. With the past of commerce in mind, the industry must now ensure merchant voices are heard in building its future – agentic as it may be.