Is the Party Already Breaking Up?

The biggest thing in retail commerce and payments in the past year has been agentic commerce. But amid all the euphoria, a more sobering and complex picture has developed.

Agentic commerce. We’ve seen this movie before.

Thirty years ago, at the dawn of the World Wide Web, Amazon.com, and retail Web sites, the Internet’s digerati hailed the looming changes that would come for “everything”—especially for the act of purchasing just about anything from electronic stores anywhere in the world.

Back then, so much was unknown about how such electronic commerce would unfold—how to make it work and how to make online shopping safe for credit cards—that the industry turned to the payment networks for guidance and direction.

The networks in turn deferred security and operational requirements for e-commerce to a bevy of technology providers (for example, Microsoft, IBM, Netscape, RSA), with little or no input from merchants or buyer users. The result was the much-maligned, costly, and fraud-prone card-not-present environment the industry has been stuck with since 2001.

Now come AI-powered shopping bots, perusing the vast reaches of the Internet to procure unprecedented deals with unprecedented ease on behalf of the digerati, optimizing everything in commerce for merchants and buyers alike. These bots offer unlimited, effortless product discovery and elimination of the “drudgery” of shopping, all while sparing purchasers from the “complexity” of paying for goods and services.

Agentic commerce, its proponents say, is another retail boom like Internet shopping and mobile commerce. A year-end report from McKinsey & Co. projected that by 2030, the U.S. business-to-consumer retail market alone could represent an opportunity—via agents—to orchestrate revenue in the range of $900 billion to $1 trillion. The implication: Everyone should jump on-board right away, or risk being left behind.

These early proponents—led by this generation’s most powerful digital payments providers, such as Mastercard, Visa, Stripe, Google and PayPal, and their AI partners—rushed to introduce agent buyers. These agents would operate on new platforms via new protocols on indefatigable computers rather than leave commerce to time-strapped and frail humans.

Better yet, these companies all promised to make their self-contained, agent-based processing platforms available as open “standards” to set the course of the next generation of friction-free digital payments.

Except…there was one thing conspicuously missing from these new ecosystems: meaningful inputs and participation by merchants and consumer organizations. Even the most merchant-friendly new platform, Google’s payments-agnostic Agent Payments Protocol (AP2), was designed with the support of 60 organizations—none of them individual merchants:

Unsurprisingly, some merchant and consumer organizations began to wonder about all the factors that might break things for them in this new paradigm—especially factors like liabilities, fraud, and chargebacks. Some merchants understood that signing up to participate on one of these platforms could limit operability to just that platform—or, require joining all of them.

On top of that, some of the AI large-language models (LLMs) were heard to say interchange rates were likely to double—to 5% on average—to cover the compensation they were looking for.

But unlike the case with buy now, pay later—where merchants knew they were paying more for what they could detect was an incremental transaction—with agentic commerce, their “customers” would be bots with little loyalty other than to the lowest possible price. And potentially, they’d never be heard from again.

So it was no surprise when Walmart announced in mid-October it planned to extend its generative AI-based customer support and shopping-assistant app, Sparky, into a full-fledged shopper-buyer that could give customers the tools to optimize their purchasing. Eschewing reliance on the “search bar” proclivity to deliver multitudes of unorganized responses to product inquiries and/or bias in favor of paid placement, the world’s largest retailer is espousing Open AI’s ChatGPT Instant Checkout capability as its own digital shopping agent.

Then, in early November, Amazon sued Perplexity—one of the more aggressive LLM players that is heavily promoting agent commerce with its Comet browser—for violating its terms by not disclosing when its AI agents are shopping on a customer’s behalf, disrupting Amazon’s “carefully curated shopping experience,” and “putting the security of our customers’ data at risk.” Perplexity claimed Amazon was “bullying” it, since other third-party applications (such as food-delivery apps and online travel agencies) do the same thing.

Coupled with questions about what problems this AI-based commerce actually solves for either buyers or sellers, who will pay for the costs of LLM participation, and who might actually get saddled with a still-unknown set of liabilities when bot buying goes bad, a backlash to agentic commerce appears to be already at hand.

Are we setting the stage for a “Son of CNP” redux, rushing once again into the cascade of complexities and uncertainties facing the payments industry without any obvious user inputs to guide us, only to fall back to the opaque and one-sided rules, rates, and rights of yesteryear?

What’s the Big Rush?

On the surface, a lot of this new functionality seems sensible, and even fun, right? An initial wave of testing using the first prototypical generation of agent buyers from the big four AI Large Language Models (Open AI, Anthropic, Perplexity, and Microsoft) seems to have the media awash in giggles and titters, though many transactions devolve to error-screen messages at this point.

Some say they can’t wait to see the functionality evolve and make their transacting lives better.

But other observers in the market appear wary of a host of potential security and operability issues, with most of the consequences to be felt by the users who pay for the payment systems: merchant and consumers. What’s the rush, indeed?

Clearly, the payoff for being first to set de facto standards for a new era of digital commerce looms as immense.  Fast-movers, the thinking goes, can lock up this wave of digital commerce, as they did with e-commerce and mobile commerce, simply porting over the established economics, rules, and functions of their existing online payments systems. And so might fast-mover merchants test the waters for a powerful new shopping experience for which they think they can drive (or protect) incremental sales.

Besides Walmart, other merchants reported to be testing or using agentic shopping include Amazon, eBay, Target, Sephora, and William-Sonoma. Moreover, a recent Boston Consulting Group (BCG) analysis says there are clear risks for retailers that do not adopt agentic buying services, mainly through disintermediation—which happens when retailers are bypassed in favor of AI platforms that complete the entire shopping journey outside of the retailer’s own e-commerce platform.

Infrastructure Requirements

As of today, the high-level infrastructure of the providers introduced so far (and there are lots of others coming) effectively packages the payment functions as-is, and wraps them with a layer of agent-management utilities. In essence, these are self-contained services that require few changes for users, but extend the providers’ control over the rudiments of transacting.

Peripherally, these providers also appear to be waving away some big concerns about new vectors of fraud and chargebacks, as well as operational questions that affect traditional buyer-seller relationships, by effectively saying, “Just trust us, like before!”  Consumers can limit agent activity by purchase amounts, time-in-service, or domain, for example. Merchant participation, it appears, requires only the ability to accept the knock of the bot to proceed with payments as they do now.

Notably, to make agentic commerce work for users, merchants will have to invest in yet another generation of online purchasing infrastructure (either by building or buying the required components/services) in order to avoid “breaking things” that now work (however suboptimally). Consumers will have to learn a bevy of new trappings for transacting from uncertain events they are likely to experience, including their own liabilities when things go awry.

That places a premium on sorting out, in advance, lots of new infrastructure that’s needed for users to transact safely and effectively, such as:

• Identifying and separating bots from buyers, and good bots from bad ones;
• Identifying and policing agents—probably leading to certification of some sort;
• Providing cross-acceptance of tokens from a potential multitude of sources;
• Retooling risk-management capabilities for real-time to handle many largely anonymous buyers

Things that Could Break

Today’s agent buyers seem more cute than capable to many as of now, and therefore are likely to make sense for just basic and simple purchases.  But as more complex transacting evolves, a host of potential and unresolved issues for merchants and consumers will need to be addressed.

Some of the many merchant issues inherent in agent buying include:

• Today’s agent-buyer prototypes deal with simple text prompts; agent buyers have no access to product SKUs, and little facility for detailed queries or interactions, so selections will largely have to be low-risk and commonplace;
• Provider program or user instructions for agent shoppers will obviate much of today’s online advertising and branding investments for merchants and product manufacturers, which can potentially reduce many products to commodities;
• Commodity product prices could well drive toward zero pricing, rendering many merchant Web sites as simple electronic catalogs, with no easily discernible differentiation for either the sellers or manufacturers;
• Merchants could lose visibility into what’s driving demand for their products, confounding normal marketing optimization strategies and techniques; upsells could be DOA without human interaction.

And here are just a few consumer issues that need to be addressed, including:

• Consumers experiencing problems with an agent purchase will have to adjust to different tests of validity, including claims of fraud or other chargebacks, which will likely soar (and increase incidence of friendly fraud);
• Buyers could find increased restrictions on purchase returns;
• Consumers who shop normally are likely to lose some or all credit or loyalty value for using agents;
• Agents could be borrowed or stolen or misused (e.g., under-18s obtaining restricted products; bypassing geographic boundaries on certain products; selling products illegally/without license by merchants catering to agents, etc.).

Where Agentic Commerce Could Help

And yet Agentic Commerce (“aCommerce”?) is the talk of the town, and many people feel it really is time for a better online purchasing experience. Many proponents of agent buying believe the existing drudgery of purchasing some products (have you bought travel or event tickets lately?) warrants all the excitement, as well as all the work that will be needed to build a new payments paradigm.

Agentic Commerce might or might not be the “agent” for such a transformation. It will certainly be worth finding out, once merchant and consumer voices can be heard and their legitimate needs reckoned

with. Merchant and consumer groups are talking up their own cross-industry working group to thrash out all the issues and concerns collaboratively.

Until then, there’s lots of work to do—and plenty of time to do it—before we all jump aboard the “fun car” and ride merrily off into what could be just the Son of CNP.

—Steve Mott is the principal at payments consultancy BetterBuyDesign.