Researchers recently tested whether AI shopping agents could be manipulated by the tricks that steer human buyers online: countdown timers, hidden fees, pre-checked boxes, misleading defaults. The agents fell for them more than 70 percent of the time. The larger models, the ones with more reasoning capacity, did worse than the small ones.
The premise of the experiment tells you more than the result does. Everything about online commerce — visual hierarchy, urgency cues, the friction engineered to walk you toward checkout — was built to operate on human psychology. When AI buyers appeared on the horizon, the first question the field asked was whether the same machinery would still work on them.
That tells you where most businesses think their competition lives. They have spent two decades optimizing for attention, persuasion, and experience design, and the study suggests those investments still function, for now, because agents are navigating pages built for people. The infrastructure being assembled underneath them points elsewhere. This month Visa, Mastercard, and Ant International announced joint work on identity-verification frameworks for agent transactions: payment networks preparing to treat software as a buyer they can recognize and vouch for.
An agent does not browse. It compares offerings against stated criteria — price, specification, availability, delivery terms — at a speed no person would attempt for a routine purchase. Anyone who has dealt with a procurement department will recognize the shape of this. Procurement makes comparison explicit: evaluation factors are set in advance, attributes are weighted, and competition sharpens wherever requirements can be written down clearly. Agent-mediated commerce suggests that logic arriving in consumer markets, where an ordinary purchase starts to behave like a structured evaluation.
The compression this implies is real and uneven. Where a product is fully described by its specifications, as with commodity electronics or standardized components, the advantages built to capture attention lose their grip. Visual merchandising, the choreography of a checkout flow — all investments in persuading a mind that an agent doesn't have.
Procurement officers still pay brand premiums for industrial equipment, though, because part of what they are buying resists specification: reliability, service quality, how a vendor behaves when a shipment goes missing. That category does not shrink when the buyer is software. It may be where advantage concentrates, precisely because systematic comparison runs out of things to compare.
Agent-mediated transactions are already occurring: real transactions, with fulfillment and refunds when something goes wrong. Which advantages survive will depend a good deal on choices being made right now in the plumbing — what an agent is able to read, what a merchant is required to expose, who absorbs the cost when a machine buys the wrong thing.
The harder problem for most organizations is their capacity to take the question seriously at all. Twenty years of building around human attention produced teams, measurement systems, career paths, and a deep store of institutional knowledge about what makes someone click. None of that becomes worthless. But it was calibrated to a buyer who could be moved. So take what you sell and sort it: which parts can be specified, ranked, and compared by a system with no psychology, and which cannot. The second column is where the surprises will be, because much of what companies count as durable advantage has never been tested against a buyer that cannot be charmed.
-
Agents already placing orders: Modern Retail documented real orders arriving at merchants through Amazon's agent-mediated "Buy for Me" system, including cases where agents ordered out-of-stock products that had to be canceled and refunded.
-
Commerce protocols for agents: The W3C and GS1 held a joint workshop treating AI agents as new intermediaries between web content and users, considering identity, delegation, commerce protocols, and what machine-readable commerce means for web openness.
-
Dark patterns hit agents harder: The DECEPTICON study at ICLR found that manipulative interface designs diverted AI agents toward adversarial outcomes at more than double the rate observed with human users, and prompt-based defenses did not consistently help.
-
Structured feeds are spreading: Stripe's seller documentation now separates agent-facing catalog information into product-data, inventory, price, and promotion feeds with recommended refresh intervals as frequent as every fifteen minutes.