Breaking the internet! Has the “AI buying spree” myth collapsed? Wallets are just a smokescreen—the real bottleneck leaves Wall Street stunned

The AI shopping stories you’ve heard about are probably wrong.

A story circulates in the community: give an AI agent a wallet, let it place orders on your behalf with full authority—this is claimed to be the most core real-world scenario for AI. It sounds plausible, but the logic doesn’t hold up under scrutiny.

Shopping, at its core, is two actions coupled together: information retrieval and value judgment. Retrieval—gathering, filtering, comparing, ranking—is standardized and something a machine can handle entirely. Value judgment—whether something is good, whether it suits you, whether the merchant is reliable—stays deeply tied to human subjective emotional depth.

Adobe Analytics’ data speaks clearly: from July 2024 to July 2025, visits generated by generative AI funneling traffic to U.S. retail websites surged by about 4700%. But the “AI wallet” narrative assumes the agent can handle both retrieval and value judgment at the same time—this is a concept switch.

Value judgment itself has two layers: evaluation and demand definition. Evaluation tests options against an existing utility function; demand definition sets the utility function itself—what dimensions matter, how to weigh them, and what “good” means. Demand definition is not completed once; it runs throughout the entire shopping journey. If a zipper is broken, should you give up? Which merchant is more trustworthy? Each filter layer is “human subjective standards × AI machine evaluation.” Automation can only replace evaluation; the reins of defining sovereignty will always remain in human hands.

Many people think that writing a standardized checklist once lets them fully let go, seriously underestimating human nature. Psychologist Paul Slovic proposed the “constructive preferences theory”: people’s preferences are not fixed in the mind waiting for you to retrieve them—they are gradually formed during the process of choice. The “preference reversal” experiments prove that two equivalent ways of researching options—choice and pricing—produce completely different product rankings, directly violating a foundational axiom of rational choice.

The “consistent arbitrariness” theory proposed in 2003 by Ariely, Loewenstein, and Prelec is even sharper: random numbers like the last digits of a social security number can anchor your psychological bid for ordinary goods; and even with rich consumption experience, this anchoring effect does not disappear. So-called “stable preferences” are nothing more than an illusion of artificially constructed order.

So human-AI interaction is absolutely not a one-time task of filling out a form; it’s continuous iteration. At each filter node, AI agents need to ask targeted questions: “You previously valued durability—at what premium does it stop being worth it?” The boundaries are defined in real time by humans.

The real dividing line is not standardized vs. non-standardized goods, but whether “the act of making a choice itself” has experiential value. For products like paper and batteries, the selection process has zero experience—so they are perfectly suited for AI’s automated repurchasing. Wine, coats, books—the selection process is the core part of the consumption enjoyment. Handing judgment to machines is effectively depriving people of the thrill of consumption.

Even if the AI provides free Q&A throughout, people still don’t want to delegate everything. For experience-based goods, the correct role of an AI agent is an information gatherer: complete retrieval, initial filtering, parameter matching, and merchant verification—reduce 200 options to 5, and leave the final decision to humans.

Someone might say: “Why not just have the AI ask me for my standards?” The opposite is true. Repeatedly interrogating your standards will distort your final choice. The jam evaluation experiment by Wilson and Schooler (1991) shows it all: those asked in advance to organize reasons for their preferences ended up with rankings that deviated more from professional tasting standards; later experiments proved that forcing people to list reasons one by one makes them choose decorative paintings with lower satisfaction afterward. Language can only describe surface traits that are easy to express; it can’t capture true internal preferences.

This yields the three-way impossibility triangle: asking AI proactively—fits real preferences, but interaction friction is extremely high and can even distort the choice; projecting from past behavior—smooth to operate, but trapped in past preferences and suppresses new exploration; humans fully autonomous end to end—keeps full choice authority, but consumes a large amount of time and energy.

The fourth path is “recognition-based interaction”: AI directly shows 3 options, and humans pick. It’s both convenient and accurate because you don’t need to abstractly name those standards that you often can’t clearly name. The classic choice overload experiment—Iyengar and Lepper (2000) showing 6 jams rather than 24, which produced a far higher conversion rate—was once treated as a rule of thumb, but later analyses overturned that conclusion. A large-scale synthesis by Scheibehenne et al. (2010) found no universal choice overload effect overall. Research analyzing 99 studies by Chernev, Böckenholt, and Goodman (2015) indicates that the overload effect appears only when product complexity is high, decision difficulty is high, and one’s own preferences are unclear. The original authors’ follow-up reconstruction suggests that with 24 products, consumers lacked time to organize their preferences.

True autonomy lies between “the agent applies my standards” and “the agent fabricates standards out of thin air based on historical records.”

Back to the “AI wallet” narrative—it mixes up three independent things: the decision maker, the executor, and the party holding the funds. Giving an AI a wallet only solves the funds-holding issue; it only becomes meaningful when the AI has decision authority.

There are three scenarios: the human decides and pays, and the agent acts only as a scout; the human decides and delegates execution (“Yes, buy that”), where the agent is responsible for checkout but doesn’t handle the money—only a limited, revocable authorization is needed once; only the third scenario—AI autonomously decides and pays without human presence—makes the wallet truly bear responsibility.

Interestingly, in 2025 the global payments industry has already rolled out layered authorization schemes. OpenAI, together with Stripe, introduced a smart-agent commercial agreement that issues payment tokens bound to a single merchant, a fixed amount, and one-time time-limited usage—so the AI doesn’t get the full card number. Mastercard launched Agent Pay in April: tokens are limited to agents, a specified merchant, and bound user authorization. Google rolled out the AP2 protocol in September: it clearly splits “user demand authorization credentials” and “AI procurement list credentials,” and both are verifiable encrypted credentials. Visa launched a trusted agent protocol in October, with the same idea.

Leading payment institutions have independently proven the core point: there’s no need to hand money over to the AI—grant only limited operational permissions.

So what are the real applicable scenarios for an AI independently custodial wallet? Retail consumption hardly needs it. The real space is standardized bulk goods and automated settlement between machines. The x402 protocol jointly launched by Coinbase and Cloudflare supports agent-to-agent automatic payments without human intervention, billed 24/7 via API calls; within months of going live, the transaction volume surpassed 100 million. Mastercard also launched machine-focused Agent Pay for high-frequency, low-latency, small-value machine settlement. This is the foundational infrastructure for a machine economy—not for individuals buying clothes.

This narrative isn’t wrong by itself, but its importance is completely inverted: an independently custodial AI wallet only has value in scenarios where goods are highly homogeneous and the per-transaction amount is relatively low.

This also explains the shift in the crypto sector focus over the past two years: it’s no longer the “consumer freedom” story aimed at individuals; it’s digging deeper into institutional underlying infrastructure—stablecoin clearing, asset tokenization, and enterprise-grade services. The procurement department of enterprises is a clear institutional embodiment of “paperwork procurement”: detach subjective aesthetics, decide purely by specifications, price, and contract terms—essentially an enterprise version of intelligent agent procurement. An AI wallet only automates processes that enterprises have already matured; it doesn’t upend behavior patterns.

Today, when many enterprises outsource office supplies procurement, the platform’s core competitiveness isn’t low price—it’s reducing process costs. Enterprises are willing to accept single-item markups in exchange for lower labor costs. AI agents can drive ordering labor costs close to zero while also searching across fragmented suppliers; procurement platforms can retain only compliance verification functions—which is exactly the real bottleneck.

The precise phrasing should be: tool-to-scenario matching. Self-custodial wallets fit enterprise procurement of standardized goods and automated settlement between machines; for personal consumption, there’s no need to custody funds—only one-time, limited-range payment tokens are needed.

Enterprise procurement is not all standardized either. Strategic choices like those of law firms and acquisition targets—like choosing wine—must be decided by humans personally. An independently custodial AI wallet only has value at the layer of standardized goods, while standardized business concentrates inside enterprises.

The real bottlenecks have never been payments. The technology for moving funds is already mature; the choke points are in two places.

First, there’s a lack of trustworthy data sources. The premise for automated judgment is that the data is real and reliable. Fake reviews flooding the market is already an industry-wide public problem. The U.S. Federal Trade Commission issued new rules in 2024 (effective October 21) explicitly banning fake reviews; it clarifies that generative AI significantly lowers the threshold for mass-producing fake reviews, with per-violation fines up to $51,744. Fake physical goods are equally severe: OECD and the EU Intellectual Property Office data for 2025 shows that in 2021, global trade in counterfeits was about $467 billion, accounting for 2.3% of total global trade; in the EU, counterfeit imports account for 4.7% of total imports, with clothing, shoes/bags, and luxury goods being the hardest hit. Single-item traceability documents, verifiable genuine reviews, and independent third-party appraisal are prerequisites for AI-safe product judgment.

Second, the power to define human needs cannot be automated. As long as the demand criteria are defined by humans, subsequent filtering, comparing, and settlement can be automated. Defining your own needs is inherently something that cannot be handed to a machine. If an AI generates the demand standards for you, the resulting preferences are not truly yours.

Equipping an AI agent with a wallet only solves the simplest part of the funds process. The direction worth deep investment is to automate filtering and evaluation in a safe and controllable way, while returning two core rights to humans—defining evaluation criteria and enjoying the pleasure of final selection.

For experience-type consumer goods, once your procurement channels become “commoditized,” your product is no longer the product itself—choice is. Platforms need to optimize the degree of information standardization so AI can retrieve it: complete, verifiable traceability, clean structured data, and guide users funnel traffic to their own platforms. Then, protect the core advantage that automation can’t replace: expanding consumers’ boundaries of taste discovery, and the joyful experience of finally picking products. When the information-gathering is handled by an AI agent, the platform’s core competitiveness focuses on building a better choice experience.

Follow me: Get more real-time crypto market analysis and insights!

#GUSD年化升至3.8% $BTC $ETH $SOL #GateDEX fully integrates RobinhoodChain #China Soft International teams up with the Moon’s Dark Side to deploy AgenticAI