The real bottleneck is never payment.

Written by: Anderl

Translated by: Chopper, Foresight News

Currently, there is a notion circulating in the AI and crypto industry: configure wallets for AI agents and let them shop on behalf of people, claiming this is the most core application scenario for AI. This narrative sounds simple and perfect, as if it's the future, but its internal logic does not hold up under scrutiny. This perspective confuses the challenges of shopping with simpler tasks and places the simplest payment process at the core.

Let's set aside payment for now and return to the shopping behavior itself.

Two Core Activities of Shopping

Shopping essentially consists of two actions that are currently bound together: information retrieval and value judgment. Retrieval (collection, filtering, comparison, preliminary sorting) possesses standardization attributes and can almost be entirely delegated to machine agents; however, value judgment (whether a product is good, whether it suits me, whether the seller is trustworthy) is the part deeply tied to human subjective emotions.

Data has proven that information retrieval is rapidly transitioning to AI. Adobe Analytics data shows that from July 2024 to July 2025, visits to U.S. retail websites driven by generative AI skyrocketed by approximately 4700%. However, the narrative of the "AI wallet" assumes that intelligent agents can simultaneously handle both retrieval and value judgment, which is the key to the conceptual confusion in the entire argument. Retrieval can be completely delegated to machines, but value judgment can only be partially delivered under specific conditions; in most scenarios, it cannot be fully entrusted.

Value Judgment Divided into Two Layers

More critically, value judgment itself is not a single dimension but is divided into two parts. One part is assessment, which checks various options according to a utility function. The other part is demand definition, which is to first set the utility function: which dimensions are important, what their weights are, which values are binding, and what the final meaning of "good" is.

Demand definition is not something that is completed only once at the beginning but runs throughout the entire shopping process. Compliance standards for products are determined by you; whether to abandon a product due to a broken zipper is determined by your judgment criteria; which seller to choose depends on the values you care about. Each filtering layer follows the logic of "human subjective standards × AI machine evaluation". Automation can only replace the assessment phase, while the sovereignty of defining demand is always in the hands of humans.

Many people mistakenly believe that humans only need to write the standard list once and then let go entirely, which severely underestimates the logic of human decision-making. Extensive research in the field of decision-making confirms that human preferences are not fixed and unchanging. Psychologist Paul Slovic proposed the theory of constructed preference: there is not a ready-made, fixed set of preferences waiting to be retrieved; preferences are gradually formed during the choice-making process. The "preference reversal" experiments validated this: two equivalent survey methods—choice and pricing—resulted in completely different rankings of products, violating the fundamental axioms of rational choice.

Alili, Lowenstein, and Prelec proposed the theory of "coherent arbitrariness" in 2003: even a random number with no connection, such as those following social security digits, can anchor people's psychological bidding for ordinary products; and this anchoring effect does not disappear with consumer experience or market transactions. The so-called "stable preferences" are merely an illusion of order constructed by humans.

Thus, human-machine interaction cannot be about filling out a standard list once; it requires continuous iterative communication. AI agents should throw targeted questions at each filtering node: "You previously valued durability; at what premium price does durability no longer seem worthwhile to you?", defining this boundary in real-time with humans.

The Real Boundary: Is Purchasing a Chore or an Enjoyable Experience?

The industry tends to distinguish scenarios by "standardized goods / personalized goods", where the divergence is not whether standards can be quantified but whether the act of making a choice itself possesses experiential value.

For items like printer paper, batteries, or goods that need regular restocking, the selection process has no experiential value. No one wants to spend energy comparing two nearly indistinguishable cartridges. Such products are naturally suited for complete delegation to AI agents, as automated reordering incurs no loss of experience.

For experiential consumption, however, the situation is quite the opposite. Wine, furniture, coats, and books—selection is part of the enjoyment of consumption. If decision-making power is given to machines, although time costs are saved, it directly deprives the core enjoyment of consumption. Even if AI answers questions throughout the process for free, people are still reluctant to fully delegate.

For experiential products, AI agents should not make decisions entirely on their own but switch to being information gatherers: completing searches, initial screening, parameter matching, seller qualification verification, extracting common issues from a vast number of reviews, narrowing down 200 options to 5, and leaving the final choice to humans.

The Triple Contradiction Dilemma: AI Inquiry, Historical Inference, Autonomous Choice

Some may say: Can't AI directly ask for my judgment standards? This approach is precisely the inefficient model that ordinary people often dislike; more crucially, repeatedly questioning standards can distort human final choices.

Wilson and Schooler conducted a jam tasting evaluation experiment in 1991: participants who were asked in advance to articulate their reasons for preferences ended up with rankings that deviated more from professional tasting standards. Subsequent experiments confirmed that forcing people to list reasons for their choices results in selections with lower post-choice satisfaction. Language can only describe surface characteristics that are easy to express, failing to capture true inner preferences; experiences like taste and aesthetic judgment can only be perceived and are hard to define in words.

This creates a triple dilemma that cannot be simultaneously accommodated:

• Proactively asking AI: aligns with current true preferences but has high interaction friction and can even distort choices;
• Inferring based on historical behavior: operates smoothly but can be confined to past preferences and stifle new consumption explorations;
• Human autonomous judgment throughout: fully preserves choice rights but consumes a lot of time and energy.

A fourth compromise solution can avoid the aforementioned drawbacks: recognition-based interaction rather than item-by-item inquiry. AI directly presents 3 options, and humans only need to make a direct selection. This method is both convenient and accurate since it does not require abstractly naming standards that you often cannot articulate.

Classic choice overload experiments can corroborate this point. Iyengar and Lepper conducted a jam tasting stall experiment in 2000; when 6 types of jam were displayed, the purchase conversion rate was much higher than when 24 types were presented. However, this theory is controversial, as later analyses and experiments have refuted this conclusion. Scheibehenne and others found in 2010 that no general choice overload effect exists across studies; further analyses covering 99 studies in 2015 showed that overload effects emerge only when product complexity is high, decision difficulty is great, or personal preferences are unclear. The original authors later noted that facing 24 products, consumers lacked sufficient time to clarify their preferences. True autonomy lies between "agents applying my standards" and "agents fabricating standards based on my historical records."

Returning to the "AI Wallet": Payment is the Least Important Component

Understanding the above logic reveals the narrative loophole of "giving AI a wallet". This narrative confuses three completely independent things: decision-making主体, execution主体, and funding holding主体. "Giving AI a wallet" only addresses the funding holding issue; funding custody only makes sense when AI also possesses decision-making power.

Three scenarios arise. In the first, the person makes the decision and pays. In this case, the agent does not make the payment but acts as a scout. In the second scenario, the person makes a decision and delegates execution tasks to the agent ("Yes, buy that"). At this point, the agent is responsible for checkout but does not need to hold the funds, only providing a limited, revocable authorization for the approved purchase. It is only in the third scenario that when the agent makes decisions and payments autonomously, with no human present for checkout, does the wallet itself take on the responsibility of payment.

Interestingly, by 2025, the global payment industry has already implemented layer-based authorization schemes, distinguishing "authorization" from "funding custody":

• OpenAI and Stripe launched a smart agent business agreement, generating shared payment tokens bound to a single merchant, fixed amounts, and limited-time one-time use, preventing AI from accessing complete credit card numbers;
• Mastercard released Agent Pay in April 2025, generating special tokens limited to agents, designated merchants, and binding user authorization rules;
• Google launched the AP2 smart payment protocol in September 2025, clearly separating "user demand authorization credentials" from "AI purchasing list credentials", both are verifiable encrypted credentials, perfectly corresponding to the aforementioned layered logic of "demand definition / machine evaluation";
• Visa proposed a trustworthy agent protocol in October 2025, following a similar approach to the above schemes.

Leading payment institutions collectively affirm the core premise of this article: there is no need to hand over funds to AI; merely granting limited operational authority suffices.

So, what is the true applicable scenario for AI to independently manage wallets? Consumer retail scenarios hardly require AI wallet management; the real ground for this scheme lies in the automated settlement of standardized bulk goods between machines. Coinbase and Cloudflare jointly launched the x402 protocol, filling the gap in traditional card payment channels: supporting automated payments between agents without human intervention, operating 24/7, and charging based on API calls, with transaction volumes exceeding 100 million within months of launch. Mastercard concurrently introduced Agent Pay targeted at machines, servicing high-frequency, low-latency settlements. This represents the foundational infrastructure of the machine economy rather than individual shopping.

This narrative itself is not wrong, but its importance is completely reversed: an AI independently managing wallets only has significant value in scenarios where goods are highly homogeneous and per-transaction amounts are low.

Where Wallets Truly Belong

This also explains the shift in the crypto space focus over the past two years: no longer promoting the narrative aimed at liberating personal consumption but digging deep into institutional foundational infrastructure—stablecoin settlement, asset tokenization, and enterprise-level services. This is not abandoning the AI shopping track, but returning to the areas where wallet custody models really fit: the enterprise side. Procurement departments are a systematic manifestation of "pure chore procurement," stripping subjective aesthetics and personal feelings from the procurement process entirely, relying solely on specifications, prices, performance, and contractual terms to make decisions, essentially functioning as human versions of smart agent procurement. AI wallets merely automate processes that enterprises have already matured, presenting no disruption to behavioral patterns.

Today, many enterprises outsource office supplies and low-value consumables procurement; platforms like Mercateo and Amazon Business do not compete primarily on price but on lowering process costs: unifying procurement catalogs and consolidating invoices. Enterprises are willing to accept slight premiums on individual items in exchange for significantly reduced procurement labor costs, where the process cost of low-value consumables often exceeds the product's intrinsic value. AI agents can reduce ordering labor costs to nearly zero while searching among scattered vast suppliers; procurement platforms only retain compliance verification functions: supplier admission review, unified reconciliation, and anti-counterfeiting verification, with verification being the true bottleneck of the entire process.

Therefore, the landing strategy cannot simply be summarized as "wallet services for enterprises rather than individuals"; a more precise statement should be: tool matching with scenarios.

• Autonomous wallet management (AI makes decisions, holds funds, completes payments): suitable for standardized goods enterprise procurement and automated settlements between machines, primarily in high-frequency B2B and M2M resale scenarios;
• Personal consumption scenarios: no need for fund custody, only granting one-time, limited-range payment tokens that AI can settle only after human order confirmation.

"Giving AI a wallet," as a promotional title aimed at C-end, targets just a tiny niche scenario; the true core market for this scheme lies in enterprise backend automation systems.

It should be supplemented that enterprise procurement is not entirely made up of standardized goods. Some procurement decisions similarly cannot be left to AI for autonomous handling; judgments about law firms, acquisition targets, and key suppliers belong to strategic choices, with significant subjective consequences, and like choosing wine, these must be decided by humans, rendering autonomous wallet management useless in such scenarios.

This rule applies across all fields: AI independently managing wallets provides value only at the level of standardized goods, while standardized business occurs primarily within enterprises, characterized by the largest volume and most concentrated demand.

The Real Bottleneck is Never Payment

Technologies for funds circulation have long matured, and there are no bottlenecks in the payment process; the real bottleneck for AI shopping implementation lies in the other two aspects.

First, there is a lack of trustworthy data sources. The premise of automated judgment is that data is real and reliable. Once the underlying information is distorted, autonomous decision-making by AI will magnify the negative impact of false information at machine speed. The rampant issue of fake reviews has already become a public problem in the industry. The U.S. Federal Trade Commission introduced new rules in 2024 (effective October 21) explicitly prohibiting false product reviews, clarifying that generative AI significantly lowers the threshold for bulk manufacturing fake reviews, imposing fines of up to $51,744 for violations, and regulatory bodies had issued multiple warning letters by the end of 2025.

The issue with counterfeit physical goods is equally severe. According to data from the OECD and the EU Intellectual Property Office in 2025, the global trade in counterfeit goods was approximately $467 billion in 2021, accounting for 2.3% of total global trade; counterfeit goods constituted 4.7% of the EU's imports, with apparel, shoes, bags, and luxury goods being the hardest hit, all of which belong to the category of experiential consumer goods. Product traceability certificates, verifiable genuine reviews, independent third-party evaluations, and item-level circulation documentation (the EU anti-counterfeiting directive and the U.S. Drug Supply Chain Security Act have long mandated traceable packaging for pharmaceuticals) are prerequisites for AI to safely judge products.

Second, the authority to define human needs cannot be automated. As long as demand standards are defined by humans, subsequent filtering, comparison, and settlement can all be automated. However, defining one's own needs can never be left to machines. If AI generates demand standards for you, the resulting preferences do not belong to you.

Equipping AI agents with wallets only addresses the simplest financial segment. The truly valuable direction to pursue is the safe and controllable automation of screening and evaluation while returning the two core rights to humans—defining judgment criteria and enjoying the ultimate choice experience.

Conclusion

Finally, it is important to emphasize that for experiential consumer goods, once procurement channels become commoditized, the products become ubiquitous. At that point, your product is no longer the item itself; the choice is. Platforms need to optimize their own information standardization levels to facilitate AI retrieval: ensuring complete traceability that can be verified and clean, structured product data to guide AI in directing users to their platforms. On this basis, maintaining the core advantages that automation cannot replace, broadening the new consumption explorations that expand user aesthetic boundaries, and ensuring the ultimate pleasure of selecting products. Handing over information collection to AI agents allows platforms to concentrate their core competitive advantage on creating a superior choice experience.