Why Is AI Agent Shopping Difficult to Popularize?
TechFlow Selected TechFlow Selected
Why Is AI Agent Shopping Difficult to Popularize?
The real bottleneck has never been payment.
Navigating Web3 tides with focused insightsWritten by: Anderl
Translated by: Chopper, Foresight News
Currently, a narrative circulates within the AI and crypto industries: equipping AI agents with wallets to handle shopping on behalf of users is the core use case for AI. This argument sounds concise and perfect, as if it were the future, but its internal logic does not hold up to scrutiny. This viewpoint confuses the difficult and simple aspects of shopping, placing the simplest payment process at the core.
Let us set aside payment for now and return to the act of shopping itself.
The Two Core Behaviors of Shopping
Shopping is essentially two actions now bound together: information retrieval and value judgment. Retrieval (collecting, filtering, comparing, preliminary sorting) has standardized attributes and can almost be entirely handed over to machine agents; whereas value judgment (whether this product is good, suits me, or if the merchant is trustworthy) is an aspect deeply bound to human subjective emotion.
Data has proven that information retrieval is rapidly shifting towards AI. Adobe Analytics data shows that from July 2024 to July 2025, traffic driven by generative AI to US retail websites surged by approximately 4700%. However, the "AI wallet" narrative defaults to the intelligent agent handling both retrieval and value judgment simultaneously, which is the key conceptual switch in this entire logic. Retrieval can be fully handed over to machines, but value judgment can only be partially delegated under specific conditions, and in most scenarios, it cannot be delegated at all.
Value Judgment is Divided into Two Layers
More critically, value judgment itself is not a single dimension but is divided into two parts. One part is evaluation, which tests various options based on a utility function. The other part is requirement definition, which means setting the utility function first: which dimensions are important, what their weights are, which values are binding, and what the ultimate meaning of "good" is.
Requirement definition is not completed once at the beginning but runs through the entire shopping process. Product compliance standards are decided by you; whether to abandon a product due to a broken zipper is decided by your judgment criteria; choosing which merchant depends on the values you prioritize. Every layer of filtering follows the logic of "human subjective standards × AI machine evaluation." Automation can only replace the evaluation step; the sovereignty of defining requirements always remains in human hands.
Many people mistakenly believe that humans only need to write a standard list once to completely let go, which seriously underestimates human decision-making logic. Extensive research in the field of decision-making confirms that human preferences are not fixed. Psychologist Paul Slovic proposed the constructive preference theory: there is no set of ready-made, fixed preferences waiting to be retrieved in our minds; preferences are gradually formed during the process of making choices. The "preference reversal" experiment confirms this: two equivalent research methods, choice and pricing, will yield completely different product rankings, violating the basic axioms of rational choice.
Ariely, Loewenstein, and Prelec proposed the "coherent arbitrariness" theory in 2003: even random numbers unrelated to anything, such as the last few digits of a social security number, will anchor people's psychological bids for ordinary products; and this anchoring effect does not disappear with consumption experience or market transactions. The so-called "stable preferences" are just an illusion of order constructed by humans.
Therefore, human-machine interaction cannot be a one-time filling of a standard list but requires continuous iterative communication. The AI agent poses targeted questions at each filtering node: "You previously valued durability; at what premium price point does durability no longer make sense for you?" Then humans define this boundary in real time.
The Real Dividing Line: Is Procurement a Chore, or the Enjoyment Itself?
The industry is accustomed to dividing scenarios using "standardized products/personalized products," but the divergence lies not in whether standards can be quantified; the real dividing line is: does the act of making a choice itself possess experiential value.
For printing paper, batteries, or products that need regular restocking, the selection process has no experiential value. No one is willing to spend energy comparing two almost indistinguishable ink cartridges. Such products are naturally suitable for being fully handled by AI agents; automatic machine repurchasing will not lose any experience.
For enjoyment-type consumption, the situation is exactly the opposite. Red wine, furniture, coats, books; selection itself is part of the consumption fun. If judgment power is handed over to machines, although time costs are saved, the core fun of consumption is directly deprived. Even if AI provides free answers throughout the process, people are unwilling to fully delegate.
For enjoyment-type products, AI agents should not make decisions with full authority but switch to information gatherers: completing retrieval, preliminary filtering, parameter matching, merchant qualification verification, extracting common product issues from massive reviews, refining 200 options down to 5, and then leaving the final choice to humans.
Triple Contradictory Dilemma: AI Inquiry, Historical Deduction, Autonomous Choice
Some may say: Can't AI just directly ask my judgment criteria? This method is precisely the inefficient mode that ordinary people dislike in daily life, and more fatally, repeatedly questioning criteria will distort people's final choices.
Wilson and Schooler conducted a jam tasting experiment in 1991: groups required to sort out reasons for preferences in advance gave rankings that deviated more from professional tasting standards; subsequent experiments proved that forcing people to list reasons for choices one by one would cause them to select decorative paintings with lower post-event satisfaction. Language can only describe surface features that are easy to express and cannot capture inner real preferences; feelings like taste and aesthetics can only be relied upon through perception and are difficult to define with text.
This forms a triple dilemma that cannot be balanced simultaneously:
- Proactive AI inquiry: Aligns with current real preferences, but interaction friction is extremely high, even distorting choices;
- Deduction based on historical behavior: Operation is smooth, but it will be confined to past preferences, killing new consumption exploration;
- Human autonomous judgment throughout: Fully retains choice power, but consumes a large amount of time and energy.
A fourth compromise solution can avoid the above disadvantages: recognition-based interaction, rather than item-by-item inquiry. AI directly displays 3 options, and humans only need to choose directly. This method both saves effort and is accurate because it does not require abstractly naming those standards that you often cannot name.
Classic choice overload experiments can corroborate this. Iyengar and Lepper conducted a jam tasting booth experiment in 2000; when 6 jams were displayed, the purchase conversion rate was far higher than with 24 jams. However, this theory is controversial, and subsequent multiple analysis experiments overturned this conclusion. Scheibehenne et al. summarized a large number of experiments in 2010 and found that there was no universal choice overload effect overall; Chernev, Bockenholt, and Goodman's 2015 analysis covering 99 studies showed that the overload effect only appears when product complexity is high, decision difficulty is high, and personal preferences are vague. The original authors later reviewed and mentioned that when facing 24 products, consumers lacked sufficient time to sort out their own preferences. True autonomy exists between "the agent applying my standards" and "the agent fabricating standards out of thin air based on my historical records."
Returning to "AI Wallet": Payment is Just the Least Important Step
Clarifying the above logic makes it clear to understand the narrative loopholes of "equipping AI with a wallet." This statement confuses three completely independent things: the decision-making entity, the execution entity, and the fund-holding entity. "Equipping AI with a wallet" only solves the fund-holding problem; fund custody only makes sense when the AI simultaneously possesses decision-making power.
Three situations. In the first situation, humans decide and pay themselves. In this case, the agent does not pay but acts as a scout. In the second situation, humans make the decision and delegate the execution task to the agent ("Yes, buy that"). At this time, the agent is responsible for checkout but does not need to custody funds, only requiring a limited scope and revocable authorization for this approved purchase. Only in the third situation, when the agent autonomously decides and pays without human presence for checkout, will the wallet itself bear the responsibility for payment.
Interestingly, the global payment industry has already implemented layered authorization solutions in 2025, completely distinguishing "authorization" from "fund custody":
- OpenAI jointly launched an intelligent agent commercial agreement with Stripe, generating shared payment tokens bound to a single merchant, fixed amount, and limited-time one-time use; AI cannot obtain complete bank card numbers;
- Mastercard released Agent Pay in April 2025, generating dedicated tokens limited to specific agents, designated merchants, and bound user authorization rules;
- Google launched the AP2 intelligent payment protocol in September 2025, clearly splitting "user requirement authorization credentials" and "AI procurement list credentials"; both types are verifiable encrypted credentials, perfectly corresponding to the "requirement definition/machine evaluation" layered logic mentioned above;
- Visa launched the Trusted Agent Protocol in October 2025, with an idea consistent with the above solutions.
Leading payment institutions have all proven the core viewpoint of this article: There is no need to hand over funds to AI; granting limited operation permissions is sufficient.
So, where is the true applicable scenario for AI independent custody wallets? Consumer retail scenarios hardly need AI custody wallets; the true landing space for this solution is in standardized bulk commodities and automatic machine-to-machine settlement. Coinbase and Cloudflare jointly launched the x402 protocol, filling the gap in traditional card payment channels: supporting automatic payments between agents without human intervention, 7×24 hours, billed by API calls, with transaction volume exceeding 100 million transactions within months of launch. Mastercard simultaneously launched Agent Pay for machines, serving high-frequency, low-latency small-amount machine settlements. This is the underlying infrastructure of the machine economy, not for personal shopping.
This narrative itself is not wrong, but the importance is completely inverted: AI independent custody wallets only have significant value in scenarios where products are highly homogeneous and single transaction amounts are low.
Where the Wallet Should Really Be Placed
This also explains the shift in focus of the crypto track in the past two years: no longer focusing on the consumer liberation narrative for individuals, but instead deepening institutional underlying infrastructure — stablecoin clearing, asset tokenization, enterprise-level services. This is not abandoning the AI shopping track, but returning to the field where the wallet custody model is truly adapted: the enterprise side. Enterprise procurement departments themselves are the institutional embodiment of "pure chore procurement"; the procurement process strips away subjective aesthetics and personal feelings, relying entirely on specifications, price, fulfillment, and contract terms for decision-making, essentially being a manual version of intelligent agent procurement. AI wallets just automate processes that enterprises have long matured; there is no behavioral model subversion.
Nowadays, many enterprises outsource office supplies and low-value consumables procurement; the core competitiveness of platforms like Mercateo and Amazon Business is not low prices, but reducing process costs: unified procurement catalogs, consolidated billing. Enterprises are willing to accept a small premium per item in exchange for a significant decrease in procurement labor costs; the process cost of low-value consumables is often higher than the value of the products themselves. AI agents can reduce the labor cost of placing orders to near zero while retrieving dispersed massive suppliers; procurement platforms only retain compliance verification functions: access supplier audit, unified reconciliation, anti-counterfeiting verification, and verification is precisely the real bottleneck of the entire process.
Therefore, the landing strategy cannot be simply summarized as "wallets serve enterprises rather than individuals"; the precise expression should be: tools match scenarios.
- Autonomous custody wallets (AI autonomous decision-making, holding funds, completing payment): Adapted to standardized product enterprise procurement, automatic machine-to-machine settlement, mainly focusing on B2B, M2M high-frequency repurchase scenarios;
- Personal consumption scenarios: No need to custody funds, only grant one-time, limited-scope payment tokens; AI can settle only after humans confirm orders.
"Equipping AI with a wallet" as a consumer-facing promotional title targets only a very small niche scenario; the true core landing market for this solution is the enterprise backend automation system.
Supplementary note: Enterprise procurement is not all standardized products. Some procurement decisions also cannot be handed over to AI for autonomous processing; judging law firms, acquisition targets, and core suppliers belongs to strategic choices, where subjective consequences have significant impacts. Like individuals choosing red wine, it must be decided by humans personally; autonomous custody wallets have no use in such scenarios.
This rule applies to all fields: AI independent custody wallets only exert value at the standardized product layer, while standardized business is concentrated within enterprises, with the largest volume and most intensive demand.
The Real Bottleneck Has Never Been Payment
Fund transfer technology has long been mature; there are no bottlenecks in the payment step; the real bottlenecks for AI shopping landing lie in two other aspects.
First, lack of trusted data sources. The premise of automated judgment is that data is true and reliable; once underlying information is distorted, autonomous decision-making AI will amplify the negative impact of false information at machine speed. The proliferation of fake reviews is already an open industry problem. The US Federal Trade Commission issued new regulations in 2024 (effective October 21) explicitly prohibiting fake product reviews, clarifying that generative AI significantly lowers the threshold for batch manufacturing fake reviews, with a maximum fine of $51,744 per violation; by the end of 2025, regulators had issued multiple warning letters.
The problem of counterfeit physical products is equally severe. According to 2025 data statistics from the OECD and the European Union Intellectual Property Office, the scale of global counterfeit trade in 2021 was about $467 billion, accounting for 2.3% of total global trade; EU imported counterfeits accounted for 4.7% of total imports, with clothing, shoes/bags, and luxury goods being heavy disaster areas, which happen to belong to experience-type consumption categories. Single-item traceability credentials, verifiable real reviews, independent third-party authentication, and single-item level transfer evidence (EU Anti-Counterfeiting Directive, US Drug Supply Chain Security Act have long mandated implementation for drug packaging) are prerequisites for AI to safely judge products.
Second, human requirement definition power cannot be automated. As long as requirement standards are defined by humans, subsequent filtering, comparison, and settlement can all be automated. The act of defining one's own requirements is naturally impossible to hand over to machines. If AI generates requirement standards for you, the preferences ultimately formed do not belong to you.
Equipping AI agents with wallets only solves the simplest fund step. The directions truly worth deep cultivation: safely and controllably automating filtering and evaluation, while returning two core rights to humans — defining judgment standards and enjoying the fun of final choices.
Summary
Finally, it must be emphasized that for experience-type consumption products, once procurement channels are commoditized, products will be everywhere. At that time, your product is no longer the product itself; the choice is. Platforms need to optimize their own information standardization level to facilitate AI retrieval: complete verifiable traceability, clean structured product data, guiding AI to drive traffic to their own platforms. On this basis, hold onto the core advantages that automation cannot replace, broaden the new consumption exploration of user aesthetic boundaries, and the pleasant experience of finally selecting products. Hand over information collection to AI agents; platform core competitiveness focuses on creating a better choice experience.
Join TechFlow official community to stay tuned
Telegram:https://t.me/TechFlowDaily
X (Twitter):https://x.com/TechFlowPost
X (Twitter) EN:https://x.com/BlockFlow_News
