Summary
While the industry widely believes payment technology is the bottleneck for AI agent shopping, deeper analysis reveals the real obstacles lie in the non-delegable nature of value judgment, the constructive dynamics of user preferences, and the complexity of regulatory compliance. As development costs decline and fintech applications proliferate, regulators face new challenges from AI-driven complaints.
The Cognitive Fallacy of AI Agent Shopping
A popular narrative in the AI and crypto industries suggests that equipping AI agents with wallets to handle shopping autonomously represents AI's core application scenario. This seemingly perfect story actually conflates the difficult and simple aspects of shopping, misplacing the simplest element—payment—at the center.
A deeper analysis of shopping behavior reveals two core actions: information retrieval and value judgment. Information retrieval (collection, filtering, comparison, preliminary ranking) has standardized properties and can be almost entirely delegated to machine agents. Adobe Analytics data shows that between July 2024 and July 2025, generative AI drove a roughly 4,700% surge in traffic to U.S. retail websites, proving that information retrieval is rapidly shifting to AI.
However, value judgment—whether a product is suitable, whether a merchant is trustworthy, whether quality meets expectations—is deeply bound to human subjective emotions and cognition. The "AI wallet" narrative assumes intelligent agents can simultaneously handle both retrieval and value judgment, which is where the logical sleight of hand occurs.
The Dual-Dimensional Dilemma of Value Judgment
Value judgment itself is not unidimensional but consists of two layers: evaluation and need definition. Evaluation involves examining various options according to a utility function, while need definition involves setting the utility function itself: which dimensions matter, how to weight them, which values are binding, and what "good" ultimately means.
Need definition is not a one-time static process but a dynamic adjustment throughout the shopping journey. Product compliance standards, tolerance for quality defects, and value orientations in merchant selection—each filtering layer follows the logic of "human subjective criteria × AI machine evaluation." Automation can only replace the evaluation component; the sovereignty of defining needs always remains with humans.
Psychological research further reveals the deeper reasons for this dilemma. Psychologist Paul Slovic's theory of constructed preferences indicates that human preferences are not fixed but are formed gradually during the choice process. "Preference reversal" experiments confirm that choice and pricing—two equivalent survey methods—yield completely different product rankings, violating the basic axioms of rational choice.
Ariely, Loewenstein, and Prelec's 2003 theory of "coherent arbitrariness" further reveals that even random numbers like the last digits of a Social Security number can anchor people's psychological pricing for ordinary goods, and this anchoring effect does not disappear with consumer experience or market transactions. The so-called "stable preferences" are merely artificially constructed illusions of order.
The Dividing Line Between Enjoyable Consumption and Routine Procurement
The industry habitually divides scenarios by "standardized goods/personalized goods," but the real dividing line is whether the act of making a choice has experiential value.
For printer paper, batteries, or goods requiring regular replenishment, the selection process has no experiential value. No one wants to spend energy comparing two nearly identical ink cartridges. These goods are naturally suited for complete delegation to AI agents; automated repurchasing does not sacrifice any experience.
For enjoyable consumption, the situation is exactly the opposite. Wine, furniture, coats, books—the selection itself is part of the consumption pleasure. If decision-making authority is handed to machines, while time costs are saved, the core pleasure of consumption is directly stripped away. Even if AI provides free consultation throughout, people are unwilling to fully delegate.
For enjoyable goods, AI agents should not make decisions autonomously but should switch to the role of information gatherer: completing retrieval, preliminary filtering, parameter matching, merchant credential verification, distilling product weaknesses from massive reviews, narrowing 200 options to 5, and then leaving the final choice to humans.
This insight applies equally to institutional digital asset management. When custody service providers like Cobo offer AI-assisted decision tools to institutional clients, they must clearly define automation boundaries: risk parameter monitoring, compliance checks, and liquidity analysis can be automated, but final asset allocation decisions, counterparty selection, and risk preference settings still require human professional judgment.
Regulatory Challenges of KYA and Agentic Commerce
The upcoming Agentic Commerce and KYA (Know Your Agent) event on July 29 will delve into AI agent commercialization and identity verification issues. The urgency of this topic stems from new regulatory-level challenges.
The U.S. Consumer Financial Protection Bureau (CFPB) points out that AI-driven complaints will become a new regulatory challenge. When AI agents initiate large-scale complaints, disputes, or refund requests on behalf of users, traditional manual review mechanisms will struggle to cope. Regulators need to establish new identity verification and liability attribution frameworks, clarifying the legal entity for AI agent actions, authorization boundaries, and accountability mechanisms.
The concept of KYA (Know Your Agent) is an extension of the traditional KYC (Know Your Customer) framework into the AI era. How to verify an AI agent's identity? How to confirm it has obtained legitimate user authorization? How to audit the compliance of its decision logic? These questions are particularly sensitive in the digital asset domain.
For platforms like Cobo that provide institutional-grade custody services, establishing a KYA framework means adding AI agent review dimensions on top of existing KYC/AML systems: verifying the agent's developer, auditing its code logic, monitoring its transaction behavior patterns, and establishing circuit breakers for anomalous behavior. This is not only a technical challenge but a major upgrade in compliance obligations.
The Jevons Paradox and Fintech Proliferation
Cheaper developer output brings more developers; cheaper fintech development spawns more fintech applications. This Jevons paradox effect is about to sweep the entire industry.
In the 19th century, economist William Stanley Jevons observed that after steam engine efficiency improved, coal consumption increased rather than decreased. Efficiency gains lowered usage costs, thereby stimulating broader applications. The same logic is playing out in fintech: AI tools have lowered development barriers, enabling more teams to rapidly build financial applications, leading to an explosion in application numbers.
The latest update of Hermes Agent to the Quicksilver version, reducing cold start time by 80% and integrating LLM inference streams and smart approval functions, is a microcosm of this trend. Improved development efficiency will not reduce the total number of fintech applications but will instead catalyze exploration of more application scenarios.
This trend poses new challenges for regulators: How to maintain effective oversight in an environment of application proliferation? How to prevent systemic risks without stifling innovation? For the digital asset custody domain with strict compliance requirements, platforms need to build smarter risk control systems that support innovation while ensuring compliance baselines.
Reflection on Real User Needs
A survey shows that 36% of Americans want AI agents primarily to "leave your brand"—cancel subscriptions and switch services. This data reveals real user attitudes toward automated commercial behavior.
Users do not want AI agents to be brand extensions but to be agents of their own interests. This means the value of AI agents lies not in facilitating more transactions but in helping users make decisions more aligned with their own interests, including rejecting transactions, canceling services, and switching platforms.
This insight has profound implications for business model design. If AI agent incentive mechanisms are tied to platform interests, their credibility will be fundamentally questioned. Only when AI agents genuinely stand on the user's side—helping identify unnecessary subscriptions, discover better alternatives, and avoid impulsive consumption—can they earn user trust.
For digital asset service providers, this means AI-assisted tools should not become channels for pushing products but tools for helping users make rational decisions. When designing AI features, platforms like Cobo need to ensure their recommendation logic is transparent, conflicts of interest are auditable, and user interests take priority over platform revenue.
Payment Is Surface-Level, Compliance Is Core
Returning to the original proposition: Is the bottleneck for AI agent shopping payment technology? The answer is clearly no. Payment technology is already mature enough. The real challenges are:
- The non-complete delegability of value judgment: The dynamic constructive nature of human preferences determines the boundaries of automated decision-making
- Identity verification and authorization mechanisms: Establishing KYA frameworks requires coordinated innovation across technology, law, and regulation
- Attribution of compliance liability: When AI agents act on behalf of users, how is legal responsibility defined?
- Building user trust: How to ensure AI agents genuinely represent user interests rather than platform interests?
These challenges are particularly prominent in the digital asset domain. Institutional-grade custody services need to solve not only technical automation issues but also build a trustworthy AI agent ecosystem within compliance frameworks. This requires platforms to maintain a clear awareness of the irreplaceability of human judgment while advancing automation, seeking balance between efficiency and security, innovation and compliance.
AI agent commercialization will not happen overnight. Its success depends on correctly defining automation boundaries, establishing trustworthy identity verification mechanisms, and building transparent liability attribution frameworks. Payment is merely the simplest link in this complex system; the real challenges have just begun.
Source: link