Is a Wallet a 'Wallet' or an 'Agent'? What is Needed to Entrust Money to AI｜HashHub Research

This article is reprinted with permission from HashHub Research. Author: Derio Tencho Date of writing: August 27, 2026

Table of Contents

1. Premise
2. Can AI Really Shop if It Has a Wallet?
3. Transferring 'Authority' Instead of the Private Key
4. No Need to Entrust Everything to AI
5. Who Bears the Loss When AI Makes a Mistake?
6. Summary of Part One
7. References

Premise

When we hear that "AI has a wallet," it may seem like a story for the near future, but in terms of technology, systems for AI to perform remittances, asset exchanges, and payments for services have already begun to be implemented.

Specifically, Coinbase has been preparing an environment through AgentKit for AI agents to hold wallets and execute on-chain operations such as remittances and swaps. Stripe is also developing systems to support commercial transactions via AI, and Circle's Agent Wallets allow AI to handle USDC and other currencies while imposing spending limits and usage restrictions.

However, when we say "AI has a wallet," we are not talking about AI owning property like a human. At this point in time, it mainly involves granting AI certain operational permissions over assets managed by humans or corporations. Even so, if AI can assemble transaction details and execute payments, the experience from a human perspective will change significantly.

In this context, I personally wonder whether allowing AI to operate a wallet and humans being able to safely entrust money to it are truly the same thing.

For example, the rapidly expanding area is "what AI can do." On the other hand, it could be argued that the technically possible range and the range where humans can safely delegate may not necessarily align.

Considering this, in this article, I would like to examine the changes from the perspective of "how far can humans entrust money to AI," rather than focusing on the functions of AI wallets, dividing it into two parts.

Can AI Really Shop if It Has a Wallet?

To conclude, even if AI can execute payments from a wallet, that alone does not mean shopping can take place.

When we shop online at sites like Amazon or Rakuten, we first search for products, then check prices, stock, and delivery conditions before placing an order, and finally make a payment. If AI is to handle this flow, it must not only be able to read product information but also ensure that the seller can receive orders from AI and verify on whose behalf the AI is acting, processing everything from payment to delivery.

The Google Universal Commerce Protocol (UCP) has emerged to enable AI to handle this entire series of commercial transactions. UCP aims to facilitate interactions between AI and sellers regarding product information, checkout, payment, buyer identification, and post-order processing.

On the crypto side, x402 is attempting to simplify the payment aspect, allowing AI to access paid APIs, receive payment conditions presented by the counterpart, and return signed payment information to proceed directly to payment.

However, x402 only deals with "how to pay," and budget management, such as how much to allow AI to spend, is not included in x402's scope. In other words, the mechanisms for automating payments and allowing AI to make payments need to be considered separately.

Based on the official specifications of x402, created by the author.

Moreover, this challenge is not only being tackled by crypto. For instance, Stripe is establishing a system that connects product discovery to checkout and payment via AI, and Visa's Trusted Agent Protocol uses cryptographic signatures to help sellers distinguish between legitimate AI agents and malicious bots. In short, for AI to shop, it must not only be able to pay but also be accepted by sellers as a "legitimate customer representative."

From this perspective, what is needed is not to have sellers also hold crypto wallets, but to create an environment that naturally connects orders, authentication, and payments from AI to existing commercial transactions. While AI is already becoming "able to pay," for it to become capable of "shopping normally," the entire commercial transaction process surrounding it must be adapted to AI. Therefore, while the ability for AI to hold a wallet is a significant step, it does not complete Agentic Commerce by itself.

Transferring 'Authority' Instead of the Private Key

So, what if we were to transfer only the range of use to AI instead of giving it the wallet itself?

In traditional wallets, holding the private key is strongly linked to being able to move the assets of that account. For example, even if you want to entrust a payment of 1,000 yen, if you hand over the private key itself, it could allow operations on assets beyond that. In everyday life, it would be like asking someone to "buy a bottle of milk at the convenience store" while handing over not just the wallet but also the bankbook and personal seal—there's no need to go that far.

This structure is changing with smart accounts and account abstraction. For example, ERC-4337 and EIP-7702 are beginning to establish systems that allow wallets to have programmable mechanisms, enabling operations to be permitted conditionally rather than just checking whether the private key is held.

This allows the permissions granted to AI to be finely divided. You can set a usable amount, permit only specific assets, limit the usage period and payment destinations, and revoke permissions when they are no longer needed, thus delegating only the necessary operations instead of entrusting the entire wallet.

Importantly, the discussion of entrusting money to AI is no longer a simple choice between "to hand over the private key or not." Therefore, what will be questioned going forward is not just what to allow AI to do, but rather "what and to what extent to permit."

So, how far should humans actually entrust to AI? The next chapter will consider this "authority" in practical usage scenarios.

No Need to Entrust Everything to AI

As seen in the previous chapter, the permissions granted to AI can now be finely segmented. Therefore, there is no need to think of actual usage as a binary choice between "entrusting everything" or "entrusting nothing."

For example, you could let AI handle product searches and comparisons, while a human verifies before purchase. Automate small payments and only approve when the amount exceeds a certain threshold. Alternatively, you could delegate only the recurring payments that occur monthly. Such usage patterns can be considered.

In fact, this is the same with humans. You might tell the organizer of a drinking party, "I’ll leave the choice of the restaurant to you, but keep it under 5,000 yen per person. Don’t pick a place too far from the station. Show me before confirming the reservation. And don’t add champagne without asking." Why does it become a binary choice of "full delegation" or "no delegation" the moment it’s AI? When you think about it this way, the idea of finely segmenting AI's authority is not such a special concept.

In other words, you can keep the judgments you want to retain in your hands and delegate everything else to AI.

And here’s the interesting part,

Restricting the autonomy of AI does not necessarily diminish its value. This is because even if there is an AI that can do anything freely, if it is too frightening to use, the number of tasks that can be entrusted to it will not increase. In fact, having a clear boundary of "this much can be entrusted" makes it easier to use in everyday situations, as seen in the example of organizing a drinking party.

Therefore, "by slightly reducing the freedom of AI, the range of tasks that can be entrusted to humans actually expands." The author believes this perspective becomes crucial when dealing with AI that handles money.

However, a new problem arises here.

Even if the authority is properly restricted, there can still be mistakes made by AI within that range. For example, what happens if an AI that is allowed to spend up to 10,000 yen makes a wrong purchase of 9,800 yen?

The next consideration must be about "who bears the responsibility when a failure occurs within the permitted range," rather than "what is allowed."

Who Bears the Loss When AI Makes a Mistake?

When considering accidents involving AI wallets, it is essential to distinguish between "hacking" and "judgment errors."

If a third party steals authentication information and moves assets without permission, it is generally easier to categorize this as fraudulent use. However, the tricky part is when "a legitimate AI makes a mistake in judgment within the given authority."

For instance, if an AI is permitted to freely buy daily necessities up to 10,000 yen and mistakenly purchases a 9,800 yen item from a fake site thinking it was a legitimate store, the amount and authentication are within the rules, but the user still incurs a loss.

What is happening here is not an "accident due to improper authority," but rather an "accident where the authority was correct, but the judgment was wrong." In the earlier example of organizing a drinking party, it would be akin to reserving a different store with the same name while adhering to all conditions of a budget of 5,000 yen, proximity to the station, and non-smoking, resulting in cancellation fees. There is no violation of rules, but the judgment is incorrect. In this case, it cannot be simply stated that there is no problem because the authority was respected.

Such transactions are considered legitimate from the perspective of blockchain and payment systems. What is needed is a mechanism to verify "what the user requested from the AI and what the AI executed" afterward. For example, Google’s Agent Payments Protocol (AP2) records the requests and conditions given to the AI, as well as the purchased items and prices, allowing for verification up to the payment stage.

Source: https://youtu.be/yLTp3ic2j5c?t=340

However, there is another issue that remains.

The ability to prove what happened and who bears the loss are two separate matters. For instance, the AP2 does not establish specific rules for dispute resolution or compensation. At least currently, it cannot be processed as "the AI made a mistake, so it is the AI's responsibility," and ultimately, responsibility must be attributed to someone on the human side, such as the user, service provider, or corporation.

This point is somewhat similar to discussions around autonomous driving. In autonomous driving, discussions have not only revolved around how smartly a car can drive but also how to design responsibility and compensation in the event of an accident. For example, in the UK, a framework has been established where, in certain cases, "insurance companies first compensate for damages" in accidents involving autonomous driving. (Reference: UK Legislation)

Similarly, in AI wallets, what will become important is not just creating "fail-proof AI" but also how to handle situations after a failure occurs. Stopping authority, tracking transactions, recovery, and linking to refunds or compensation. Only when such mechanisms are in place can humans more easily entrust larger amounts or more critical tasks to AI.

Thus, as the range of tasks that can be entrusted to AI expands, the design of responsibility regarding "who bears the burden when a failure occurs" will become an unavoidable issue in AI wallets in the future.

Summary of Part One

The technology for AI to operate wallets, find products and services, and execute payments is already entering the implementation stage. However, what has become clear throughout this first part is that the speed at which "what AI can do" increases is not the same as the speed at which "what humans can safely entrust" increases.

Once payments can be made, the next question becomes "how far to allow usage." And once fine-grained authority settings are possible, the next issue will be "if the AI makes a mistake within the granted authority, who bears the responsibility?" Considering this, it becomes difficult to view wallets in the AI era merely as "wallets." On the other hand, calling AI a mere "agent" still leaves the authority and responsibility unclear.

With this in mind, what may become crucial moving forward is not what AI can do, but how humans manage the economic authority they grant to AI and where they can reclaim it.

So, how will wallets that take on this role differ from traditional wallets? And how closely will crypto and existing payment infrastructures approach similar designs when considering AI agents? In the second part, we will delve deeper into the role of wallets in the AI era starting from this question.

References

• Circle Launches AI Infrastructure to Power the Agentic Economy -- Circle
• Agent wallets -- Circle Docs
• Agent Platform Terms of Use -- Circle
• Coinbase for Agents: Your AI Agent Can Now Trade and Pay with Coinbase -- Coinbase
• AgentKit Overview -- Coinbase Developer Documentation
• Coinbase Developer Platform Terms of Service -- Coinbase
• Stripe launches the Agentic Commerce Suite to help every business thrive in the AI-enabled commerce era -- Stripe
• Agentic Commerce -- Stripe Docs
• Agentic commerce: How AI agents are changing the way businesses buy and sell -- Stripe
• X402 Protocol Specification -- x402 Foundation
• Universal Commerce Protocol Integration Overview -- Google for Developers
• Powering AI commerce with the new Agent Payments Protocol (AP2) -- Google Cloud
• Agent Payments Protocol Specification -- Google
• Visa Unveils Trusted Agent Protocol for AI Commerce -- Visa
• ERC-4337: Account Abstraction Using Alt Mempool -- Ethereum Improvement Proposals
• EIP-7702: Set Code for EOAs -- Ethereum Improvement Proposals
• ERC-7715: Request Permissions from Wallets -- Ethereum Improvement Proposals
• Spend Permissions -- Base Documentation
• AI agents Powered by Safe Smart Accounts -- Safe Docs
• Human approval for AI agent actions -- Safe Docs
• AI agent with a spending limit for a treasury -- Safe Docs
• Automated and Electric Vehicles Act 2018, Section 2 -- UK Legislation

※ Disclaimer: This report should not be considered as any form of legal or financial advice.
HashHub Research × CoinPost | Report List
View CoinPost published reports →
HASHHUB RESEARCH

HashHub Research is one of the leading research and analysis services specializing in the cryptocurrency and blockchain sector in Japan.
We provide multifaceted reports on macroeconomics, protocol analysis, market trends, and more for institutional investors and businesses. This report is part of exclusive content aimed at account holders of SBI VC Trade.
View other reports →
📋 Sample Reports
Read HashHub Research reports for free with a free account

SBI Group's VCTRADE Service Easy Account Opening Steps

1
Fill out the application form Enter personal information such as your name and address in the designated form
2
Submit identification documents Take and submit a photo of your face and documents (such as My Number Card) from your smartphone
3
Login and Deposit After the review is completed, once you receive your login ID, use instant deposit services to transfer funds
Easy identity verification via smartphone
Start trading as early as the same day

※ This does not guarantee same-day account opening
Open an account for free
No various fees, possible to start saving from 500 yen, a wide range of available stocks

【Important Notes】

The content provided by this service is for informational purposes only and does not constitute investment recommendations or legal/tax advice. We ask that cryptocurrency investments be conducted at the user's own research and responsibility. The information provided is obtained from reliable sources, but we do not guarantee its accuracy or completeness. Furthermore, regarding external links, we cannot be held responsible for any issues arising from the content or use of the linked sites. For inquiries regarding the content of this article, please contact HashHub Inc. (https://hashhub.tokyo/)

This content is provided for general informational purposes only and doesn't constitute financial, investment, legal, or tax advice. Any events, rewards, online promotions, or related information mentioned herein should not be considered a recommendation, solicitation, or invitation to purchase, sell, trade, or otherwise deal in any crypto assets. Crypto assets are highly volatile and may result in loss. The availability of WEEX services, products, and related events may vary by region. You are responsible for ensuring that your participation is in accordance with applicable local laws and regulations.