When AI Starts Spending Money: Who Will Underwrite Agent Transactions?

Original Article Title: Payments in the Agentic Economy

Original Article Authors: Saurabh Deshpande, Oliver Jaros, Decentralised.co

Original Article Translation: AididaoJP, Foresight News

In the article "Internet Pricing," we previously discussed: When payment is frictionless, machines will automatically make payments. Humans have not fully embraced micro-payments because focusing on the measuring process requires effort and attention. But machines are different; they only see 1s and 0s. Their cognitive capacity or task switching does not affect their execution. If breaking down to sub-cent levels can make the process more efficient, they will do so, unlike humans.

In our previous article, we ended with a question: What happens when the agent messes up? The agent's intent is not important. The key is that we cannot oversee the agent step by step.

This leads us to a dilemma: new technology has not inherited a significant advantage of the old infrastructure, such as the ability to reverse payments in case of errors. This article aims to explore this issue. We will discuss what agents need to achieve autonomy, who is building the foundation for it, and why new startups are emerging at the intersection of blockchain payment channels and autonomous agents.

Emerging Standards

Any business activity involves three parties: the buyer, the seller, and the intermediary facilitating the transaction. The intermediary can be a platform like Amazon or a payment processing network like Visa.

Buyer

Consumer applications typically handle funds or transactions and take a cut. But what happens when the consumer is an AI acting on our behalf? Several emerging standards are currently seeking answers to this.

ChatGPT has 700 million active users, all trying to obtain information or services through AI. While we have not yet directly bought and sold goods through an agent interface, we commonly use it to "discover" products. Whether buying running shoes or finding a hotel in El Calafate, I use AI for price comparison. If direct purchases could be made on the same interface, it would undoubtedly be much more convenient. This is precisely the goal of OpenAI partnering with Stripe to introduce the Autonomous Commercial Protocol (ACP).

Source: OpenAI

This is currently the most direct way for agents to handle funds: user full control. After the user places an order, ChatGPT sends the necessary information to the merchant's backend through ACP. The merchant then decides to accept or reject the order, processes the payment through the existing payment service provider, and handles shipping and customer service as usual.

You can think of ACP business as follows: you authorize an intern to spend a fixed budget, and you have the final say in choosing which product/service to purchase from which merchant and completing the payment.

OpenAI and Stripe have ACP, while Google has introduced the Agent Payment Protocol (AP2). Before delving into AP2, let's take a step back. Google aims to address the "interoperability" issue. Currently, AI agents operate independently: Gemini doesn't talk to Claude, and ChatGPT doesn't know what's happening in Perplexity.

Ideally, when tasks become more complex and require collaboration, we hope these agents can communicate using a common language. To achieve this, Google has developed A2A (Agent-to-Agent Protocol), allowing different agents to communicate and coordinate.

But just being able to converse is not enough. Agents also need to be able to use tools, access APIs, and services. The Model Context Protocol (MCP) enables agents to use tools such as Google Calendar, Notion, Figma, and others.

Source: Level Up Coding

The MCP defines a common language. As long as they both "speak" MCP, agents can use any tool without the need for additional custom code. The protocol was created by Anthropic, but the specification is open and is quickly being adopted by various companies. The MCP server essentially acts as a translation layer, positioned in front of a company's existing APIs, exposing services in a standardized format to any MCP-compatible agent.

Returning to AP2, it can be understood simply as follows: MCP gives agents the ability to access data, files, and tools; A2A gives them the voice to communicate with each other; and AP2 provides them with a wallet to securely spend money.

All of these protocols place the user at the center, with agents having limited spending authority. This addresses distribution and process issues, but it still doesn't address: what happens if an agent makes a mistake?

Seller

The story is not only about the buyer. Sellers are also setting new standards, focusing on how machines can pay for access to APIs, data, and content.

Currently, the most discussed standard is the x402 standard, an open protocol developed by Coinbase. It resurrects the long-defined but never used HTTP status code 402 from 1997 — "Payment Required." x402 combines this with stablecoin payments, enabling microtransactions to settle economically, thus giving new life to this status code.

x402 turns an HTTP request into a payment request. Whenever payment is required, the server makes the demand. As the agent has a preset budget, it pays the server and receives the data in the same process. This enables "pay-as-you-go" or "pay-per-call" to become feasible in machine-to-machine transactions.

With x402, the agent can precisely pay for what is needed at the moment. For example, paying 2 cents to read a paid article or paying a fraction of a cent for an API call. Transactions settle on-chain within seconds, without the need to establish a long-term relationship.

Source: Coinbase's x402 Whitepaper

Cloudflare has adopted this concept and built a more specific "pay-per-fetch" system. It also uses HTTP 402 at its core, but the key is Cloudflare's market dominance, with 20% of global internet traffic passing through its network, giving it significant influence.

"Pay-per-fetch" leverages Cloudflare's edge network, requesting payment before providing content to AI crawlers. This turns access to content into a metered force. Publishers are facing a plummet in traffic as people no longer land on websites via search engines but directly consume AI-generated summaries. Through this system, publishers can charge the AI lab every time the crawler accesses content.

Card networks are also attempting to extend existing payment channels to handle agent transactions. Visa has introduced the MCP Server and Acceptor Agent Toolkit. Mastercard has a project called "Agent Pay." Both are in early pilot stages, but their significance lies in Visa and Mastercard's global distribution networks, issuer relationships, and extensive merchant acceptance networks. The basic idea is to register agents, set spending controls, and enable agents to transact on the existing human credit card payment network.

Urgent Need to Fill the Trust Gap

All of the above standards assume that payments will proceed smoothly and the outcomes will meet expectations. Both ACP and AP2 involve human intervention at the checkout stage, providing a certain level of security. The x402 variant deals with machine-to-machine data access, with generally lower risk. The card issuer extends its familiar protection mechanisms, but at the cost of slow settlement and high fees.

For large-scale micropayments, speed is the primary goal. Card payment networks take several days to settle, with merchants paying a percentage of the transaction amount in fees. Cryptocurrency channels settle in seconds and cost less than a cent. However, this efficiency comes with irreversibility—once a cryptocurrency payment is made, it cannot be undone.

Traditional commerce has built a whole infrastructure around "things can go wrong." When an issue arises with a credit card purchase, you have a process to follow: contact the bank, raise a dispute, have the card issuer investigate and possibly withhold funds, and ultimately reach a decision on a refund or support for the merchant. In 2025, there were 261 million disputed transactions worth a total of $340 billion.

However, agents operating on stablecoin channels have none of these protections.

When agents begin to collaborate, the situation becomes more complex. When hundreds or thousands of multilayered agent workflows intertwine, clarifying responsibilities could become a nightmare.

Card issuers will not take on this risk, at least not under their current profit models. Visa and Mastercard's agent initiatives still charge standard interchange fees, and settlements still take days. They could move to instant stablecoin settlements, but that would mean giving up the dispute resolution system that forms the basis of their charges.

The dispute resolution mechanism in traditional finance was not inherent. The first credit card (Diners Club) emerged around 1950, but consumers had to wait another 24 years to gain transaction dispute rights. The modern infrastructure we take for granted today was gradually built up as issues arose.

Autonomous agents don't have as much time to waste. API requests make up 60% of the dynamic HTTP traffic Cloudflare handles. Bots and automated traffic represent nearly half of internet traffic. ChatGPT's 700 million users can now check out directly on Etsy via ACP, and Shopify integration is on the horizon. Transaction volumes already exist, users have the potential demand for agents to handle tasks, and agents for commercial purposes are not far off.

Therefore, we face a choice: continue with the slow settlement of traditional financial infrastructure, or consciously build trust infrastructure to match the swift blockchain settlements. The former will constrain agent potential, while the latter is an opportunity and a natural extension of autonomous agent business development.

So, What's the Specific Process?

As expected, this involves two parts: pre-transaction and post-transaction.

Pre-Transaction: Should Proxy Transactions Be Allowed?

This depends on three factors: identifying the counterparty, fraud detection, and using reputation scores to determine pricing and access rights.

In the U.S., Plaid connects nearly half of bank accounts, verifying millions of accounts daily. When you verify your identity on Venmo, you are using Plaid.

Currently, any agent interacting via API, scraping web pages, or initiating payments lacks peer-to-peer authentication. The server only sees a vague ID (such as a wallet address or API key) and does not know who the caller is. Without a universal cross-service identity, there is no way to build trust, and each interaction starts from a "zero-trust" standpoint.

By 2024, U.S. adults are projected to lose around $47 billion due to identity fraud.

We need a "Know Your Agent" (KYA) layer, similar to what Plaid provides for fintech's identity infrastructure. It should issue persistent and revocable credentials that bind agents to the humans or organizations behind them.

Card networks have spent decades building systems that can identify suspicious patterns from millions of transactions. They understand typical human spending behavior and can flag anomalies in real time. If an agent is compromised and conducts unauthorized transactions across multiple merchants, there is currently no shared fraud graph to detect it.

Visa states that after investing $11 billion in security from 2019 to 2024, their systems prevented $400 billion in fraud attempts. Stripe processes over $1.4 trillion in payments annually and uses this data to train its Radar fraud prevention system. During Black Friday and Cyber Monday in 2024, Radar blocked 20.9 million fraudulent transactions worth $9.17 billion.

Proxy transactions currently lack such fraud detection layers. When a proxy makes an x402 payment, there is no shared system to flag anomalous behavior, such as an unexpected surge in spending or unusual frequency.

Without persistent identity and reputation, each agent interaction starts from scratch. Reputation is deeply embedded in human commerce: the ads you see are based on browsing history, Uber drivers' acceptance is influenced by ratings, and your credit score follows you to every financial institution. This should be true for agents as well.

Post-Transaction: What If Something Goes Wrong?

Chargeback is a card network's way of handling disputes: when a customer disputes a transaction through the bank, funds are pulled back from the merchant. However, this process is often abused. In 2023, chargebacks cost merchants around $117.47 billion. For every $1 lost to chargeback, merchants typically incur an additional $3.75-4.61 in other costs (including fees, lost goods, and administrative expenses).

Source: Coinbase's x402 Paper

Merchants win only 8.1% of actively contested disputes. 84% of customers believe initiating a chargeback with their bank is easier than seeking a refund from the merchant.

Agent-initiated stablecoin transactions settle in seconds and are currently irreversible. Cloudflare has proposed a delayed settlement extension for x402, allowing for a "waiting period" to be set before final fund transfer.

Developers have been working on laying the groundwork for this infrastructure. At the ETHGlobal Buenos Aires hackathon, a team created Private-Escrow x402. Their escrow solution involves: buyers prepaying funds to a smart contract, signing a "payment intent" off-chain at the time of payment. A coordinator batches hundreds of such signatures into a single settlement transaction, reducing Gas fees by 28x.

However, this is just a foundational component that still needs to be productized.

Who Will Build All This?

This reminds me of the era dominated by telecom operators. They had billing relationships with every mobile user but missed out on the value created by smartphones. App distribution and mobile advertising generated billions of dollars in revenue, a domain that could have been captured by telcos.

Payment networks now find themselves in a similar position. Visa and Mastercard have spent decades building the trust infrastructure that the self-sovereign agent economy lacks. However, their business model relies entirely on interchange fees, which are predicated on their control of the payment rails. They invest heavily in maintaining this infrastructure, funded by a few percentage points of transaction volume. Offering consumer protection for stablecoin transactions would essentially mean subsidizing competitors' payment rails with their own revenue.

If payment networks don't act, the next contenders in line are OpenAI, Google, Anthropic, and other AI labs. They all want widespread adoption of their agents. But being centralized identity registries mean that they would be held accountable when agents misbehave. They wouldn't want to be the arbiters of "you booked the wrong hotel."

They would rather have a third party build the identity and recourse infrastructure for them to plug directly into, much like accessing payments or search engines today.

Cloudflare is in a unique position. They have handled massive amounts of web traffic, run bot detection, and have an "AI Audit" tool that allows publishers to track bot visits. Going from "identifying bots" to "verifying proxy identity and reputation" is not a huge technical leap.

But Cloudflare has always touted itself as neutral infrastructure. Once it starts issuing trust scores or arbitrating disputes, it becomes more like a regulatory body — a different business altogether and a different set of responsibilities.

Three Entry Points for Startups

You can't beat OpenAI on model quality, and you can't surpass Cloudflare on traffic. You need to find parts of the technology stack that their business models (at least for now) don't allow them to touch but still hold value. I think there are three entry points: identity, recourse, and attribution.

Proxy identity is the most direct. The registration model has been validated. While Plaid is a classic example, but very apt: they did identity verification for bank accounts. Startups can do the same for proxies: issue credentials, build reputation, and allow merchants to verify reputation before receiving payments. Its moat comes from network effects: once enough merchants verify through your registry, the proxy has to maintain a good reputation record.

Recourse mechanism is more challenging because it requires taking on risk. You can think of it as insurance: charge a small fee for each transaction, and bear the loss in case of issues. Scale is key. Card interchange rates range from 1.5% to 3%, which includes dispute resolution costs. Stablecoin channels cost much lower than this, so a recourse layer can very well offer comparable protection with a 0.5% fee and still have room for profit.

Attribution mechanism is the most forward-looking but will inevitably emerge. When proxies start influencing purchase decisions, brands will pay to impact recommended content. An auction mechanism can be designed. But it has a "cold start" problem, requiring brands, proxies, and merchants to participate in the market together to function, a challenge that the first two entry points do not have.

The importance of these three entry points varies with the development stage of the proxy economy:

· Identity becomes critical when proxies no longer require manual approval per transaction.

· Recourse becomes crucial when proxies start handling real funds.

· Attribution, on the other hand, will only kick in once the volume of transactions between proxies is sufficient to support an advertising market.

This leads to a practical development trajectory:

Source — Chart generated using Claude

Startup to Build Partial Agent Economy Infrastructure

The development of agents can be divided into three stages:

· As an interface for interaction

· Performing under human supervision

· Autonomous exchange with each other

We are currently in the first stage. An example is ChatGPT's Etsy checkout integration: we browse products in a chat interface (though not exclusive), the agent recommends options, but the final decision is made by a human. Trust is fully borrowed from existing infrastructure.

This stage favors existing giants as it is a distribution game to capture user entry points. The value accrues in the hands of players who own the purchasing decision interface.

The hallmark of the second stage is agents gaining more autonomy. Agents no longer just suggest itineraries but directly book flights, rental cars, hotels. We provide goals or constraints, agents execute, and we accept the results.

At this point, the trust layer becomes indispensable. Without recourse mechanisms, users won't authorize agents; without authentication, merchants won't accept agent payments.

This is where the opportunity lies for startups. Existing giants may lack the incentive to build trust infrastructure for stablecoin channels as they have significant room for growth in the current stage (still self-led). OpenAI generated $13 billion in revenue this year. In comparison, Tether made $10 billion in profit in just the first ten months of 2025, with higher annual profit expected.

An identity, recourse, attribution layer will be constructed by new companies dedicated to solving specific issues at the boundary of agent capability and user authorization.

The third stage is the autonomous agent economy. Your agent does not need approval for daily decisions; it can negotiate with other agents, bid on computational resources, participate in ad auctions, settle thousands of small transactions continuously. Stablecoins, due to their ability to handle the volume, speed, and granularity of machine-to-machine transactions, will become the default settlement layer.

The competitive focus in this stage will no longer be the best model or the fastest public chain but who has built the most trusted infrastructure: the agent's "passport," the "court" for dispute resolution, the "credit system" allowing overbalance transactions. These institutions serving software will determine which agents can participate in the economy under what conditions.

Conclusion

We have laid the pipes for agents to "spend" money, but we have not yet constructed the mechanisms to validate whether they "should spend" it. HTTP 402 has been dormant for thirty years, only to be revived by the feasibility of microtransactions. The technical challenges have been resolved. However, the trust infrastructure that underpins human commerce, such as identity verification, fraud detection, and dispute resolution, lacks equivalent agent-friendly versions. We have addressed the easy part. It will take time to enable agents to transact with confidence.

Original Article Link