Agentic Payment from Visa's Perspective

In the Agent era, what Visa truly wants to sell is not just card swiping, but the very fact that "this payment is legally authorized."

Written by: Will Awang

In July 2026, Visa and the on-chain data agency Artemis jointly released the report "Agentic Payments from the Ground Up," commissioned by Visa. Three months prior, Visa's Chief Product and Strategy Officer Jack Forestell publicly referred to the Agentic Web as the biggest growth opportunity he has seen in his over twenty years in the payments industry—immediately followed by the card organization funding a data agency to tell the market: machine payments are real, here is the data, and here are the gaps.

The nature of this document is more akin to a data portion of a strategic narrative rather than neutral research. The most valuable aspect of the report is that Artemis has systematically cleaned the on-chain data for x402 and MPP for the first time—about 90% of the on-chain surface flow consists of inflated numbers and tests; the most restrained part is that it lists issues related to responsibility allocation and regulatory qualifications, which Visa itself is involved in, and then simply concludes.

Duke Fintech previously addressed three questions in three studies that this report did not elaborate on: the distinction between substitution and creation in Agentic Commerce, the responsibility adjudication gap after Agent errors, and why the machine economy is inherently immune to local regulation. This article uses the Visa report as the main thread, and where it stops, we insert our judgments—everything introduced with "Duke Fintech believes" reflects our views.

Report Introduction

At its inception, Visa's goal was not merely to establish a credit card network but to build a global secure value exchange system. Today, Agentic Commerce is becoming a new frontier of this system—software programs are beginning to autonomously discover, evaluate, and procure goods and services, with payments transitioning from human-facing settlement tools to foundational capabilities embedded in machine workflows. In response to this demand, new protocols such as x402 and MPP are rapidly emerging, integrating credit card networks, stablecoins, and machine payment protocols.

1. What is Agentic Commerce

1.1 Two Types of Transactions

Agentic Commerce is a commercial form where software programs autonomously discover, evaluate, and execute transactions. It does not merely "automate" traditional e-commerce but allows software to become an economic entity capable of independently procuring, invoking resources, and making continuous payments for the first time. Based on current applications, transactions can be divided into two categories.

Macro Transactions: Agents act on behalf of humans or enterprises to complete traditional procurement, essentially executing the commercial intent of the human—such as an IT Agent automatically adjusting SaaS seats based on employee onboarding and offboarding, or a travel Agent booking flights and hotels according to budget and policy. These transactions involve higher amounts, and the merchants and payment channels are similar to traditional e-commerce.

The real new challenge is: authorization occurs before the transaction, execution happens under no supervision, and the system must prove who the Agent represents, what can be purchased, and where the budget boundaries lie.

Micro Transactions: Agents autonomously purchase data, computing power, models, and APIs to complete tasks—such as a research Agent paying a few cents to multiple data service providers to retrieve and cross-verify financial reports. These transactions occur frequently, at low prices, and instantly, with no accounts, subscriptions, or long-term contracts between buyers and sellers; the Agent discovers a service, evaluates it, pays, invokes, and leaves immediately. The requirements for millisecond-level responses, extremely low fees, and machine identities are the direct reasons for the emergence of new protocols like x402 and MPP.

Duke Fintech believes that the distinction between these two types of transactions can be further refined: Macro represents a substitution relationship—taking shares from existing human shopping behaviors; Micro represents a creation relationship—these transactions did not exist in the past, not due to a lack of demand, but because the decision-making cost for humans was higher than the transaction value. The Agent brings decision-making costs close to zero, allowing transactions to emerge spontaneously. This also means that all market forecasts measured by "the speed at which human shopping is taken over by Agents" only cover the first half.

As for how large the other half is, Circle CEO Jeremy Allaire provided a coordinate in "The Agentic Economy": a company is essentially "organized cognition with a logo," with labor costs accounting for one-quarter to half of revenue, and nearly all in knowledge-based companies—this cost is precisely what AI targets. When each function is refined into clearly defined skills, a skill that is clean enough to be arranged internally within a company is also clean enough to be discovered and hired externally—the open agent labor market does not need to be created; it is the overflow of millions of companies optimizing themselves.

By then, the minimum price unit will be compressed for the first time to a level finer than internal divisions within a company: a verification, an invocation, a translation can all be individually priced, transacted, and settled. In other words, Micro transactions measure not shopping, but the hiring of machine labor—using Allaire's words, models become cost items, and agents become businesses.

1.2 Why is it Exploding Now

The HTTP 402 "Payment Required" status code was written into internet standards as early as 1997 but has long remained unimplemented: traditional payments have high fixed costs, making small transactions commercially unviable, leading the internet to rely on advertising and subscriptions for monetization for a long time.

Duke Fintech believes that the high cost of settlement is only half the cause of death; the deeper half is the psychological transaction cost for humans— for a payment to be established, the cost of making a decision must be lower than the payment amount, and making a judgment of "is it worth it" for a few cents is itself worth several cents. Micro payments have not failed in the human economy; they are mathematically unfeasible; Agents have flipped this inequality.

The real turning point comes from AI Agents. The report anchors the crossing of capability thresholds around mid-2025—starting from models like Claude 4.5 and GPT Codex 5.2, Agents begin to autonomously discover unfamiliar APIs, evaluate prices, and decide whether to pay, without manual approval for each transaction. The demand for programmatic, on-demand purchases of data and computing power thus appears on a large scale for the first time. Meanwhile, dedicated chains like Tempo offer fees at the level of thousandths of a cent and finality in 500 milliseconds, while the gas costs of Base and Solana have dropped to a fraction of a cent.

However, the report emphasizes: dedicated chains have existed for many years and have never spontaneously given rise to a pay-per-use API market—it is the Agents that create the pull, and the infrastructure merely allows the demand to be commercialized.

The report also issues a blunt warning to traditional institutions: the risk of inaction for incumbents is disintermediation. If a few platforms simultaneously control the demand entry and transaction routing of Agents, they can pull payment traffic outside the existing network—especially in the micro-payment sector, where the card economic model is weakest. Incumbents participating only through large agent transactions will miss the part of the agentic economy with the highest transaction frequency.

2. Why Machine Transactions Need a New Payment Architecture

The difference between traditional e-commerce and micro machine transactions is not just smaller amounts, but a fundamental change in the entire transaction logic.

Subscriptions and pay-per-use billing can alleviate some issues but still require prior registration, negotiation, and establishing cooperative relationships, which cannot meet the Agent's temporary transaction model of "discovering interfaces—evaluating prices—immediate payment—completion of invocation." Therefore, the value of new protocols is not to replace credit cards but to add a layer of payment architecture suitable for machine interaction on top of the existing payment system.

Duke Fintech believes that the weight behind this table is much heavier than "smaller amounts, higher frequency": Agents are not faster consumers; they are a completely different species of consumption.

• Consumption Atomization—Humans bundle consumption into monthly subscriptions, annual fees, or one-time purchases, essentially exchanging prepayments for simplified decision-making. Agents do not have this need; they can precisely calculate the marginal value of each invocation;
• Decision Flow Streamlining—Human payments have a tolerance for waiting; filling out forms, entering passwords, and waiting for confirmations are acceptable. Agents complete hundreds of tool invocations in milliseconds, and each step may require payment;
• Dehumanization of Subjects—The most fundamental change is that the payer is no longer a human, while every design in the traditional payment system, from KYC to CVV verification to chargeback mechanisms, is based on the assumption that "users are humans." The first two are efficiency issues that engineering can solve; the third is an assumption issue that requires a complete rewrite of the entire system.

3. x402 and MPP: Architecture and Real Data

Before entering the two protocols, let’s first supplement a map previously drawn by Duke Fintech.

Today, Agents spend money through three different paths, and the essence of the differences lies in the responsibility residing in three different places:

• Wallets and Credits—Users recharge first and spend until the balance runs out; responsibility is prepaid, but the cost is being trapped in an ecological closed loop—this is the fastest and easiest path to implement, and most so-called "AI payments" are shopping assistants within the ecosystem;
• Card Tracks—Responsibility is borne by the network rules of Visa and Mastercard, with fifty years of contentious frameworks still in place; all the TAP, AP2, ACP, and UCP in Chapter Five are under construction on this track;
• On-chain Direct Payment—Wallet-to-wallet, with immediate finality; responsibility is either suspended or replaced by code and reputation.

Summary

The data section of Visa's report focuses on the two most important specimens on the third path. x402 will launch in May 2025, incubated by Coinbase and Cloudflare, and will be handed over to the Linux Foundation for open governance in April 2026; MPP will launch in March 2026, co-developed by Stripe and Tempo, with Visa participating in the protocol construction, currently undergoing the IETF standard process.

3.1 Common Foundation: HTTP 402

Both protocols are based on HTTP 402. The typical process is as follows:

1. The client requests server resources;
2. The server returns a 402 status code, providing the price, currency, and payment conditions;
3. The client evaluates the conditions and attaches a payment voucher;
4. The server verifies the payment and opens the resources.

Payments are directly embedded in HTTP requests and responses, eliminating the need to redirect to a settlement page and not requiring the buyer and seller to establish an account relationship in advance. The server only needs to return machine-readable payment requirements to become a paid interface.

3.2 Core Differences

x402 introduces intermediary service providers (facilitators) in transactions: after proxying payment funds, the intermediary verifies transaction conditions and completes the settlement. The intermediary layer undertakes payment verification, merchant access, risk control, and fund settlement, charging fees accordingly, making it easier for x402's industrial value to be deposited in intermediaries and settlement service providers.

MPP emphasizes direct settlement between buyers and sellers, with the underlying Tempo achieving approximately 500 milliseconds for finality. It supports two modes: immediate charging (Charge) allows for immediate payment delivery, with a single transaction atomically split among up to ten payees, and market revenue sharing completed directly at the protocol layer; session payment (Session) first deposits escrowed funds, signing off-chain vouchers based on consumption, with batch settlement and balance refunds after completion—suitable for ongoing services where total prices cannot be determined in advance, transforming "monthly subscriptions" into task-level, real-time, and terminable cost structures. By weakening intermediaries, MPP's long-term value is more likely to be deposited in upper-layer applications, aggregation platforms, and development tools.

Both types of protocols are still in their early stages, and while their functions may converge, their trust models and value deposition locations differ.

3.3 x402 Data: 11 Months Post-Launch

First, let's clarify the numbers mentioned in the report. The on-chain original data for x402 shows a cumulative flow of $135.7 million and 17.83 million transactions—but after filtering out wash trades and test transactions, only the following remain:

• $15 million corrected flow
• 109.6 million valid transactions
• 422,000 valid buyers
• Approximately 5,300 valid merchants

(Agentic Payment from Ground Up)

The real market is an order of magnitude smaller than the on-chain surface data. This is also why discussions about Agentic Commerce must first undergo a data cleansing process—Noah Levine, a partner at a16z crypto, pointed out a more fundamental issue: the 30-day transaction volume for the same x402 shows $24 million reported by x402.org, $3 million by Allium, and less than $2 million by Artemis, with a tenfold difference among the three data sources. When there is no consensus on how to measure this market, all growth narratives must be questioned—this report uses the strictest of the three measures.

November 2025 is the peak transaction volume, with approximately 38 million transactions and $5.15 million in corrected flow in a single month. By March 2026, transaction volume dropped to about 2.1 million transactions, with flow still at $1.64 million, and the average transaction price reached a new high since launch—x402 is no longer just a "few cents API payment" protocol, but is expanding into a broader range of amounts.

(Agentic Payment from Ground Up)

Duke Fintech has a different interpretation of this curve: the boundaries of x402 are defined by the potential disputes over goods. Surviving transactions focus on buying tokens, adjusting APIs, and renting computing power—goods that are consumed upon delivery and cannot be reversed; x402 has no authorization, no recourse, and no dispute channels, making "bare" a virtue for goods that are consumed immediately. If there is any possibility of error with the goods, the bare protocol cannot enter.

Concentration: The top 1% of buyers contribute about 90% of the flow, with only 0.02% of buyers contributing about 48%. A few professional institutions dominate most of the business, which is a typical characteristic of an early market.

Distribution: Base accounts for about 90% of valid transactions and 93% of corrected flow. In terms of transaction count, agents purchasing services from each other is the largest category; in terms of flow, a large number of uncategorized long-tail interfaces dominate—supply side is highly fragmented, and aggregation and discovery tools are still lacking.

(Agentic Payment from Ground Up)

3.4 MPP Data: First 33 Days Post-Launch

After adjustments, MPP has generated a cumulative flow of $25,000, approximately 115,000 valid transactions, over 2,800 valid buyers, and about 90 valid merchants, with an average of about 4,000 valid transactions per day. Immediate charging accounts for about 97% of transaction count, with an average price of about $0.17; the session mode has an average price of about $2.15, still primarily for developer testing.

The current ratio of buyers to sellers is about 32:1, with demand growing significantly faster than supply. The report mentions that the number of merchants in the first few months after x402's launch was also at a similar level, only to increase sharply afterward—Duke Fintech believes that this 32:1 ratio may not be a bad signal, but rather a time lag on the supply side.

4. Trust, Identity, and Responsibility

---

4.1 The Biggest Issue is Not Payment, But Trust

The payment industry has upgraded six generations of authorization technology over fifty years—from signature impressions, magnetic stripes with real-time authorization networks, EMV chip encryption, NFC, to Tokenization, and now today's Agentic Token—what has changed are the verification methods, but the verification has always been about the same thing: is the person at the transaction site the cardholder? Agentic Payment does not break the means, but the premise: the payer is no longer the person making real-time decisions, but a software program that holds authorization and executes autonomously; authorization and transaction are separated in time, with intent expressed in natural language, which cannot be verified by risk control engines. This brings four types of risks:

1. Mis-purchase: Agents misunderstand, choose the wrong service provider, or overpay. Humans can correct this before settlement, but uncontrolled agents can generate a large number of erroneous transactions in a short time;
2. Malicious attacks: Prompt injections can alter the execution logic of agents, inducing unauthorized purchases or fund transfers. The higher the autonomy, the greater the loss if compromised;
3. Unclear responsibility: Authorizers, agent platforms, model service providers, wallet service providers, and merchants may all claim they are not at fault;
4. Chain failures: When agents purchase services from other agents in large quantities, any failure in the chain can render previously completed payments worthless.

x402 and MPP can solve part of the payment and delivery issues, but cannot independently answer "who is responsible when an agent makes a mistake."

4.2 Current Solutions

Protocol layer: x402 verifies delivery through intermediaries before settlement; MPP controls risks through session escrow; flexible limits and spending caps restrict single losses.

Identity layer: Many agents still rely on API keys and wallet addresses for authentication, which can only prove access rights but cannot demonstrate historical performance capability. Even Visa itself does not shy away from this—Forestell admitted in a public dialogue in March 2026 that agentic transactions would be riskier than e-commerce and mobile payments back in the day: there is an additional agent in the middle, which needs identity, needs to be protected and verified, and requires more data to ensure security. The industry is addressing this from two directions—cross-chain identity and reputation standards in the crypto ecosystem, and the existing authorization, KYC, and anti-fraud capabilities of the card system.

Strategy layer: Tools like Coinbase Agent wallet, Turnkey, Privy, and Safe provide whitelists, spending limits, and merchant category restrictions; the card system binds limits and merchant restrictions to agent identities through tokenized vouchers.

4.3 Unresolved Issues

The industry still lacks three foundational infrastructures: a universal agent identity standard across crypto and card ecosystems; dispute, refund, and evidence mechanisms suitable for high-frequency micro-payments; and a clear division of responsibilities among authorizers, agent platforms, model service providers, and merchants.

These gaps are both the main risks of Agentic Commerce and the areas where traditional payment institutions have the greatest capacity to fill.

Duke Fintech believes that behind these three gaps lies the same structural shortfall. Delegated payments legally cannot avoid four questions: whether authorization is valid, where the boundaries are, who is responsible for crossing the line, and who ultimately bears the loss—these four questions have been the backbone of agency law for two thousand years. The first two questions are the territory of proof, whether authorization exists and whether boundaries are clear, which cryptography can answer; the latter two questions are the territory of adjudication, who is responsible for crossing the line and who bears the loss, which only rules can answer.

Yet today, almost all players are crowded into the first two questions: signature chains, audit trails, and cryptographic proofs of performance, nailing down "what happened" to a cryptographic level, while saying nothing about "who should be accountable for what happened." The warning from the dispute resolution company Chargebacks911 speaks to this issue:

The industry is building from the wrong end------the entire front-end authorization framework is online, but the infrastructure for post-transaction disputes, how to classify disputes, how to allocate responsibilities, and how to accept evidence has hardly been touched.

In the empty space of adjudication, the first thing to rise is not the law, but the price list: the conditions for liability transfer by card organizations, the compensation thresholds of insurance, and the protection clauses for registered agents by Amex are all pricing the question of "who is responsible for the mistakes"------they anchor the same variable. The precision of the authorization boundary is the price of responsibility.

Industry Standards and Payment Network Layout

Agentic Commerce requires multi-layer protocols for communication, identity, authorization, payment execution, and clearing, and currently, there is no single standard covering the entire chain.

Visa TAP allows merchants to confirm that visitors are authorized agents through encrypted signature HTTP messages. It does not require merchants to reconstruct their payment systems but adds a machine identity layer on top of the existing web and card systems------this is the most direct path for traditional payment networks to enter Agentic Commerce.

AP2 records procurement intentions, specific shopping carts, and final payment authorizations through encrypted authorization instructions, transforming natural language intentions into verifiable and auditable structured authorizations, while remaining neutral towards channels like credit cards, transfers, and stablecoins. Duku Fintech believes its ceiling is here: budgets, slots, and time windows can all enter the fields, but the reasons you haven't articulated cannot------the authorization record captures what you said but cannot record why you said it.

ACP and UCP: ACP is driven by Stripe and OpenAI, adopting a closed market model where the agent platform and merchants mutually pre-approve; UCP is driven by Google and Shopify, allowing agents to complete purchases directly within search and Gemini, while merchants retain customer relationships. Both are essentially competing for the agent's commercial entry------whoever controls the process of discovering products, confirming transactions, and executing payments may grasp the next generation of e-commerce traffic distribution.

Visa Intelligent Commerce aims to become a unified entry point: a single integration that is compatible with protocols like TAP, x402, MPP, ACP, UCP, and provides tokenization, identity verification, and programmable spending control. This reflects that the strategy of card organizations is not to guard card transactions but to become the intermediary hub connecting agents, merchants, and various settlement channels------at the Visa Payments Forum in June 2026, Visa even demonstrated a conceptual product allowing agents to pay directly with tokenized credentials in the command line, with Forestell stating the goal is to "make cards the best payment method in the command line": card organizations have already extended their reach into the realm of machine-native transactions.

The long-term trend is integration rather than replacement. MPP supports both stablecoins and credit cards; Visa has launched card specifications compatible with MPP; Stripe has integrated x402 stablecoin payments on Base. In the long run, large-scale agent consumption will still rely on the mature merchant network and dispute system of credit cards, while machine-native micropayments are more suitable for low-cost on-chain settlements, with both types of channels coexisting in the same workflow------this coexistence is the structural division of replacement and creation discussed in the first chapter: the card network has not lost any existing categories, and new categories are not within its range.

Both sides' public statements align with this. Forestell provided the shortest version of positioning at the Visa Payments Forum: AI transforms the front end of business, stablecoins reshape the back end of funds, and Visa's role is to ensure both ends operate safely, reliably, and globally; the report's author, Tim Conard, Visa's on-chain data head, summarized it succinctly in an official accompanying article: what Visa aims to do is not force the ecosystem to choose sides or require a new infrastructure.

The opposing camp has given a mirrored statement------Circle CEO Jeremy Allaire wrote in "The Agentic Economy" that the on-chain environment does not need to defeat card networks in debate; it just needs to be a place where the agent economy can operate natively. He even acknowledged that "stablecoins only need to provide funding, while processing and fees remain with Visa and others" is a realistically possible equilibrium. Both sides are leaving room for each other.

Duku Fintech believes that reading these statements and reports together clarifies the sequence: Forestell's positioning is the conclusion, and this report is the argument supporting that conclusion------strategy comes first, followed by data.

But beyond the statements, there is a more honest document: investment records. Visa Ventures invested in Nekuda, which generates authorization documents, and Payman, which handles bank-side approval breakpoints------the former's authorization is born at the moment of user-agent dialogue, outside VisaNet; the latter's funding flow does not go through card tracks at all. One investment buys upstream, and another buys outside the track, indicating that Visa is clear that its agreements only cover the segment "when money passes through me"------looking at an institution's agreements reveals what it wants you to believe; looking at its investments shows what it does not believe.

Putting together the command line concept, investment records, and the statement of "not forcing sides to choose," Visa's substantive actions become apparent.

Duku Fintech believes that Visa is doing something that is easily misunderstood:

It is not competing with stablecoins; rather, it is quietly repositioning itself from a settlement network to an authorization infrastructure------VIC manages tokenization, TAP manages identity, verifiable records manage audits, and MPP card specifications manage multi-track access, with four lines betting simultaneously, all covering the links of the authorization chain. In the Agent era, what Visa truly wants to sell is not card swiping, but the fact that "this payment is legally authorized" itself. The card network may not win the settlement of the machine economy, but it may win the authorization of the machine economy.

Application Scenarios

The original report's scenario chapter mainly consists of an eight-industry list and three model subsections. Section 6.1's industry expansion is an extended analysis by Duku Fintech based on the report framework and does not represent the statements of Visa or Artemis; sections 6.2--6.4 are based on the original report.

6.1 Eight Industry Scenarios

The most valuable part of Agentic Commerce is not the protocols themselves, but the transactions that the protocols enable, which were previously impossible to establish. These can be categorized into eight types by industry:

1. Corporate procurement and back-end operations: IT, administration, finance, travel, and procurement agents complete repetitive procurement payments within established policies and budgets, relying more on mature credit card and expense management systems;
2. Sales and customer data: Sales agents purchase corporate information, email verification, and CRM completion for individual potential customers, comparing multiple offers to use the best------turning fixed subscriptions into task-level variable costs;
3. Research and financial data: Purchasing financial reports, on-chain data, and regulatory information around the same issue, with multi-source cross-validation replacing single data vendor binding, especially suitable for high-value, low-frequency, long-tail data;
4. Content creation: Marketing agents pay ten image service providers 5 cents each, only using the best output------embedding payments into the production and selection process;
5. Engineering and computing power: Code agents temporarily procure GPUs, inference, testing, and sandbox environments, exiting once tasks are completed without needing to open accounts or sign contracts in advance;
6. Security and compliance: Calling KYC, sanction screening, on-chain risk analysis, and deep forgery detection on a per-instance basis. This is the highest value and most sensitive------the legality of data sources and the responsibility for results must be resolved simultaneously;
7. Public opinion and monitoring: Continuously purchasing social signals and competitive dynamics based on data volume, duration, or trigger counts, more suitable for temporary projects than fixed monthly fees;
8. Maps and logistics: Continuously purchasing route, inventory, and price data------paying first for decision-making capabilities, then deciding whether to purchase real-world goods.

Duku Fintech believes that these eight scenarios will not arrive simultaneously; the order of implementation is determined not by technological maturity but by the risk-bearing party: scenarios backed by agreements (API, data, computing power) run the fastest------immediate delivery, clear billing, and no refund disputes, with almost all real transactions on x402 and MPP concentrated here; scenarios backed by enterprises (B2B procurement, internal standardized spending) follow as the second batch------budget pools and approval chains existed in manual form before agents appeared; consumer retail backed by individuals is the slowest because it requires a complete consumer protection framework. The most visually impactful scenarios will land the latest.

6.2 Streaming Pay-per-Use

Not all services can be priced in advance: continuous hours of computing power, minute-based transcription, and continuously received market data cannot report a total price before the transaction. MPP's conversational payment is designed for this------merchants earn income based on actual delivery, and buyers do not need to commit to fixed subscriptions for fluctuating demands.

6.3 Agent Purchasing from Agent

Agent mutual services are the most representative scenarios of Agentic Commerce and have already become the largest category counted by transaction volume on x402. An agent does not need to possess all capabilities but can outsource sub-tasks to other specialized agents------the main agent purchases searches from retrieval agents, who then purchase reports from data agents, with analysis, compliance, and reporting each being outsourced layer by layer. A single user request thus forms a multi-layered machine supply chain.

Duku Fintech believes that payments here are both a settlement tool and a coordination mechanism for resource allocation and professional division among agents. This also explains why the number of transactions grows much faster than the growth of consumption amounts: the total cost of a single task may only be a few dollars, but it is broken down into dozens or even hundreds of machine payments in the background.

6.4 Long-tail Services and "API as Merchant"

Machine payments have significantly lowered the barriers for digital services to become merchants: developers no longer need to build complete SaaS products, but can simply package their capabilities into machine-readable payment interfaces. Many services that were previously unable to be commercially independent have now emerged—such as single document conversion, corporate registration inquiries in certain countries, authenticity checks for individual images, and risk scoring for specific addresses. This model is already being productized: Merit Systems' AgentCash packages wallet management, merchant discovery, and over 420 paid APIs into a single integration, connecting to Agent frameworks like Claude Code.

However, the low barrier to interface deployment also means that similar services can be easily replicated—individual long-tail merchants find it difficult to establish barriers. The real barriers will be discussed in the conclusion.

7. Compliance and Regulation

7.1 How to Qualify Protocol Roles

The x402 intermediary may temporarily hold funds, verify deliveries, and transfer payments, and may be classified as funds transmission, payment processing, or custodial services in different jurisdictions. The absence of a unified intermediary in MPP brings another type of problem: without a clearly defined intermediary, who bears the obligations of KYC, anti-money laundering, transaction monitoring, and dispute resolution?

Cross-border transactions further amplify uncertainty: agents, service providers, wallets, and infrastructure may be distributed across multiple countries, while existing rules are based on clear transaction subjects, defined jurisdictions, and lower transaction frequencies.

Circle CEO Jeremy Allaire provides a more thorough statement in "The Agentic Economy": the three layers of this economy—currency, contracts, and the agents doing the work—are all software running on the internet, with no inherent geographical attributes; the work executed in the cloud has "no hometown"; globality is not an added feature of the system, but a structural attribute of its constituent materials.

Duke Fintech believes that following this statement, the failure of fiat currency tracks here is based on two premises: agents lack subject qualifications, which is a human dimension; and actions have no place of occurrence, which is a spatial dimension—this is not a jurisdictional vacuum, but an overabundance of jurisdiction, where an action may simultaneously fall under multiple legal domains of the consumer's location, the data subject's location, and the market's location, yet there is no place of occurrence to make the final decision. For payment institutions, compliance frameworks may shift from backend functions to sources of competitive differentiation.

7.2 Stablecoin Regulation

The x402 heavily relies on USDC, and MPP also considers on-chain settlement as an important component, thus stablecoin regulation directly affects the expansion speed of Agentic Commerce. In the United States, the GENIUS Act has established a federal licensing and reserve framework for payment stablecoins; the regulatory status is clear, and stablecoins are a channel that banks and acquiring institutions can confidently build upon. If other jurisdictions remain fragmented for a long time, the costs of cross-border implementation will significantly increase.

Allaire also provides a breakdown that explains why the U.S. is willing to adopt this framework: the channel through which money flows and the money that flows through the channel are two different things—the protocol layer is neutral and stateless, while the money that flows above it is still denominated in sovereign currency and issued under the laws of that country.

Duke Fintech believes that following this breakdown, stablecoins are not the opposite of fiat currency; they are fiat currency itself; what is being replaced is not the currency, but the set of national banking channels that weld currency within national borders. The way stablecoins win is not by defeating fiat currency, but by liberating fiat currency from national channels—this is almost a plan for the globalization of the dollar that does not require a CBDC for Washington. The GENIUS Act is less about accepting something new and more about signing off on this plan.

7.3 Exclusive Challenges of AI Agents

Current KYC and anti-money laundering rules are centered around human clients. Agentic Commerce must re-answer: who is the verification subject—authorizers, agent platforms, wallets, or model service providers; who is responsible for monitoring the continuous transactions of agents; who bears the obligations for reporting suspicious transactions and tax obligations; how thousands of cross-border micropayments can be audited and tracked.

Therefore, an infrastructure for agent transactions that is traceable, auditable, and capable of bulk reporting will become a new commercial value point.

And the vacuum is not abstract. The United States still lacks specific regulations for agent transactions, and the binary framework of "authorized or not authorized" from Reg E / Reg Z is being rigidly applied in gray areas—the act of contracting itself is not the problem; the ESIGN Act recognized contracts concluded by electronic agents back in 2000, and the vacuum exists after the contract is formed: the distribution of liability for agent overreach lacks specific rules and precedents. The EU's AI Liability Directive (AILD) was officially withdrawn on the eve of large-scale commercial use of agents in October 2025, citing the inability of the 27 member states to reach a consensus; the strong customer authentication in PSD3 is designed for individuals holding mobile phones, while agents do not have fingerprints.

The gaps are filled with private orders: currently, the closest globally consistent rules that can be directly executed in the payment chain are still the card organization rules—its enforcement power does not come from courts, but from "I can refuse this transaction"; Amex allocates liability through commercial commitments, and institutions like Munich Re have begun selling AI liability insurance.

Duke Fintech believes that the liability rules for agent payments are being written by protocol specifications, card organization rules, and policy terms—legislation is occurring outside of the law.

8. Future Opportunities

For financial institutions, the most realistic path into machine payments is hybrid settlement: retaining bank cards and existing merchant experiences at the front end, while using stablecoins for 24/7 settlement, faster returns, and lower cross-border costs at the back end. Moving up a layer, x402, MPP, AP2, TAP, and ACP are currently independent, but in the future, agents will need to automatically select authorization, identity, and settlement methods within a single workflow—platforms capable of cross-protocol identification and routing will become the key control layer.

Cross-border transactions will be the default attribute of machine transactions rather than an edge case, but the foreign exchange, compliance, and cross-border settlement that adapt to machine speed have not yet matured, which is the clearest near-term opportunity for global payment networks.

The supply side will also change: currently, agents are mainly purchasers, but as deployment costs decrease, more agents will directly provide data, models, and professional services to other agents, leading to rapid expansion on the supply side. However, Duke Fintech believes that this prediction has a legal premise that a report has not pointed out—the lack of subject qualifications mentioned in Chapter 7 is most evident here: the foundation of the fiat account system is that "account holders must be natural persons or legal entities"; agents in the banking system are always just a set of API permissions on someone else's account, able to spend money for others, but cannot have money, receive payments, or pledge themselves. Spending money for others is a payment issue, which is sufficient for adaptation; having money oneself is a subject qualification issue, and the account structure of fiat currency fundamentally cannot provide that. For agents to become merchants, they must either wait for the law to provide subject qualifications or move to on-chain wallets—this is also why the supply-side expansion of "agents becoming merchants" is likely to occur first on-chain.

9. Conclusion

Agentic Commerce has already moved from concept to real transaction stage: x402 and MPP are beginning to carry actual transaction flows, with per-query, streaming payments, and agent mutual services forming. New protocols have solved the issues of low cost, machine speed, and programmable payments, but to enter the formal financial system, it still relies on the long-accumulated capabilities of traditional payment institutions in identity, risk control, compliance, and dispute resolution.

Returning to the map of three paths at the beginning of Chapter 3, we can now outline their respective boundaries. Duke Fintech believes that the three paths are not competing for the same piece of cake, but are drawing boundaries along the density of disputes: zero disputes go to chains and stablecoins, complex disputes go to cards, and disputes that are avoided go to wallets. The real battlefield is at the intersection of these boundaries—agents first compare prices with credits, then place orders with cards, and settle with stablecoins to suppliers, with the three paths converging in the same transaction.

Due to the ease of replication on the supply side, long-term value will not remain in individual transactions or single merchants, but will concentrate on three layers: traffic distribution (who controls the entry for discovering agents and routing transactions), identity and trust (who can verify agents, merchants, and authorization chains), and underlying clearing (who can complete global settlements at low cost, securely, and in compliance).

Traditional institutions have decades of accumulation in the latter two layers, while new protocols win on speed, cost, and programmability—the future leaders are likely to be platforms that combine both types of capabilities. The end point of competition is not in individual transaction fees, but in who can master agent identities, transaction entry points, cross-protocol routing, and global compliance infrastructure.