Scenarios, contradictions, and the ultimate outcome as seen by 16 AI payment practitioners

AI payment is no longer just a concept. x402, MPP, Tempo, AP2—over the past year, Coinbase, Stripe, Google, and Visa have built protocol frameworks at different levels. Real on-chain data, real merchant integration, and real model misinterpretations are also beginning to emerge. Last Saturday, Zhiwubuyan organized a closed-door meeting on Agent Payment. Sixteen guests from payment infrastructure, wallet services, major companies' payment businesses, and investment institutions answered four questions in three hours: Where exactly does AI payment occur? How can AI spend money securely? How does this business make money? Where will the competition between major companies and startups lead? The following are the core judgments that emerged from this discussion: The most mature scenario for Agent payments is API calls, with single transactions of $0.01 relying on frequency to sustain volume; There is a fundamental conflict between the uncertainty of AI output and the certainty requirements of the financial industry; this is the most fundamental technical contradiction of Agent payments; The security framework of Agent payments is shifting from identity verification to intent verification; The chargeback mechanism in Agent payments... In scenarios where this fails, three-layer arbitration will become a new paradigm for payment security; The design philosophy of large companies is to distrust agents and only trust transactions; The real bottleneck in agent-based payments is not the payment itself, but the upstream transaction process, which has not yet been rebuilt for agents; Startups play the role of component suppliers for large companies, not end-user service providers.HostHazel HuHost of the podcast "Zhi Wu Bu Yan", core contributor to the Chinese Public Goods Foundation GCC, X: withhazelhu; Jike: A careless YueyueIvy ZengHost of the podcast "Zhi Wu Bu Yan", exploring real-world use cases of Agentic Payment, focusing on Fintech growth, previously worked in post-investment management at a VC firm, and was responsible for regional growth of 2C products at a new type of bank. X: IvyLeanInThomas ZhengHead of Capital Markets at Zhi Wubuyan, with 6+ years of experience as a primary market financing advisor, serving multiple leading projects in the industry and facilitating win-win cooperation within the industryReal-world scenario—Agent payments are already happening, but the form is different from expectationsAPI calls are currently the most mature on-chain scenario for Agent paymentsThomas ZhengAnalysis of on-chain data from ClawRouter (an application that uses USDC payment to pay LLM APIs) shows that API call scenarios exhibit high-frequency, small-amount characteristics: as of early April 2026, approximately 1,400 unique addresses generated 530,000 transactions, totaling approximately $28,000. Considering the platform also offers a free model, actual usage is likely underestimated—the free portion generates approximately 1 million API calls per month. Image: ClawRouter Official WebsiteImage: ClawRouter Official WebsiteData from a payment infrastructure startup also shows that since starting to deploy the Agentic Payment native payment layer last September, API calls have accounted for about half of the total. Credit limit authorization is the basic authorization model for Agent payments. The unexpected success of the A2A (Agent 2 Agent) bonus campaign has driven the innovation and popularization of authorization mechanisms. The core of this authorization model is the credit limit rather than approval: users pre-authorize a credit limit to AI, and within that limit, AI can autonomously access funds without requiring confirmation for each transaction. "Within this limit, AI can access your money without your confirmation." Offline consumption hasn't fully taken off yet; the problem isn't payment, it's the experience. Exploration in the online-offline settlement field has covered 50 million real merchants, with scenarios including booking airline tickets, topping up mobile phone credit, and buying gift cards. However, C-end consumption scenarios still face the dual challenges of cultivating user habits and leaping forward in user experience. Experts and KOLs' Distillation Agents Already Have Mature Business Models. Successful cases have validated this approach: renowned doctors and KOLs distill their professional knowledge and content into agents, which users can use when they cannot meet in person. For example, a self-media practitioner distilled their past content into an app, charging 199 yuan per month, and sales have been excellent—a 15-minute live stream with them costs several thousand or even tens of thousands of yuan, while using their agent version only costs tens to hundreds of yuan. Image: Self-media practitioners distill their past content into apps. Transaction Agents Find PMFs Faster Than Payment Agents. Data in the Crypto field shows that transaction scenarios are currently the concentration of real user demand, and their business model naturally possesses a commission-based attribute. Analogous to the early history of blockchain development, those scenario builders who preemptively deployed merchants and stablecoins when gas fees were high, such as Tron, found it difficult for users to migrate even after gas fees increased. The phenomenon of hundreds of millions of users using Qianwendian milk tea during the Spring Festival sparked discussion: were users using it because of a better experience or because of the 25 yuan subsidy per order? The information density of dialogue-based communication is limited; future C-to-B scenarios may require seamless dialogue through smart glasses, demanding a leap in user experience. Attendees listed scenarios that better address user pain points: Procurement scenarios: Scenarios with strict budget control requiring comparison of multiple suppliers (e.g., Alibaba's AI e-commerce agent - Accio) Complex tasks: Scenarios requiring multi-step coordination, such as wedding planning and travel booking Ticket-buying scenarios: High-time-demand needs such as concert tickets Image: Alibaba's AI E-commerce Agent - Accio From a traffic acquisition perspective, Agent payment is similar to early SEO and short videos—representing new traffic opportunities. Those who initially researched SEO, though starting from humble beginnings, consistently found ways to acquire early traffic from SEO. The significance of the "Jinguyuan Dumpling Restaurant" incident may be comparable to the early days of buying pizza with Bitcoin, which will still be remembered by people many years later. Jinguyuan Dumpling Restaurant Skill Background Story: "On April 7, 2026, amidst the popularity of OpenClaw, the owner of the dumpling restaurant developed an AI capability module called 'Jinguyuan Dumpling Restaurant SKILL'. This AI skill is designed for AI Agents rather than directly for humans. After installation, the AI assistant can autonomously query menu information, business hours, queuing rules, and even obtain a number online. Due to excessive queues during the Winter Solstice of 2025, the food delivery platform's server mistakenly flagged the store's interface as abnormal and blocked it. The owner hopes to optimize the future queuing experience through AI."Image: Meituan queue skill at Jinguyuan Dumpling RestaurantTrue Agent Payments Haven't Started YetFrom a macro perspective, discussing true Agentic Payments may indeed be premature. We can use a child's development as an analogy: currently, they are like children aged 1 to 5, their income comes from their parents, their disposable income is authorized by their parents, what they buy is decided by their parents, and they haven't yet developed their own intentions. The consensus among participants is that genuine Agent payments are currently concentrated in productivity scenarios: API calls: Large models or API purchases are needed to improve productivity. Enterprise scenarios: Agents in procurement and finance teams for improving enterprise productivity. Vibe Coding: Scenarios for rapid development of demos or products. Identity and Authorization—The Uncertainty of AI vs. Financial CertaintyAgent payment security requires a four-layer framework: identity, risk control, compliance, and arbitrationPayment security can be broken down into three dimensions: identity, risk control, and compliance. AI payments should also follow this framework, with arbitration added as the fourth layer of protection.I. Identity Layer: Identity Verification is Shifting to Intent VerificationIssuing ID cards to agents, establishing a credit scoring system (building a 5-dimensional scoring standard based on the agent's professionalism, adoption rate, effectiveness, token price, etc.), and completing identity verification. Establishing a traceable and verifiable decentralized DID identity system through blockchain. Building on this, traditional payment authentication is shifting towards intent verification in agent scenarios. Intent verification needs to consider whether the agent's payment is reasonable, whether the behavioral process meets the requirements, whether it satisfies the final intent, and whether it complies with compliance requirements. II. Risk Control Layer: The Uncertainty of AI and the Certainty of Finance are Fundamentally ConflictingThere is a fundamental contradiction here: the uncertainty of AI output conflicts with the financial industry's high requirements for certainty and the cost of trial and error. In real-world scenarios:- Errors in amount recognition have been exposed (0.01 USDC may be read as 10,000 USDC). The editor also encountered the same situation; AI easily misreads the amount of USDC, reading 0.1 USDC as 10,000 USDC. This is because the field returns a raw integer, but USDC supports 6-digit precision (USDC microunits), so the displayed result is multiplied by 1,000,000. - This is an issue where consumers are easily misled (food delivery descriptions claiming "it can cure all diseases," leading many consumers to place orders). Image: AI mistakenly reads 0.1 USDC as 10,000 USDCMeanwhile, developing supply chain poisoning strategies presents a new challenge for risk control. Since the explosive popularity of OpenAI, for example, poisoning a certain npm package, users may not directly use the package, but the packages they depend on use it. Risk control needs to cover multiple layers such as identity authorization layer (anti-money laundering), model side (drift, illusion), and execution chain (poisoning attack). The design philosophy of major tech companies is to assume all agents are malicious by default. They don't pursue "verifiable agents," but rather "verifiable transaction chains." By introducing a mandate protocol, tasks are broken down, constraints are set, and cross-validation is performed. The anti-fraud architecture encompasses layered data, zero-knowledge proofs, the zero-trust principle, and self-verification mechanisms. III. Compliance Layer: The Semi-Decentralized Lightning Network is a Better Solution for Micropayments Traditional finance and blockchain both face bottlenecks when handling massive concurrency. When designing an agent, it should first be defined as a micropayment. The security of micropayments can be designed in a way that is neither too centralized nor too decentralized. The Lightning Network, dormant for many years, may be reborn in the era of Agentic Payments due to its extremely high theoretical TPS. IV. Arbitration Layer: A Layered Arbitration Mechanism Will Replace the Traditional Chargeback Mechanism The traditional credit card chargeback mechanism in the Visa network is difficult to implement in Agentic Payments, necessitating the establishment of a new layered arbitration mechanism: Layer 1: AI automatically arbitrates clear disputes (double charges, incorrect amounts, undelivered services) Layer 2: AI arbitration panel handles the parts requiring judgment (service quality, authorization boundaries) The third layer: Human participation in arbitrating complex disputes. Business models – securing niches, repricing AI, risk control, and licensing. Startups are currently securing niches "out of passion." Before a business model is proven successful, entrepreneurs honestly say, "Out of passion, securing a position, waiting for the right opportunity"—this is how one API platform entrepreneur describes the current stage. The transaction scenario inherently possesses a take-profit attribute. Analogous to the early development of blockchain, those scenario builders who preemptively deployed merchants and stablecoins when gas fees were high, such as Tron, found it difficult for users to migrate even after fees increased. The business model of the trading scenario in the crypto industry inherently possesses a take-profit attribute. Bill aggregation is key to solving the uneconomical nature of small-amount payments. If using a card for payment, merchants may lose money on transactions under $10. In the Agentic Payment scenario, small payments are common; the solution is to perform bill aggregation to increase the single settlement amount. **Pricing by results is only suitable for quantifiable piecework** Users may only call one API, but the results can vary greatly. How should AI services be priced? Participants believed that pricing by results is only feasible for simple piecework (such as the number of work orders a customer service agent resolves); in uncertain scenarios (such as the quality of leads acquired by a sales agent), it is highly subjective. Pricing by results only holds true for a small number of piecework tasks. Mainstream scenarios will likely remain within the old logic of pay-per-call/subscription until the verifiability of agent output is proven. Pricing your AI product: Lessons from 400+ companies and 50 unicorns | Madhavan RamanujamThe key to Vibe Coding's commercialization lies in subscription and usage conversionThe goal is to enable new AI companies or ordinary developers to quickly commercialize products developed through Vibe Coding. Many independent developers find it easy to create product demos, but developing a viable business model is much more difficult. The key lies in how to convert the cost of each user's large-scale use into a monthly subscription plan, or a subscription plus credit model. Competitive Landscape – The Offensive of Large Companies and the Strategies of Startups Stablecoins are Combating Traditional Card Organizations Before acquiring stablecoin company Bridge, Stripe's valuation fell from a high of $92 billion to below $70 billion. After the acquisition, its valuation quickly returned to the $90 billion range, with its latest funding round priced at $159.1 billion. Its stablecoin back-end clearing and settlement service is priced at 1.5%, far lower than the average rate of 2.8% to 3% for traditional card organizations, and may even drop to 1% in the future. In contrast, the business models of traditional payment companies are very fragile (such as Visa's heavy reliance on transaction fees), while PayPal and others, fearing impact on their core businesses, have hesitated in their stablecoin initiatives and failed to achieve significant scale. For a long time to come, the business model may not involve ordinary end-users using these tools themselves, but rather large companies providing unified encapsulation. Large companies may become customers, and entrepreneurs may become suppliers, piecing together the developed tools and then selling them at higher prices. This trend inevitably increases the degree of centralization in the industry. AI tax is the inevitable form of high-frequency, low-value payments within 3-5 years. Some participants believe that AI taxation will serve as a source of UBI (Universal Basic Income) and unemployment benefits, and high-frequency, low-value AI payments will become the underlying infrastructure. Possible tax collection methods include: Introducing the concept of "AI penetration rate" and collecting taxes progressively based on AI penetration rate; Collecting taxes based on token usage, using it as a tax base similar to value-added tax; The real bottleneck isn't in payment, but upstream—the transaction process hasn't yet rebuilt the Agent system. Payment issues seem solvable through protocols and user wallets. However, the biggest problem is that transactions cannot be completed. All payments require a prior transaction; scenarios like e-commerce or airline ticket purchases cannot be completed through an Agent. Without a transaction Agent, subsequent payments cannot proceed. Breaking Through to the Consumer Market: The Importance of Ground Promotion and the Boundaries of Startups Why is OpenClaw suddenly so popular? In China, it was built on ground promotion, driven by large companies selling cloud services and engaging in such promotions. Just like the early promotion of mobile payments, a key reason why even the elderly could use it was the subsidies offered by ground promotion teams—"You install the app, I'll teach you how to use it, and I'll really give you 50 yuan." However, for startups, many needs may take a long time to be met. An AI payment infrastructure entrepreneur stated that after assessing this, they initially decided not to look for user scenarios. They believe that the cost of user education should not be borne by one or two startups, but by the entire industry. If the industry itself is not viable, then it's meaningless; if it is viable, the costs should be spread out by large companies, who can then enjoy the benefits. Conversely, they focus on abstraction—abstracting all accounts, wallets, and even bridges, chains, and payment networks in the industry, so users don't need to understand them. Once they understand this, they know where the small team's winning edge lies and which costs to avoid. This is perhaps the question all Agent payment participants need to answer at this stage: not "Will Agent payments succeed?", but "Before it succeeds, at which layer are you prepared to stand?" Protocol layer, wallet layer, identity layer, authorization layer, transaction layer, clearing and settlement layer—at each layer, there are bettors and waiting players. Large companies are preparing to swallow the entire chain, while startups are preparing to be pieced together. Those who survive are likely those who neither overestimate their ability to independently support a sector nor underestimate their value at any particular layer.