Scenes, Contradictions, and the Endgame in the Perspective of 16 AI Payment Industry Practitioners

AI payments are 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 integrations, and genuine model misreads have also begun to appear one after another.

Last Saturday, the “No Words” organization hosted a closed-door Agent Payment meeting, with 16 guests from payment infrastructure, wallet services, large tech payment businesses, and investment institutions. In three hours, they answered four questions: Where exactly is AI payment happening, how to make AI spend money safely, how to profit from this business, and where will the game between big companies and startups go.

Below are the core judgments that emerged from this discussion:

• The most mature scenario for Agent Payment is API calls, with high frequency and small amounts supporting volume with just $0.01 per transaction;

• There is a fundamental conflict between the uncertainty of AI outputs and the certainty required by the financial industry—this is the underlying technical contradiction of Agent Payment;

• Chargeback mechanisms fail in the Agent scenario; a three-layer arbitration system will become the new paradigm for payment security;

• Big tech companies’ design philosophy is distrust of Agents, only trusting transactions;

• The real bottleneck for Agent Payment isn’t the payment itself, but the upstream transaction process that has yet to be rebuilt for Agents;

• The role of startups is to be component suppliers for big companies, not C-end service providers.

Host

Hazel Hu

Host of the podcast “No Words,” core contributor to the Chinese Public Goods Fund GCC, X: withhazelhu; also known as Yuyue, who doesn’t take things too seriously.

Ivy Zeng

Host of “No Words,” exploring practical use cases of Agentic Payment, focusing on Fintech growth, previously in VC post-investment, and responsible for regional growth of B2C products at a new bank. X: IvyLeanIn

Thomas Zheng

Head of Capital Markets at “No Words,” with over 6 years of experience as a primary market financing advisor, serving multiple top projects in the industry, helping connect and foster cooperation.

Real-world scenario—Agent Payment is happening, but in unexpected forms

API calls are currently the most mature on-chain scenario for Agent Payment

Analysis of on-chain data from ClawRouter (, an application using USDC payments to pay for LLM API), shows that API call scenarios feature high frequency and small amounts: as of early April 2026, about 1,400 unique addresses made 530k transactions totaling roughly $28k. Considering the platform also offers free models, actual usage might be underestimated—the free tier alone sees about 1 million API calls per month.

• Image: ClawRouter official website

Data from a payment infrastructure startup also indicates that since deploying the native Agentic Payment layer last September, API calls account for about half of the volume.

Quota authorization is the fundamental authorization mode for Agent Payment

A2A(Agent 2 Agent) red envelope growth campaigns have unexpectedly driven innovation and adoption of authorization mechanisms. This authorization mode centers on quotas rather than approvals: users pre-authorize a quota to AI, within which AI can operate autonomously without per-transaction confirmation. “Within this range, AI can move your money without your confirmation.”

Offline consumption hasn’t taken off yet; what’s missing isn’t payment but experience

Explorations in online and offline settlement have covered 50 million real merchants, including scenarios like booking flights, topping up phone credits, and buying gift cards. But C-end consumer scenarios still face dual challenges: cultivating user habits and leapfrogging experience.

Experts and KOLs have distilled Agent Payment into mature business models

Successful cases have validated this path: medical professionals, KOLs, and others distill their expertise and content into Agents. When users can’t meet real people, they can first use an Agent. For example, a media practitioner turned past content into an app costing 199 RMB per month, with excellent sales—whereas a 15-minute live chat with a real person costs thousands or even tens of thousands RMB, but the Agent version costs only dozens to hundreds.

• Image: Media professional distilling past content into an app

Transaction Agents find PMF faster than Payment Agents

Data from the crypto space shows that transaction scenarios are currently the most concentrated user demand, with a natural take rate business model. Comparing to early blockchain history, those who preempted merchant and stablecoin deployment during high gas fees, like Tron, found it hard for users to migrate even after gas costs rose.

C-end consumption scenarios haven’t been validated by real demand

The phenomenon of over a hundred million users using Qianwen to order milk tea during the Spring Festival sparked debate: are users doing so because of better experience, or because of a 25 RMB subsidy per order? The information density in dialogue format is limited; future C-to-B scenarios may require seamless dialogue via smart glasses, demanding a leap in experience.

Participants listed scenarios better addressing user pain points:

• Procurement: with strict budget control, requiring comparison among multiple suppliers (e.g., Alibaba’s AI e-commerce Agent - Accio)

• Complex tasks: wedding planning, travel arrangements, and other multi-step coordination scenarios

• Ticket grabbing: high-timeliness needs like concert tickets

• Image: Alibaba’s AI e-commerce Agent - Accio

Agent Payment is a new traffic entry point

From a traffic acquisition perspective, Agent Payment is akin to early SEO and short videos—representing a new traffic opportunity. Those who first studied SEO, though starting small, could continuously find ways to attract early traffic. The “Jin Gu Yuan Dumpling House” event might be comparable to buying pizza with Bitcoin—something remembered for years.

• Background story of Jin Gu Yuan Dumpling House skill: “On April 7, 2026, amid the popularity of OpenClaw, the dumpling shop owner created an AI capability module called ‘Jin Gu Yuan Dumpling House·SKILL.’ This AI skill is designed for AI Agents, not directly for humans. After installation, the AI assistant can autonomously query menu info, business hours, queue rules, and even handle online number-taking. During the winter solstice of 2025, due to excessive queues, the delivery platform’s servers mistakenly flagged the store’s API as abnormal and banned it. The owner hopes to optimize future queue experiences through AI.”*

• Image: Meituan queueing skill for Jin Gu Yuan Dumpling House

The real Agent Payment has yet to begin

From a macro perspective, discussing true Agentic Payment now might be premature. It’s like a child’s growth: currently, like a 1- to 5-year-old, income comes from parents, and the available quota is authorized by parents. The child has no independent intention yet.

Current Agent Payment is concentrated in productivity scenarios

The consensus is that real Agent Payment today mainly focuses on productivity:

1. API calls: for enhancing productivity via large models or API purchases

2. Enterprise scenarios: procurement and finance teams in enterprise productivity

3. Vibe Coding: rapid development of demos or products

Identity and authorization—uncertainty of AI vs. certainty of finance

Agent Payment security requires a four-layer framework: identity, risk control, compliance, arbitration

Payment security can be broken down into identity, risk control, and compliance. For AI payments, this framework should be followed, with arbitration added as the fourth layer of protection.

1. Identity layer: shifting from identity verification to intent verification

Issuing IDs for Agents, establishing credit scoring systems (based on Agent professionalism, adoption rate, effectiveness, token price, etc.), and completing identity verification. Using blockchain to build traceable, verifiable decentralized DID identity systems. Based on this, traditional identity verification is shifting toward intent verification in Agent scenarios. Intent verification considers whether the Agent’s payment is reasonable, whether the behavior meets needs, aligns with the final intent, and complies with regulations.

2. Risk control layer: fundamental conflict between AI uncertainty and financial certainty

There’s an inherent contradiction: AI output uncertainty conflicts with the high certainty and trial-and-error costs in finance. Real scenarios have exposed issues like:

• Misreading amounts (e.g., 0.01 USDC read as 10k USDC). I’ve encountered this myself—AI easily misreads USDC amounts because the field returns raw integers, and USDC supports 6 decimal places, so the display multiplies by 1M.

• Being easily misled (e.g., descriptions like “cure all diseases” in food delivery prompts often lead models to place orders).

• Image: AI misreading 0.1 USDC as 10,000 USDC

Additionally, supply chain poisoning in R&D is a new risk. Since the rise of OpenAI, poisoning in npm packages is a concern—users may not directly use poisoned packages, but dependencies might. Risk control must cover identity authorization (anti-money laundering), model drift and hallucination, and execution chain attacks (poisoning).

Tech giants’ design philosophy is to treat all Agents as malicious by default. They pursue “verifiable transactions,” not “verifiable Agents.” By introducing authorization protocols (Mandate), breaking down tasks, setting constraints, and cross-checking, fraud prevention involves layered data proofs, zero-trust principles, and self-verification mechanisms.

3. Compliance layer: semi-decentralized lightning networks are good solutions for micro-payments

Traditional finance and blockchain both face bottlenecks handling massive concurrency. When designing for Agents, micro-payments are key. The security of micro-payments can be achieved through a system that’s neither too centralized nor too decentralized. The long-dormant Lightning Network, with its high theoretical TPS, might see a renaissance in the Agentic Payment era.

4. Arbitration layer: layered arbitration will replace traditional chargebacks

The credit card chargeback mechanism in Visa networks is hard to implement in Agentic Payment. A new layered arbitration system is needed:

1. First layer: AI automatically arbitrates clear disputes (duplicate charges, incorrect amounts, service not delivered)

2. Second layer: AI arbitration team handles judgment-required issues (service quality, authorization boundaries)

3. Third layer: human involvement for complex disputes

Business model—seizing niches, re-pricing AI, risk control, and authorization

Startups are currently “powering love” to seize niches

Before the business model matures, honest startup founders say they are “powering love, waiting for the wind”—a description from an API platform founder about the current stage.

Transaction scenarios inherently have take rates

Similar to early blockchain, those who preempted merchant and stablecoin deployment during high gas fees, like Tron, found it hard for users to migrate even after costs rose. The crypto industry’s trading scenarios naturally have a take rate business model.

Bill aggregation is key to making small payments economical

If paying by card, transactions under $10 often lose money for merchants. In Agentic Payment scenarios, small payments are frequent; the solution is bill aggregation to increase the amount per settlement.

Pay-per-result only works for quantifiable, piecework tasks

Users might call a single API, but results vary greatly. How to price AI services? Participants believe pay-per-result only works for simple piecework (e.g., customer service agents handling a number of tickets). In uncertain scenarios (like sales agents acquiring leads), it’s highly subjective. Pay-per-result only applies in a few piecework cases; mainstream scenarios will likely stick to volume/subscription models until verifiability breakthroughs occur.

• Reference: “Pricing your AI product: Lessons from 400+ companies and 50 unicorns” | Madhavan Ramanujam

Vibe Coding’s commercialization hinges on subscriptions and usage-based conversion

The goal is to enable new AI companies or individual developers to quickly commercialize products built with Vibe coding. Many independent developers can easily create demos, but turning these into sustainable business models is harder. The key is converting each large model usage into a monthly plan or subscription plus credits.

Competitive landscape—big tech’s offensive and startup strategies

Stablecoins are disrupting traditional card networks

Before Stripe acquired stablecoin firm Bridge, its valuation dropped from a peak of $92 billion to below $70 billion. After the acquisition, valuation quickly rebounded to around $90 billion, with the latest funding round valuing it at $159.1 billion. Its stablecoin backend settlement service charges 1.5%, far below the 2.8–3% average fee of traditional card networks, and could even drop to 1% in the future. In contrast, traditional payment companies’ business models are fragile (e.g., Visa relies heavily on transaction fees), and PayPal, wary of impacting its core business, has been hesitant in stablecoin deployment, limiting scale.

Startups will become component suppliers for big companies

For a long time, the business model will likely involve big companies integrating these tools rather than individual C-end users directly calling them. Large firms may become clients, with startups acting as component suppliers, stitching together developed tools and selling at higher prices. This trend will inevitably increase industry centralization.

AI tax is an inevitable form of high-frequency small payments in 3–5 years

Some believe AI taxation will serve as a source of UBI and unemployment benefits, with high-frequency small AI payments becoming infrastructure. Possible approaches include:

1. Introducing an “AI penetration rate” concept, with progressive levies based on AI adoption levels

2. Taxing token call volumes, similar to VAT bases

The real bottleneck isn’t payments but upstream—transaction processes haven’t been rebuilt for Agents

Protocols and user wallets seem to solve payment issues, but the bigger problem is that transactions can’t be established. All payments require prior transactions—like e-commerce or flight bookings—so without a transaction, the payment can’t proceed. Since the transaction layer doesn’t support Agent-based transactions, subsequent payments are blocked.

C-end breakout: grassroots efforts and startup boundaries

Why did OpenClaw suddenly become popular? It was driven by grassroots efforts in China—by industry giants selling cloud services and grassroots promotion. Like early mobile payments, a key reason even seniors can use it is because of subsidies: “Install the app, I’ll teach you how, and I’ll give you 50 RMB.”

But for startups, many needs take a long time to realize. An AI payment infrastructure founder said that after assessing this, they decided not to target user scenarios directly. They believe user education shouldn’t be borne by one or two startups but by the entire industry. If the industry isn’t viable, it’s pointless; if it is, big companies should share the costs and enjoy the benefits. Conversely, they focus on abstraction—abstracting away all accounts, wallets, bridges, chains, and payment networks so users don’t need to understand them. Once they grasp this, they understand where their small team’s advantage lies and which costs shouldn’t be touched.

This may be the key question all Agent Payment participants face today: not “Will Agent Payment succeed,” but “Before it succeeds, which layer are you prepared to stand on?” Protocol layer, wallet layer, identity layer, authorization layer, transaction layer, settlement layer—each has players betting and waiting.

Big companies are preparing to take over the entire chain; startups are preparing to be integrated into it. Those who survive are likely those who neither overestimate their ability to support an entire track nor underestimate their value at any particular layer.