Why AI Agents Still Won't Pay for Anything - Jan Curn, Apify

Despite over a year and a half of agentic payment protocols being developed by major finance and tech players, actual adoption is nearly non-existent due to ...

Why AI Agents Still Won't Pay for Anything: A Structural Analysis of Agentic Payment Adoption

Abstract

Agentic payment protocols - standards enabling autonomous software agents to transact without human intervention - have proliferated rapidly since 2024, with more than a dozen competing specifications released by major payment networks, platform vendors, and cryptocurrency infrastructure providers. Despite forecasts projecting a $15 trillion agentic economy by 2028, measured on-chain volume remains negligible: approximately $1 million over 30 days for X402 and $1,500 for MPPP. This synthesis, drawn from marketplace operator data and protocol telemetry, analyzes the gap between protocol supply and transaction demand through a four-stage agentic payments funnel (Need → Setup → Discovery/Trust → Buy), identifying structural friction at each stage. It further examines the technical evolution of payment schemes - from fixed-price exact settlement to cryptographic escrow-based batch settlement - as evidence that the protocol layer is only now acquiring primitives required for realistic pay-per-use workloads. Practical implications for marketplace operators, service providers, and protocol designers are discussed.

1. Introduction

The emergence of autonomous agents - large language model (LLM)-driven systems capable of planning and executing multi-step tasks - has prompted speculation that machine-to-machine commerce will displace substantial portions of human-initiated transaction volume. Gartner projects agentic payments will constitute a $15 trillion economy by 2028, with 80% of business-to-business (B2B) activity conducted autonomously. This projection has catalyzed intense standards competition among incumbent payment networks, platform providers, and cryptocurrency infrastructure firms.

The central empirical observation motivating this analysis is a stark discrepancy: protocol availability has expanded dramatically while usage has not. The research question is therefore not whether agentic payment rails can be built, but why functioning rails remain unused despite substantial investment from Mastercard, Visa, Google, OpenAI, Stripe, Coinbase, and Alipay, among others.

This analysis draws on observations from a marketplace operator hosting approximately 80,000 AI tools ("actors") spanning data extraction, e-commerce, travel, search, and automation, which distributed nearly $2 million to creators in a single recent month. That this volume flows almost entirely through conventional API-key authentication rather than agentic protocols is itself a core finding. The remainder of this paper establishes the standards landscape (§2), analyzes adoption friction through a funnel model (§3), examines protocol-level technical evolution (§4), and discusses implications (§5-6).

2. Background and Related Work

Agentic payment standardization has accelerated from a single niche proposal to a crowded field within roughly eighteen months. The observed timeline includes L402 (Lightning Labs), Mastercard Agent Pay and Agent Pay for Machines, X402 (Coinbase, May 2025), KYA Pay/Skyfire (Visa, June), AP2 (Google), ACP (OpenAI/Stripe, September), TAP (Visa, October), UCP (Google/Shopify, January), MPPP (Stripe/Temple), and several Asian-market entrants including ACTP, AMP (Alipay), APOP (UnionPay), and an OKX agent payment protocol.

This density of competing specifications indicates that major financial incumbents regard agent payment infrastructure as a strategically contested layer. Attention has shifted accordingly: agentic payments occupied a single speaking slot at the prior year's AI Engineer World Fair and a dedicated track in the current year, suggesting the industry consensus has moved from speculative interest to active infrastructure building - even as end-user adoption lags.

3. Core Analysis

3.1 The Structural Case for Crypto-Native Rails

Three properties position cryptocurrency rails advantageously relative to traditional payment infrastructure. First, crypto rails have no single controlling owner, unlike card networks or bank-mediated transfers, which reduces platform lock-in risk for protocol designers. Second, crypto transaction costs are lower than credit card, PayPal, or bank debit fees for the microtransactions characteristic of agentic workloads. Third, and most significantly, traditional payment rails carry chargeback risk - a mechanism poorly suited to agentic contexts where a potentially unbounded number of agents could abuse a chargeback facility at scale. Crypto substitutes optional merchant-initiated refunds for mandatory chargebacks, shifting dispute resolution control to the service provider. Consistent with this analysis, the two most adopted standards - Coinbase's X402 and Stripe's MPPP - both rely on stablecoin settlement.

3.2 The Adoption Gap

Despite structural advantages, measured adoption is minimal. X402 scan data shows approximately $1 million in payment volume over 30 days; MPPP scan data shows approximately $1,500 over the same period. This is notable given that agentic payment directories already list established service providers - including the marketplace operator studied here, along with Exa, Perplexity, and Fireworks - indicating that supply-side listing is not the binding constraint. These same companies generate revenue far exceeding $1 million per month from agents, but overwhelmingly via API keys rather than autonomous discovery and payment. The bottleneck is therefore not protocol availability but transactional demand routed through the new rails.

3.3 The Agentic Payments Funnel

Four sequential friction points explain the gap. Need: most agentic systems rely on fixed accounts and API keys, which remain more convenient than agentic payments; consequently, agentic payment protocols compete with API keys rather than with Visa. Demand-side prerequisites include more trusted models, always-on personal agents (e.g., Meta's muse, OpenAI's agent products), and the eventual withdrawal of token subsidies. Illustrating the subsidy effect, Claude's $100/month plan provides access to approximately $14,000 worth of tokens if fully utilized, substantially reducing any incentive to purchase external services piecemeal.

Setup: crypto wallets remain opaque and harder to configure than entering a credit card number. A plausible unlock condition is direct integration of crypto payment capability into major agent platforms such as Claude or ChatGPT, removing wallet setup as a user-facing step.

Discovery/Trust: historically, commercial trust was established via personal networks, then via trusted intermediaries and marketplaces - exemplified by Claude's verified connector directory and the marketplace operator's own actor ecosystem. No equivalent trust infrastructure yet exists for arbitrary agentic services hosted on unaffiliated URLs, leaving agents without a reliable mechanism to vet unfamiliar payment counterparties.

Buy: early agentic payment systems supported only fixed-price-per-API-call billing, which is insufficient for realistic jobs requiring pay-per-use or variable pricing - a limitation addressed at the protocol level, discussed below.

4. Technical Insights

The X402 protocol's evolution illustrates how payment schemes have incrementally acquired primitives necessary for variable-cost billing. The base flow proceeds: an API request returns a 402 HTTP status with a payment-required header; the client submits a signed payment; a facilitator (e.g., Coinbase) verifies the signature; work is performed; the transaction settles on-chain; and a payment confirmation is returned. A structural risk in this flow is that if work is performed prior to settlement confirmation, the provider bears loss exposure should payment fail to materialize.

The exact scheme, launched in May, supports only fixed pricing and avoids this risk by requiring settlement certainty before service delivery - at the cost of inflexibility for usage-based billing. The subsequently released upto scheme attempted to support variable costs but failed to prevent double-spending prior to on-chain settlement, rendering it unsuitable for genuine pay-per-use billing. An interim workaround - charging an upfront amount (e.g., $10) and refunding the unused portion after service delivery (e.g., refunding $8 of $10 when only $2 of work was performed) - requires provider trust in eventual refund execution and incurs additional blockchain transaction fees for both charge and refund legs.

The batch settlement scheme, added within the week preceding this analysis, uses cryptographic escrow on-chain to enable secure microtransactions settled after work completion, addressing the double-spending vulnerability of upto without requiring upfront overcharging. Independently, MPPP introduced a sessions construct to address the same pay-per-use billing problem. These parallel developments suggest that variable-cost, usage-based billing is the binding technical requirement for protocol maturity, and that it has only recently become available - implying that low current adoption volumes partly reflect immature tooling rather than purely demand-side friction.

5. Discussion

The funnel analysis suggests that agentic payment adoption is gated less by protocol sophistication than by the absence of compelling agent-side demand. Token subsidies embedded in flat-rate subscription plans (illustrated by Claude's $100/$14,000 token ratio) create strong substitution effects that undercut incentives to transact externally, regardless of how frictionless the payment rail becomes. This implies that demand-side catalysts - subsidy withdrawal, proliferation of always-on personal agents, and increased model trustworthiness for autonomous financial decisions - may matter more than further protocol refinement in the near term.

Simultaneously, the trust infrastructure gap identified in the Discovery stage parallels earlier phases of e-commerce and API-economy development, where intermediated marketplaces (eBay, app stores, and in this context, the actor marketplace discussed in §1) substituted for absent peer-to-peer trust mechanisms. No comparable trust layer yet exists for arbitrary agentic service endpoints, suggesting an open design space for marketplace operators and platform vendors.

Finally, the proliferation of over a dozen competing standards raises coordination costs: service providers face uncertainty regarding which protocol to implement, and the fragmentation itself may suppress adoption independent of any single protocol's technical merits.

6. Conclusion

This analysis indicates that agentic payment protocols have reached reasonable technical maturity - evidenced by the rapid progression from fixed-price exact schemes to escrow-based batch settlement - while adoption remains gated by demand-side and trust-infrastructure deficits rather than protocol limitations alone. Practically, marketplace operators and service providers should anticipate that API-key-based monetization will remain dominant until subsidy economics shift and platform-level wallet integration reduces setup friction. Protocol designers should prioritize interoperable trust signals analogous to existing marketplace verification systems, since technical settlement capability alone appears insufficient to drive transaction volume.

Sources

• Why AI Agents Still Won't Pay for Anything - Jan Curn, Apify - Original Creator (YouTube)
• Analysis and summary by Sean Weldon using AI-assisted research tools

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub