TRM finds AI agents drive at most 7.5% of x402 commerce

TRM Labs says most visible x402 payment activity cannot be attributed to AI agents, even after filtering probable non-commerce flows.

Key points

TRM Labs identified about $52.7 million across 198.9 million x402 settlements on Base, Solana and Polygon since May 2025.
After filtering probable non-commerce flows, TRM estimated that AI agents accounted for 0.6% to 7.5% of the remaining $25.62 million by value.
USDC represented 99.6% of the value studied, but the methodology cannot conclusively distinguish autonomous agents from ordinary scripts.

Most payment activity visible on Coinbase-developed x402 rails cannot be confidently attributed to autonomous AI agents, according to a TRM Labs analysis published September 9 and highlighted by Decrypt on September 13. TRM identified about $52.7 million across 198.9 million settlements on Base, Solana and Polygon since May 2025, then found that roughly half the value disappeared after screening out likely non-commerce flows.

A payment protocol does not identify the buyer

x402 turns the web’s HTTP 402 Payment Required response into a programmatic checkout. A server quotes a price for an API or digital resource, the client returns a signed payment authorization, and a facilitator verifies and settles the transaction onchain. Coinbase documentation says both people and machines can use the protocol, including ordinary software, automated scripts and AI agents.

That flexibility creates a measurement problem. A scheduled job can make the same request and leave the same blockchain record as an AI system choosing among services. Transaction count and settled value therefore show use of the rail, but they do not establish how much economic activity came from autonomous decision-making.

TRM narrowed $52.7 million to probable commerce

TRM began with settlements handled by known x402 facilitators. It excluded addresses paying themselves, concentrated bulk flows involving one or two payers, and sellers with fewer than 10 distinct buyers. That process left $25.62 million classified as likely commerce. The filters are analytical judgments, not proof that every removed transaction lacked a commercial purpose or every retained transaction represented a genuine customer.

The firm then applied two models. A permissive test counted facilitator-broadcast payments whose amounts varied and averaged below $1. A stricter test also required sustained activity across multiple months plus either a public agent registration or payments to more than one seller. Those tests produced an estimated agentic share of 0.6% to 7.5% of the screened commerce by value.

The estimate has an important blind spot

TRM cautioned that its approach may undercount single-purpose agents. An AI service repeatedly purchasing one resource at a fixed price can resemble a conventional script, while voluntary on-chain agent registries cover only part of the market. The result is best read as a modelled range under stated rules, not a definitive census of AI-controlled wallets.

The asset mix was less ambiguous. TRM reported that USDC accounted for $52.47 million of the $52.68 million total it measured, or 99.6%. Agent-like payment volume remained small at roughly $5,000 to $11,000 per month, even as the number of participating addresses moved sharply during 2026.

Why the findings matter to developers and compliance teams

For developers, the study separates technical readiness from demonstrated demand: x402 can settle machine-readable micropayments, but visible protocol totals should not be presented as equivalent to AI-agent spending. For merchants, attribution affects product design and customer acquisition. A service built for autonomous buyers needs machine-readable descriptions, discoverable pricing and reliable proof of what was purchased, rather than a conventional checkout page. For compliance teams, many low-value automated payments can create high transaction counts while voluntary identity claims leave responsibility unclear. Better registration, counterparty reputation and monitoring designed for machine-scale traffic may be needed before agent commerce can grow to higher commercial scale without weakening controls.

Sources

AI-generated editorial image; not a photograph of the reported event. Prepared with AI assistance and source verification.