Blockchain intelligence firm TRM Labs has released new research analyzing nearly 200 million settlements under the Coinbase x402 protocol since May 2025, finding that the actual amount originating from AI agents may account for only 0.6% to 7.5%. As Binance, Amazon, and Stripe continue to invest heavily in the "agent economy," this report raises significant questions about the AI payments hype.
(Prior context: Are AI Agents the Next Billion Users in Crypto? USDC Takes the Lead in Machine-to-Machine Payments)
(Background supplement: Ramp surges $44 billion: Stripe capitalizes on AI new task红利)
The latest research from blockchain intelligence firm TRM Labs has cast doubt on the AI agent economy: the majority of transaction volume attributed to Coinbase’s promoted x402 settlement protocol does not originate from AI agents. According to TRM Labs’ official report, researchers analyzed settlement records on Base, Solana, and Polygon since the protocol’s launch in May 2025, covering approximately $52.7 million and 198.9 million transactions processed by known x402 operators. After excluding self-transfers and anomalous flows, the amount resembling genuine commercial transactions amounted to just $25.62 million.
Of the $25.62 million, TRM estimated the share attributable to AI agents using two models: approximately 7.5% under a loose standard and only 0.6% under a strict standard. In monetary terms, the settlement contribution from AI agents ranges between $150,000 and $1.92 million, representing a negligible portion of the total protocol volume.
How can you tell if a payment was made by AI?
TRM's logic is that genuine AI agents explore and compare across different services and products, whereas an address that repeatedly pays the same amount to the same service appears more like a script than an agent. The report acknowledges this is an intentional modeling choice that may underestimate real-world scenarios, as many current agents are single-purpose and only repeat payments to the same service—behavior that would be classified as a script.
Researchers also noted that the operation of x402 does not inherently require AI: the payer obtains a quote, signs authorization, and the operator verifies and submits an on-chain transaction—scheduling tasks, load testing, and even self-trading can produce identical on-chain records; therefore, the total protocol volume is insufficient to represent the true scale of commercial activity.
Where do the remaining funds come from? USDC accounts for 99.6%.
TRM first excluded large cash flows from self-payments, single or dual payers, and sellers with fewer than 10 buyers, resulting in $25.62 million in "potential commercial" flows. A second round of filtering applied the proxy characteristic of "amounts under $1 and highly variable"; the strict version further required this pattern to persist across months and included publicly registered agents or payments to multiple sellers.
On the timeline, activities at the end of 2025 included suspected meme coin minting and payments to a single AI analytics service; in early 2026, traffic concentrated on a single payment contract, and after mid-year, AI service payments flowed back through a proxy payment router. Throughout this period, USDC accounted for 99.6% of settlement value, approximately $52.47 million, indicating that x402 remains a stablecoin-dominated ecosystem.
Binance, Amazon, and Stripe are also making heavy bets.
Despite the cold data, investment in agent payments continues. Binance’s Agent OS, launched in August, integrates the x402 payment layer; Coinbase’s Base has targeted new agent and payment startups through a million-dollar accelerator; and in May, Amazon partnered with Coinbase and Stripe to launch AgentCore Payments, enabling agents to pay for online services using stablecoins.
TRM noted in its report that while on-chain agent registration profiles allow individuals to assert ownership of agent addresses, the current mechanism is largely voluntary and underutilized by most participants. For agent commerce to truly take off, a more accurate registration system, reputation information verifiable by agents themselves, and monitoring designs tailored for “many small transactions” rather than “few large ones” are needed. As the report concludes: “The tracks have been laid; what’s needed now is agent compliance.”
Insights from the Taiwan market
For Taiwan, the significance of this report lies in reminding regulators and the industry that the story of AI-powered payments is still in the phase of “infrastructure leading, demand lagging.” Dynamic Zone has previously reported that after Visa and Mastercard entered the agent payment space, actual transaction volumes amounted to only $136 million, aligning with TRM’s observations. As stablecoin payment pilot programs gradually emerge in Taiwan, designing a compliant framework for “machines paying machines” will become the next major focus of financial regulatory discussions.
