AI Agents Are Not Spending as Much Money Online as You Think

The idea of artificial intelligence agents independently buying products and services online has attracted plenty of attention across the technology and cryptocurrency industries. But new research suggests that the much-discussed era of AI agents spending money autonomously may not have arrived yet.

According to blockchain intelligence firm TRM Labs, only a small portion of the payment activity taking place through Coinbase’s x402 protocol appears to be generated by genuine AI agents.

The findings challenge the growing assumption that large volumes of blockchain-based payments automatically represent AI-driven commerce.

TRM Labs analyzed approximately $52.7 million in payment activity across 198.9 million settlements processed through known x402 facilitators on Base, Solana and Polygon since May 2025.

After filtering out transactions that appeared to be self-payments, bulk activity and other unusual flows, researchers estimated that only around 0.6% to 7.5% of the remaining transaction value could be attributed to AI agents.

That suggests most activity on the payment network is still being generated by conventional software, users or automated systems rather than autonomous AI shoppers.

What Is Coinbase’s x402 Protocol?

To understand the research, it helps to look at what x402 actually does.

Coinbase introduced x402 in 2025 as a way to make online payments part of ordinary web requests. The system is designed to allow a buyer to receive a price, authorize a payment and complete the transaction through blockchain infrastructure.

A facilitator handles much of the process. It verifies the payment authorization, submits the transaction to the relevant blockchain and handles the associated network fee.

The technology has attracted attention because it could provide a straightforward payment system for AI agents.

An autonomous agent could theoretically discover a service online, determine its cost, authorize a stablecoin payment and receive the requested information without requiring a person to manually approve every transaction.

That vision has helped fuel interest in agentic commerce, where software agents conduct economic activity on behalf of people or businesses.

But TRM Labs says blockchain transaction data alone cannot prove that an AI agent was responsible for a payment.

A Blockchain Transaction Doesn’t Prove an AI Agent Was Involved

One of the biggest challenges in measuring AI-driven commerce is distinguishing an intelligent agent from ordinary automation.

A conventional computer script can perform many of the same actions as an AI agent.

For example, a scheduled program could repeatedly request a service and automatically make payments through x402. Load-testing software could generate transactions. Automated applications could interact with a payment service without any artificial intelligence being involved.

From the blockchain’s perspective, these activities can look remarkably similar.

This means simply counting x402 settlements and labeling them as “AI payments” could dramatically overstate the amount of autonomous commerce taking place.

TRM Labs therefore developed a methodology designed to identify payment patterns that are more consistent with genuine AI agents.

TRM Labs Filtered Out Suspicious and Anomalous Activity

The researchers began by removing transactions that appeared unlikely to represent normal commercial activity.

This included self-payments, large flows involving only one or two participants and sellers that had fewer than 10 buyers.

After those filters were applied, approximately $25.62 million in activity remained and was considered more likely to represent genuine commerce.

Researchers then looked at the characteristics of the payments themselves.

One test focused on transactions broadcast through facilitators where payment amounts varied and generally averaged less than $1.

A stricter methodology required those patterns to continue across multiple months, along with additional signals such as public registration identifying an address as belonging to an AI agent or payments being made to multiple sellers.

The goal was to separate a genuine agent that explores different services from a simple automated script repeatedly accessing the same service.

The Research Could Still Underestimate AI Commerce

TRM Labs acknowledged that its methodology has limitations.

A genuine AI agent doesn’t necessarily need to behave like a human shopper.

For example, an agent could be specifically designed to purchase one particular service repeatedly. Such an agent might make identical or highly similar payments to the same provider.

Under TRM’s stricter criteria, that behavior could look more like a traditional automated script than an AI-powered system.

The researchers therefore warned that their estimate could understate the true level of AI-agent activity.

In other words, the 0.6% to 7.5% figure shouldn’t be interpreted as proof that only that percentage of all x402 activity involves AI.

Instead, it represents the portion of observed commerce that could be identified as agent-driven using the researchers’ particular methodology.

Activity Has Changed Over Time

TRM’s research also found that the nature of x402 activity shifted throughout the period studied.

During late 2025, some transactions appeared to be associated with activities such as meme-token minting and payments involving an AI analysis service.

In early 2026, activity became heavily concentrated around a single payment contract.

Later, payments associated with AI services began appearing again through an agent-payment router around the middle of the year.

This changing pattern highlights another difficulty in evaluating the growth of AI commerce.

A payment protocol can experience rapid changes in transaction volume without that necessarily meaning the number of autonomous AI agents is increasing at the same rate.

Large bursts of activity may instead come from individual applications, services or automated processes.

USDC Dominates x402 Payments

Another notable finding from the report is the dominance of USDC within the x402 ecosystem.

Across the full period examined by TRM Labs, USDC represented approximately 99.6% of settled value, equivalent to around $52.47 million.

The heavy use of stablecoins makes sense for machine-to-machine payments.

AI agents need predictable payment values when purchasing digital services. A highly volatile cryptocurrency could make a service that costs a few cents significantly more or less expensive within a short period.

Stablecoins such as USDC are designed to maintain a value tied to the U.S. dollar, making them potentially more practical for automated online transactions.

Companies Are Still Betting on AI Agent Payments

Despite the relatively small amount of clearly identifiable AI commerce, major technology and cryptocurrency companies continue investing heavily in the concept.

Binance, for example, included an x402 payment layer as part of its Agent OS initiative launched in August.

Coinbase’s Base network has also targeted startups working on AI agents and payment infrastructure through an accelerator program offering up to $1 million in support.

Meanwhile, Amazon announced AgentCore Payments in partnership with Coinbase and Stripe, providing infrastructure that can allow AI agents to make online payments using stablecoins.

These investments suggest that the industry believes autonomous digital commerce could become significantly larger even if today’s transaction data doesn’t yet show widespread adoption.

The Bigger Problem: Who Is Responsible for an AI Agent’s Payment?

TRM Labs also highlighted a different challenge that could become increasingly important as AI agents become more common: accountability.

If an AI agent makes a payment, who is responsible for that transaction?

Blockchain addresses don’t automatically reveal whether they belong to an individual, company, application or autonomous software agent.

Some blockchain ecosystems have introduced agent registries that allow people to publicly declare ownership of an agent address.

However, TRM Labs noted that these declarations are voluntary and currently aren’t used by most participants.

That creates a problem for businesses that want to determine whether they’re dealing with a legitimate AI agent.

Agentic Commerce Needs Better Identity and Reputation Systems

As AI agents begin interacting with thousands or potentially millions of online services, businesses will need better ways to determine who—or what—is on the other side of a transaction.

TRM Labs argues that the industry needs more accurate registration systems, reputation information that agents can independently evaluate and monitoring tools capable of handling huge numbers of small-value transactions.

Traditional financial monitoring often focuses heavily on transaction value.

AI commerce could work differently.

An autonomous agent might make thousands of tiny payments every day, each worth only a few cents. Individually, those transactions may appear insignificant, but collectively they could represent substantial economic activity.

That means compliance systems will need to monitor transaction behavior and volume, not just dollar amounts.

AI Agent Commerce Is Still in Its Early Stages

The TRM Labs report doesn’t mean AI agents aren’t capable of making payments. The underlying technology already works.

The bigger question is whether autonomous agents are actually using these systems at scale.

For now, the evidence suggests the answer is not yet.

Most x402 payment activity appears to come from sources that cannot confidently be identified as autonomous AI agents. Some may be ordinary users, while others are traditional scripts, automated applications or other forms of software.

That could change quickly as AI agents become more capable and companies build more services specifically for machine-to-machine transactions.

The Bottom Line

The promise of AI agents independently navigating the internet, purchasing services and completing transactions has become one of the most talked-about developments in digital commerce.

But according to TRM Labs, the blockchain data doesn’t yet support the idea that AI agents are responsible for most x402 payment activity.

Only an estimated 0.6% to 7.5% of qualifying commerce by value showed characteristics associated with AI-agent activity under the firm’s methodology.

The technology behind agent payments is already operational, and major companies are investing in the infrastructure needed to expand it. But before AI-driven commerce can reach a truly massive scale, the industry still needs better systems for identity, reputation, monitoring and accountability.

The payment rails may already be ready. The real challenge is proving who—or what—is actually using them.

As TRM Labs’ research suggests, the future of agentic commerce may depend not just on making AI agents capable of spending money, but on building a financial ecosystem that can reliably identify, monitor and trust those agents.