The Uncomfortable Truth Behind 90% of AI Payments
At every major financial conference this year, it seems impossible to escape the latest AI hype train.
A ubiquitous gospel has been preached that autonomous AI agents are right on the verge of taking over global finance, effortlessly throwing around serious capital and running the economy on behalf of their human owners.
But is that the reality? Are AI agents moving money at scale? Who is really using them? If the broader AI bubble bursts, will agentic payments vanish alongside all that venture capital?
BeInCrypto’s latest report, The State of AI Agent Payments 2026, tries to answer these questions by looking at real transactions on x402 – a payment protocol designed for the machines.
The research, published in collaboration with AWS, Cloudflare, Chainalysis, and Zoomex, finds that the underlying rails for machine payments are technically working, but real-world commercial demand is practically a ghost town.
Is AI Ready to Make Your Payments?
The findings reveal a massive split between marketing hype and on-chain reality. While protocols like Coinbase’s x402 and Stripe/Tempo’s Machine Payments Protocol process millions of transactions, 90.8% of all recorded agent payments are microtransactions worth less than a single penny.
In fact, a 35-day study tracking 6.4 million transactions processed just $119,947 in total value.
High transaction volumes aren’t showing us an explosion in autonomous shopping or B2B commerce; they are simply masking fractional-cent automated pings for API access, compute time, and simple data checks.
Reacting to these baseline numbers, Francesco Gi, Founder & CEO of Castle Labs, points out that the industry has simply built payment rails far ahead of actual market demand:
“BeInCrypto’s report accurately captures the current state of agentic finance: the payment infrastructure works, but demand is lacking. Most activity in the machine economy is small, repetitive, and concentrated among a handful of users, indicating that the market is still early. Solutions already exist for some problems that haven’t even arisen yet, but technology ultimately depends on adoption and use cases,” said Francesco Gi.
When looking at institutional adoption, human hesitancy is even clearer.
The report’s readiness framework gave micropayments a decent score of 60.7 out of 100, but Institutional Financial Workflows scored a bottom-tier 12.0 out of 100.
Out of 19 reviewed trading tools and financial systems, 14 operate strictly within pre-approved spending limits set by humans, while setups at major exchanges like Zoomex still require a real person to confirm every single trade order.
On trading platforms like Hyperliquid, AI-labeled tools accounted for less than 0.55% of overall trading volume.
Albert Castellana, CEO of GenLayer Labs, emphasizes that sending a few cents across a blockchain is easy, but trusting software to make complex economic choices is a completely different story:
“Most real agent transactions today are relatively small: paying for APIs, compute, data, maybe simple digital services. An agent can execute a transaction exactly as instructed and still make a terrible decision because the information was wrong, manipulated, or the other side didn’t actually deliver what was agreed. Moving money is one problem. Deciding what should happen when the transaction goes wrong is another,” Castellana said.
Castellana adds that current safety discussions focus far too much on simple wallet permissions instead of overall decision accuracy:
“You can give an agent a $100,000 limit, perfect wallet permissions, and every normal security control, and it can still spend $100,000 on something that clearly wasn’t what you intended. Getting money to move at machine speed is useful. Eventually we also need a way to resolve what happens when something goes wrong at machine speed.”
The Security Risk is Real
Security risks extend beyond setting spending caps; the report highlighted $281,200 lost in exploits where validly budgeted capital was redirected to attacker-controlled wallets.
Robin Nordnes, Founder and CEO at Raiku, argues that the biggest danger in trading isn’t just malicious prompts, but what he calls the “execution gap”:
“The risk I worry about more sits between the decision and the confirmed trade. Say an agent makes a sound, compliant trade within budget, and the transaction gets delayed or outbid on the way to the block. If the agent can’t tell whether it went through, it tries again. Then again. Every retry was individually approved, and together they can do real damage.”
Nordnes believes that once dependable execution is built, especially on high-speed chains like Solana, which handled 94.5% of all transactions in the report, agents will begin taking over trading tasks directly, and human sign-offs will disappear.
Venture capital is already betting heavily on this setup.
The report shows that 92% ($3.91 billion) of VC funding went to underlying payment and cash management infrastructure like Tempo and Mastercard/BVNK, rather than pure trading bots.
As Castle Labs’ Francesco Gi concludes, the shift from sub-cent API calls to true autonomous finance is bound to happen once agents build verifiable track records:
“Today, a penny-worth credit transaction costs $0.30, making it economically unviable to use traditional finance infrastructure for micro-payments in the machine economy. As agents build verifiable track records, we expect investors to start backing those with verifiable reputations. Stablecoin and agentic finance adoption is inevitable,” said Francesco Gi.