On this page
Crypto rails for AI: machine-to-machine payments
Now the other direction: crypto as infrastructure for AI. An that books your travel, bids on freelance tasks, or calls paid APIs on your behalf needs to move money — and the traditional system is a poor fit: software can't hold a bank account, can't pass , and card rails assume a human with a chargeback right. Public-chain payments are programmable, permissionless, work in tiny amounts, and don't care that the customer is a process instead of a person. The clearest example is , an open standard that repurposes the long-unused HTTP “402 Payment Required” status code: a server quotes a price, the client (often an AI agent) pays — typically in a like USDC — and the resource is released, with settlement verified on-chain. AI-agent commerce is one of the standard's explicit target use cases.
A machine pays for an API call
The unsolved part is custody — and it is the Wallets unit with a twist. An agent with a private key can spend; a prompt-injected or hijacked agent with a private key can drain. “Not your keys, not your coins” becomes “whose keys does the agent use, and who is liable when it misjudges?” The engineering answers under construction look like the controls you already know: per-agent wallets with hard spending caps, narrowly scoped and revocable token approvals, and multisig or policy contracts that require a human signature above thresholds. An agent's wallet should look like a debit card with a limit — not like your savings account.
Machine-to-machine payments on crypto rails are real and growing, but today they are a rounding error next to card rails and enterprise billing. What is genuinely new is the shape: permissionless programmable settlement is one of the few workable answers to “how does software pay software?” — which is why this quadrant, unlike the memecoin one, is engineering rather than narrative.