Think back to the last thing you bought online. Somewhere in that flow you confirmed the payment at a checkout page. Agentic commerce removes that moment: the purchase still happens, but you are no longer the one tapping “pay”.
It is closer than most people realise: 62% of Australian consumers are open to AI agents handling purchasing decisions, but the payment layer underneath still assumes a human at the checkout. This article explains the what and the why, so you can judge where card rails hold and where they don’t. For the broader framing, see AI agent wallets and the agentic commerce landscape.
What is agentic commerce, and how is it different from traditional e-commerce?
Agentic commerce is transacting where an AI agent discovers, decides and completes a purchase on your behalf, within limits you set. In a normal purchase a human browses, selects and checks out. In agentic commerce, the agent researches, compares and buys within guardrails you define.
The difference is who decides and pays. The agent compresses discovery, comparison and checkout into one automated decision layer, payment included. As Accenture puts it, payments move from the last step in the purchase into the middle of it.
This inverts the funnel. In traditional e-commerce, intent forms while a human browses and authorisation comes when they click pay. In agentic commerce, both resolve before any human sees an interface, and the first you know is a notification.
The agent functions as a decisioning layer: a chatbot recommends, while an agent chooses and executes within its mandate. That matters, because the payment downstream now has to handle something it was never designed for.
Why do AI agents need machine-native payment rails instead of existing card rails?
Card rails were engineered around one assumption: a human is present to authorise the transaction. You tap, enter a PIN, clear a 3-D Secure step-up. Settlement happens in batches, and chargebacks assume a person filing a dispute. Every rail in use today assumes a human is ultimately responsible.
An AI agent violates every one of those assumptions. It transacts programmatically, around the clock, with no one to answer a step-up challenge, and no jurisdiction currently grants it the legal personhood to open an account or pass a KYC check.
The economics strain just as hard. Cards charge a fixed fee on top of a percentage, typically around $0.30 per transaction, because they were priced for human-sized tickets. When the fee eats the ticket, the rail becomes a rounding error with an invoice.
Machine-native rails exist for this. They settle over stablecoins, need only a cryptographic key pair, and run 24/7 with fees in fractions of a cent, built for programmatic signing rather than CAPTCHA and two-factor prompts. We cover the machine payment rails agents actually use separately, and the wider agentic commerce picture sets the broader context. Which rail a transaction can afford comes down to its size and frequency, and that is the split that follows.
What is the difference between macro commerce and micro commerce in agentic payments?
“Agentic commerce” is two different things, and the split decides which rail a transaction can afford. Visa draws the line: macro commerce is the agent acting for a person, booking a flight or managing a subscription. It is high value and low frequency, and looks a lot like ordinary e-commerce.
Micro commerce is the other end: small, frequent payments one piece of software makes to another, often well under a dollar. A few cents for an API call, a unit of compute. Nobody creates an account, and there is no human creating the payment at all.
The difference is value and frequency, and it determines which rail each can ride. Macro commerce tolerates card interchange because a $1,500 flight can absorb a $0.30 fee. Micro commerce cannot. This is where unit economics decide everything.
Take an average agent transaction of $0.20. On card rails, the $0.30 fixed fee is bigger than the whole payment, which is why 76% of AI agent transactions fall below Visa’s $0.30 floor. On stablecoin rails settling on Layer 2, the same transaction costs around $0.0001. Micropayments only work when the fee is small enough to ignore against the ticket.
This is where metered pricing and per-use billing come in: charge fractions of a dollar per API call without the fee swallowing the margin. For Australian merchants it is local too: the NPP, Osko and PayTo were built for human-initiated payments, none of them for agent delegation, and interchange would wipe out a payment that small before it reaches you. Whether that traffic is arriving yet is the adoption question. Most real flows will use both: cards for the macro purchases, machine-native rails for the micro traffic.
Conclusion
The human checkpoint at checkout was quietly doing a lot of work: authorising the payment, providing someone to file a dispute, and carrying minimum-ticket economics. When the human leaves the loop, those assumptions leave with them.
What remains is a decision layer transacting at a granularity card rails were never designed to settle. Agentic commerce is a new transaction class in its own right, and because it splits into macro and micro there is no single right rail: cards keep the high-value, low-frequency purchases, and machine-native rails take the sub-dollar traffic.
The next question is how this fits into agentic commerce and machine payments. For now, the takeaway is simple: figure out which kind of transaction you are handling, and the rail decision mostly makes itself.
Frequently Asked Questions
Is agentic commerce just e-commerce with a chatbot in front?
No. A chatbot recommends while an agent decides and pays, and that difference in execution authority is the whole point. In traditional e-commerce a human still clicks checkout, so the payment event stays a human action. Agentic commerce delegates that decision to software that completes the purchase within limits you set, compressing discovery, comparison and payment into one automated flow. That is a different transaction class, not a new interface.
Why can’t AI agents just use virtual cards on existing card rails?
They can, but only for larger, slower transactions. Virtual cards still inherit interchange and a fixed fee floor of roughly $0.30 per transaction. Keyrock’s May 2026 data found 76% of AI agent transactions fall below that floor, so the economics collapse before the card is even tapped. Virtual cards also assume interactive authorisation and human dispute models that agent traffic does not fit.
Why can’t an AI agent just open a bank account and pass a KYC check?
Because every bank account and card is tied to a legal person, and no jurisdiction currently grants AI agents legal personhood. An agent holds no identity documents and cannot complete CAPTCHA or two-factor prompts. A 2026 METR experiment gave an advanced agent $5,000 and four days to earn money; it made nothing, blocked at every turn by the same human verification flows.
Who is liable when an autonomous agent makes a purchase the user did not intend?
Liability still sits with the human or organisation that delegated the authority, at least for now. Card rules derive responsibility from proof that a human was present, and that proof disappears when software transacts. Emerging practice pushes the answer into the delegation chain itself: signed, scoped permissions that record what the agent was allowed to do and who granted it.
What is “Know Your Agent”, and why does it matter for payments?
Know Your Agent (KYA) is an emerging identity framework that verifies the agent, not just the end user. It issues cryptographically signed credentials attesting to the agent’s identity, the organisation that deployed it and its authorised scope. KYA matters because trust, not capability, is the limit on autonomous commerce. It gives investigators an audit trail without requiring the agent to hold legal personhood.
What are agentic payment protocols like x402, AP2 and ACP, and do I need to pick one?
They solve different layers, so most stacks will use more than one. x402 revives the HTTP 402 status code so any API endpoint can be paid in stablecoins, AP2 defines how agents, merchants and wallets coordinate user authorisation, and ACP and UCP handle the checkout layer. Treat them as complementary standards rather than competing bets, and choose based on which layer you actually need.
Are Visa and Mastercard building their own agent payment rails?
Yes, both networks are extending card rails rather than conceding them. Visa’s Intelligent Commerce provides tokenised credentials and APIs for agent transactions with over 100 partners, while Mastercard’s Agent Pay for Machines applies a similar overlay. The card networks are betting that tokenisation, scoped credentials and liability shift let them keep high-value agent traffic on familiar rails.
Do I need to rip out my existing card stack to support agents?
No, and doing so would be premature for most teams. Card rails still hold for macro commerce, the high-value, low-frequency purchases agents make on a user’s behalf. The pragmatic path is a hybrid: keep cards for larger tickets, and add machine-native settlement only where per-request volumes and sub-dollar values make card economics unworkable.
How do I evaluate the unit economics of agent micropayments?
Compare the fixed fee per transaction against the average ticket size, then against your margin. If your average transaction sits around $0.30 to $0.48, as Keyrock found for agent traffic, a $0.30 card floor consumes the entire ticket. Machine-native rails on Layer 2 settle for roughly $0.0001, which is what makes high-frequency micropayments viable. Metered, per-use billing maps naturally to that model.
Is agentic commerce actually going to replace card payments, or is it mostly hype?
Neither, and the split is by transaction size. Card rails will keep the high-value, low-frequency purchases that tolerate interchange, while machine-native rails take the sub-dollar, high-frequency traffic card economics cannot serve. Analyst projections of trillions in agentic commerce by 2030 are real, but they describe a new transaction class sitting alongside cards, not the end of them.
What are the biggest security risks when agents initiate payments?
The main shift is that compromised authority can affect streams of transactions, not just one. Attackers are experimenting with synthetic delegation and prompt injection to manipulate an agent’s decisions, and Darwinium found 97% of organisations saw more AI-facilitated attacks in the past year. The defences are agent-aware: scoped, revocable permissions, verifiable agent identity and strict separation between decisioning and payment execution.
