The Wallet Learns to Think
Agentic payments won't arrive as a new rail. They'll arrive as a new kind of customer, one that never sees a checkout page.
For fifty years, every payment has started the same way: a person decides, then instructs. Tap this card. Send this amount. Approve this mandate. The whole stack, from the terminal to the scheme to the bank’s fraud engine, is built around the moment a human says yes.
That moment is starting to get quieter. Software that books, renews, compares and reorders on our behalf is already here. The natural next step is that it will help pay, too.1
From instruction to intent
An instruction is precise and narrow: pay $84.20 to this merchant, now. An intent is broad and conditional: keep the pantry stocked, stay under $400 a month, and never use a supplier I haven’t approved.
The first fits neatly into today’s rails. The second needs something that can reason about price, timing, route and risk, and then act.
Instruction: one fixed path
Intent: an agent chooses the route
When intent becomes the input, the payment becomes an output. It is chosen rather than commanded. An agent might route one purchase over card for the chargeback protection, another account-to-account for cost, and hold a third until the price drops.2
The first customer who never sees a checkout page is already being built.
What breaks
Almost everything we rely on assumes a person in the loop. Strong customer authentication assumes someone can be challenged. Disputes assume someone can remember what they meant. Fraud models assume behaviour has a human rhythm.
An agent works differently. It doesn’t hesitate, doesn’t forget and can make many small purchases while its owner sleeps. With the right guardrails, that could make it one of the most efficient customers a merchant has ever had. Without them, it is a new kind of risk to manage.
Checkout pages change too. They are designed for people: images, reviews, a big friendly button. An agent looks for different things: price, terms, delivery, returns and reputation. Businesses that can describe those clearly to software may find it easier to be chosen.
An example. A family sets one intent: keep the household basics stocked, stay under $400 a month, prefer local suppliers. On a Tuesday, the agent notices milk and nappies are running low. One store has a better price but slower delivery, another is close by. It picks the nearby store, pays within the limit the family set, and adds a single line to the weekly summary: “Ordered basics from the local store, $62. Cheaper option was two days slower.”
What has to be built
Three things stand out.
Delegated consent. A mandate that carries scope, limits and an expiry date, readable by every party in the chain. Not “this agent can spend”, but “this agent can spend up to this amount, with these merchants, for this purpose, until this date”.
Agent identity. A way for a merchant or a bank to know which agent is acting, for whom, and with what authority.3 Without it, every agent looks like a bot, and every bot looks like fraud.
Receipts that explain themselves. If software is making a month of small decisions for you, you need to be able to review them in a minute. Every payment should carry its reason: what intent it served, why this route, why now.
None of this is science fiction. PayTo already gives Australia mandates with scope and limits. Tokenisation already gives us credentials that can be restricted. Real-time rails already make account-to-account payments instant. What’s missing is the layer that ties them to intent.
Who wins
It is tempting to think the winners will be whoever builds the smartest agent. There may be more to it than that.
The winners will be the ones people trust to hold the wallet: clear limits, honest receipts and the ability to say no on your behalf. In payments, trust has always been the real infrastructure. Agents don’t change that. They make it the whole game.
The organisations that build that layer may not look like payment companies at all. They may look like the most trusted assistant you’ve ever had.