Think about everything you've handed over to AI this year. Writing, research, planning a trip, and whatever your creativity takes you. Now it can spend your money.

This is agentic commerce and it's a system in which an AI agent is authorized to discover, compare, select, and potentially initiate or complete purchases on behalf of a person or business. Over the past two years, the biggest names in payments and commerce have built the infrastructure for it, but almost nobody is using it.

Why? My take is that trust hasn't been earned, and people still want to make the purchase themselves. The industry built full payment delegation before most people were ready to take it.

I've spent 15 years in payments. So instead of reading the announcements, I went and read the rulebooks, the privacy terms and the developer documentation to work out what actually got built and why adoption is flat.

Here's what I found.

Since 1992, a system called Falcon has been reading your card transactions and scoring how likely each one is to be fraud, in milliseconds, before the sale goes through. It runs on a neural network, trained on a shared pool of card data that, by FICO's own 2009 count, covered 1.8 billion accounts, and it uses that to judge a transaction it has never seen. FICO, the company that builds it, calls it a "proven, powerful fraud detection and prevention solution" that "monitors transactions end-to-end," these days across debit and prepaid and commercial cards, wire transfers, and instant payments like Zelle and Venmo.

Here's what that means when it's your card. Every time you tap or swipe, that transaction gets measured against how you normally spend, and it comes back with a score. Your bank decides what to do with that score. Either the bank approves the transaction, declines it outright, or sends you the text asking whether that charge was really you.

It isn't checking your card against a list of stolen ones. It's noticing that this purchase doesn't look like you. FICO says the profile it keeps tracks your usual spending pace, the hours and days you normally buy, and how long it should take to get between two places you've paid from. It also watches things that aren't purchases at all, a change of address, a request for a replacement card, a recent password reset.

In a 2009 talk the Chicago Fed published, FICO's own chief executive said Falcon was managing 65% of card accounts worldwide, including roughly 90% of credit cards in the United States, and that it typically cuts an individual bank's fraud losses by half, often by more.

Falcon wasn't the only neural network already working on your money. Researchers at AT&T's Bell Labs built one that read the amount you handwrote on a check and decided what number you meant. NCR put it into its check-reading machines, and banks started using it in June 1996. By the late 1990s, it was reading millions of checks a month across US banks.

Nobody called any of this artificial intelligence, mostly because nobody outside the industry knew it was running. Today, we would instantly know this is AI and the technology arrived before the worldwide understanding of it.

So why did nobody mind? Four reasons:

They had one job. Score a transaction, read a number.

Being wrong was cheap. The worst Falcon could do to you was decline a card you wanted to use. One phone call, completely reversible.

They never asked you for anything. No setup, no permission screen, no decision to make. You didn't have to know it existed, and you didn't.

They were protecting you, not spending for you. Pointed at a loss you didn't want, not at a purchase you hadn't made.

Figure 1

That's the difference, and it was never the technology. It's what it's pointed at.

Falcon and the check reader are both doing the same job: for thirty years, AI processed decisions you had already made. It scored the transaction you'd started and read the check you'd written. That's one stage of a longer story. The next one, AI influencing what you'd decide, arrived without much of a public conversation either. Today we know this as an algorithm. Now it's being asked to do a third: decide. Not recommend the purchase. Make it.

Think about what the payments world has handed you. Tapping your phone at the register. Saving your card so you don't retype it. One-click checkout. Biometric payments. Every one of those removed a step you found annoying. You were still standing right there, deciding to buy the thing. What they sped up was the buying process.

There have always been payments that happen while you're asleep such as your monthly streaming subscriptions or your recurring gym membership. But there the merchant holds a standing instruction, and the amount was agreed in advance. At its most autonomous, agentic commerce is the first time, in an everyday purchase, that something other than you or the store gets to decide what to buy.

This is where the four things that made Falcon easy to accept come apart. An agent that can be pointed at any purchase, not just one, isn't locked into a single job the way Falcon is. A purchase you have to return or dispute is not cheap to be wrong about. A thing you set up, fund and supervise is asking plenty of you. And it isn't protecting you from a loss, it's spending your money on a purchase you hadn't made.

Visa's own rules come down to three things: identity, authority, intent. Identity is solved, more or less. Verify who the cardholder is before the agent acts. Authority is where it gets uncomfortable: the rule is that you, as a consumer, are responsible for anything the agent does, as if you did it yourself, and you have to acknowledge that in advance. Intent gets specified upfront, the agent has to act on criteria you set. Afterward you get a receipt, but nothing in the rules checks the purchase against what you asked for. I'd argue that follows from authority: if you're on the hook exactly as if you'd done it yourself, there's nothing distinct left to dispute. You keep the right to dispute an agent's purchase like any other, but there's no dispute code anywhere in the rulebook for "my agent bought the wrong thing." And the one place agents do show up in the dispute rules, the agent's login can be used as evidence that the purchase was really yours.

NielsenIQ ran a quick poll of about five hundred Americans, asking what they had actually done in the past month. 42% had used at least one AI tool while shopping and 5% had used a fully autonomous AI agent to place an order on their behalf. Same survey, same people and most of them stopped short of letting an agent buy on its own.

Figure 2

Think about how you actually use AI right now. You ask it something, you read what comes back, you push on it, you keep the half you like and throw out the rest. That relationship has a role, and the role is an advisor you supervise.

And there's a reason that arrangement feels fine, which is the same reason from earlier. Ask for a summary and get a bad one, you delete it. Ask for an image you don't like, you ask again. Being wrong costs you a minute. Every job people have handed AI so far has had that arrangement. Buying something doesn't. A purchase you didn't want is a return, a dispute, or a charge you have to explain to somebody.

The usage data says the same thing. Anthropic classified what Claude actually produces in personal conversations, and the top two results are an explanation, at 25%, and a recommendation, at 22%. People are using it to think something through, not finish it for them, and that's the same shape as the shopping numbers: more people use it for the advice than for the purchase.

People have already assigned this technology a role. Advisor, editor, second opinion. Nobody assigned it the role of spending your money without asking first.

So, what's being asked isn't more safeguards: it's a role change. And you can't hand someone a new role before they're ready for it, which is why the spending limits and the consent screens and the cryptographic proof that the agent followed your rules have all landed on that 5%.

Think about how that works anywhere else. Nobody hands a new hire the company card on day one, no matter how many certifications they walk in with.

Now, the question isn't how to make someone trust an agent with a decision. It's how that trust actually gets built.

Think about how you'd come to trust anyone with anything. Not all at once, but a step at a time. An AI agent earns your trust the same way, and each step hands it a little more authority over your money. That's the trust ladder:

1.     Explain something to me. What is this, how does it work, what am I actually looking at.

2.     Recommend something. Now that I believe you understand it, tell me what you'd pick.

3.     Get it ready, and I'll confirm it. Build the cart, fill the form, line it up. I'll look it over before the transaction is made.

4.     Buy it for me. Full delegation: you don't need my approval for every transaction.

Figure 3

Each step is earned by the one below it. You don't get to four because you're good at two. You get to four because three kept going right, over and over, until you stopped checking.

Set that against where the industry actually built. Recommendations, step two, have been around since at least the early 2000s, when Amazon was already suggesting what to buy next, which is why they feel normal. Step three got built, and built well. Mastercard names it outright: in a "human-in-the-loop" purchase, its new trust layer, Verifiable Intent, confirms you were present, approved the cart and authorized the transaction. Even Amazon's "Buy for Me," marketed as agentic AI, still has you confirm the order on Amazon's own checkout page. But step four shipped right alongside it. Verifiable Intent is designed for "more autonomous scenarios" too, and Amazon's Alexa for Shopping will already complete a purchase on its own, with your default payment method, once an item reaches a price you set. The technology for full delegation is already here. The trust for it isn't.

And the rules make it sharper. At step three you confirmed the purchase, so being responsible for it as if you'd made it yourself is simply true. At step four you didn't, and you're responsible anyway. If the agent gets it wrong on something you can send back, you're out the time, maybe a return fee, maybe a trip to the post office. Annoying, but fixable. Where it really comes apart is anything you can't return: groceries already at the door, final-sale and clearance items, gift cards, digital downloads, a plane ticket on a fare you can't refund because the agent went with the cheaper one. There's no return policy to fall back on, and Visa's rules treat the agent's choice as yours. People seem to sense it. In a 16-country survey, Accenture found travel and lifestyle purchases are where willingness to hand control to an agent falls off fastest as it gets more autonomous. American Express has promised to cover agent mistakes once its program goes live, but only when the purchase strays from what you specified, not when you asked for something like "the best." That's exactly the kind of judgment you'd hand an agent in the first place.

Step three is the one people keep asking for. Accenture found 32% would hand over the choosing inside limits they set but want the money step kept back. That's step three exactly: an agent that picks, with them still confirming before it spends.

All day through step three. Compare the options. Read the two hundred reviews so I don't have to. Find the lower price. Build the cart and show me the total. A YouGov survey found 65% of Americans already comfortable with AI comparing prices for them.

Buy it for me, but only the boring stuff. The dog food, the dish soap, the thing I buy every six weeks and would buy again without thinking. Capped on price, capped on how often, restricted to stores I've already used.

Everything else, not yet, and not because I'm nervous about the technology. Anything I'd have to explain to someone else, a gift, a thing my family is going to use. Anything where getting it wrong means a return, a dispute, or a phone call. If the agent gets one of those wrong, I'm the one answering for it, as if I'd chosen it myself. So it's a question of who owns the decision, and right now the answer is still me.

And three questions underneath all of it. Has it earned it? A long run at step three without a miss. Can it explain it? If the AI agent can't tell you plainly why it picked that one, it hasn't earned the step. Not the marketing answer. The actual reason, the way a person you trust would tell you. Step one is the foundation for a reason. Can you undo it? If it can't go back, the agent doesn't get to buy it alone. Yes to all three, and it can buy it for me. Miss one, and it gets it ready while I confirm.

"Buy it for me" will arrive. It gets there the way every step on the ladder does, by asking, every time, until the asking has earned the right to matter less. Falcon never needed your permission; it was scoring a transaction you'd already made. Buying starts from the opposite place, which is exactly why this takes longer.

Contact lenses before you run out. Office supplies refilled from vendors you already approved. A brand substitution it checks with you before making. Each one asks the first several times, then less, as it earns it. Not an agent running your financial life. A narrower decision, made a little better.

Accenture found the same thing coming the other way. Nearly a third of the people they asked said what would make them comfortable moving toward an agent that buys on its own is experimenting with low-cost, low-risk purchases first. In other words, a small one that went fine.

The foundation is solid: the identity checks, the spending limits, the consent screens, the confirm-the-cart step. I wouldn't rebuild any of it. What I'd add are a few things to earn trust in the early steps. Start with the dish soap you buy every few weeks. Don't let the agent quietly reorder it. Make it ask, in its own words, until the person tells it to stop asking: "Looks like you're running low, want me to grab another one?" Or, when nothing's low but something's on sale: "There's a sale running on this right now, want it while it's cheap?" Let the person say yes or no. Make the whole thing opt-in, one category at a time, not a blanket switch flipped for them.

And three more I'd add, non-negotiable:

Show your work. A plain-language record of what the agent did and why, not a transaction log. The same "can it explain it?" question from earlier, built into the product instead of left to a support call.

Don't use the agent as cover. A new feature isn't permission to collect more data or make pricing harder to find.

Measure the right thing. Not just how many purchases went through. How often people cancel, override, send it back, or pick up the phone. That's what actually tells you whether trust is building or eroding.

So the question isn't whether AI will touch your money. It has since 1992.

What purchases would you let an AI agent make for you today, and which ones would you never hand over without a review step? I'd genuinely like to know, because I have a feeling a lot of us draw that line in the same place.