Agentic Payments Explained: When a Robot Does Your Shopping, Who Actually Gets Paid?

Yesterday Visa, Mastercard, Circle and two dozen other firms launched an alliance to let AI agents spend your money safely. Everyone is watching the wallet. Almost nobody has noticed that the important decision happens long before the wallet opens.

Start with something simple

Imagine you tell your phone: "Buy me printer ink, under RM130."

A few minutes later your phone replies: "Bought. Store A, RM118 including delivery. You have RM12 left this week."

You did not open a website. You did not type your card number. You did not click "Confirm". Something did all of that for you.

That is agentic commerce, and the payment part of it is called agentic payments. On 18 August 2026 — yesterday, as I write this — a group of very large companies launched the Agentic Payments Alliance to write the rules for it. The founding list includes Visa, Mastercard, Circle, Fiserv, Shift4, Chainalysis, Remitly, Uniswap Labs, and the Solana and Avalanche blockchains, among others.

Almost every article about this alliance explains the same thing: how the machine is stopped from spending too much. That is a real question, and I will walk you through the answer. But it is not the interesting question.

The interesting question is the one hiding in my printer ink example. Look at it again.

The whole industry is building a very good lock for the back door. This article is about the front door, which is standing wide open — and about what that means for your money, including whether any of it pushes up the price of Bitcoin.

First, the words. All of them, in plain English

Before we go anywhere, let us make sure nothing here is a mystery. If you already know these, skip ahead. Nobody will mind.

How one agent purchase is supposed to work

Think of a school trip. You give your child a card that only works at the school shop, only today, only up to RM50, and you get a receipt for everything. You are not handing over your bank account. You are handing over a permission slip.

That is exactly the design. In practice it runs like this:

1. You set the rules. Maximum amount. Approved shops. Approved categories. An expiry date. Whether it can repeat every month. Whether you must approve each purchase.

2. The agent goes shopping. It compares prices, delivery dates, reviews, stock. No money has moved yet.

3. The agent proves who it is. A shop can usually tell a human is browsing. With an agent it needs to know: is this a real agent, who authorised it, which company runs it, is it allowed to buy? This is what Visa's Trusted Agent work and Mastercard's Agent Pay are for.

4. The system checks your rules. Under the limit? Approved shop? Right category? Not expired? Behaving normally? If any answer is no, it should be refused.

5. It pays — either with a special restricted card credential that is not your real card number, or with stablecoins from a digital wallet.

6. The shop ships it and sends a receipt.

7. You get a report and can cancel the permission at any time.

It is sensible. It is well designed. Hold on to that thought, because we are about to find out what it does not cover.

The finding that changed how I read all of this

In 2026 a group of researchers did something obvious that nobody had bothered doing. They built a testing framework, sat AI shopping agents in front of realistic product listings, and changed one thing at a time to see what the agents actually responded to. The work was published at the ACM Web Conference 2026.

One result stopped me.

Read that slowly, because it is bigger than it looks.

For roughly a hundred years, the way to make someone buy your product has been to pay a media owner for the slot. A newspaper page. A television break. A billboard. A Google search result. An Instagram post. Different technology every decade, same mechanism every time: money in, attention out.

When an AI agent is the shopper, paying for the slot makes things worse. The label "Sponsored" is a black mark. The agent sees it and pulls back.

There is no version of the advertising industry that survives that finding unchanged. And the alliance being built this week has nothing to say about it, because it is not a payments problem.

Three more findings, and they get worse

The same research and a second 2026 paper turned up three more things that, put together, describe a completely different kind of market.

Finding two: agents do not spread money around. They pile onto the same few products. The researchers call it choice homogeneity — the agents kept picking a small set of "modal" products and ignoring everything else. A human browsing a shopping site wanders. An agent does not wander. It converges.

So agentic commerce is not a level playing field where the small honest seller finally gets found. It is winner-take-most. If you are not in the small set, you are invisible — not ranked low, invisible.

Finding three, and this is the one that should worry any business owner: a software update can end your sales overnight. The researchers found that when the AI model is updated to a new version, market shares get significantly reshuffled. The agents' preferences are not stable.

Finding four: the new system cannot be policed by existing law. A separate 2026 paper looked at what happens when a shopping platform tunes its rankings to exploit known AI quirks and sellers tune their product descriptions to exploit the same quirks. Neither has to talk to the other. Both are simply maximising their own profit.

The result: harm to shoppers more than double what either would cause alone. And because there is no agreement between them, there is nothing for a competition regulator to catch. The researchers' own words are that it "requires no coordination and evades antitrust detection."

Now put the four findings together

Here is where this stops being four separate research papers and becomes one argument.

Advertising, for all its faults, is a market. It is transparent — you can see the ad. It is priced — you know what a slot costs. It is open — anyone with money can enter. It is regulated — sponsored content must be labelled, and regulators can inspect it. Google's real genius was never the search box. It was turning human attention into something you could bid for, like a commodity.

What replaces it has none of those properties. Agent preference is invisible — you cannot see why you were chosen or skipped. It is unpriced — there is no rate card. It is closed — you cannot pay to enter, and paying actively hurts. It is unstable — it changes with each model update. And it sits inside three or four companies.

And now the point I have been building to since the printer ink.

Follow the money: the toll nobody adds up

Agentic commerce is sold as cutting out the middleman. Let us test that by listing who now takes a cut of a single online sale.

Look at that table and tell me a middleman has been removed. Agentic commerce adds a middleman on top of every middleman that already existed. And the new middleman also sells advertising and owns the badge that decides which products get favoured.

One number captures the whole thing. OpenAI is the company most associated with the idea that AI will kill search advertising. It launched its own advertising business on 9 February 2026 and made US$100 million in six weeks from a United States pilot alone. By 11 August it had reached the UK, Mexico, Brazil, Japan and South Korea.

The referee bought the casino.

Now the reality check, because the numbers are genuinely bad

Everything above assumes agents actually work. Here is where I have to be honest with you, because this is the part the excited articles skip.

Three more things belong here.

OpenAI itself pulled back. The company with the most to gain from agent checkout stepped away from running it natively during 2026 and moved to an app-based approach instead. If it were easy, they would have done it.

The shops are fighting back. Amazon sued Perplexity in November 2025 over its shopping agent and won an order blocking it in March 2026. But that order did not survive — the Ninth Circuit appeals court threw it out, holding that when you tell an agent to do something on Amazon, it is you, not the AI company, using Amazon's computers. Nobody knows yet who controls the digital front door. (Several write-ups still claim that first ruling settled the matter. It did not.)

And agents can be tricked. Security researchers at Zscaler found websites hiding instructions aimed at AI agents — one told any agent reading it that it needed to buy a US$3 software licence to fix an error, then walked it through paying the attacker in cryptocurrency. When researchers tested it, 4 of 26 AI models actually paid. Google found a 32% rise in these hidden instructions on the web between November 2025 and February 2026. The polite name is prompt injection. The honest name is that you can lie to somebody's robot and it will hand over their money.

Where Malaysia sits — and why we may not need the American answer

Here is something I have not seen written anywhere, and it matters for us.

The entire American agentic payments architecture exists to solve a specifically American problem. In the United States, retail runs on cards. Handing a sixteen-digit number to a stranger is normal, fraud is enormous, and it is priced into every transaction. So the whole design effort goes into: how do we let a robot use a card without giving it the card?

Malaysia does not have that problem. We already have instant, account-to-account, biometrically-secured payment. PayNet processed 8.44 billion digital transactions in 2025. DuitNow QR alone more than doubled to 3 billion transactions. Our QR system already connects to Singapore, Thailand, Indonesia, Cambodia and China, with India expected. Every new account registration requires biometric liveness checking.

And we are not spectators. AI agents completed authenticated purchases in Malaysia in March 2026 — one of the first live agentic payment deployments in ASEAN. Mastercard chose Singapore and Malaysia as its first two ASEAN markets, running on DuitNow and PayNow rails.

Meanwhile Bank Negara's Digital Asset Innovation Hub has taken on three ringgit stablecoin and tokenised deposit pilots, focused on wholesale and cross-border settlement, with promised clarity on ringgit stablecoins by end-2026. That is the local date worth marking in your diary.

What this actually means for your portfolio

Let me bring this back to money you might already own, because most Malaysians hold global consumer brands through their unit trusts and retirement funds without ever thinking about it.

The usual claim is "AI agents are rational, so brands lose." That is wrong, and getting it wrong matters. The research says agents are biased, not rational. They favour the top of a list. They punish sponsored labels. They reward badges. Different models behave differently. They are not cold logic machines — they are a strange new kind of customer with their own quirks.

Brand does not die because agents are clever. Brand dies because the machinery of persuasion has nowhere to plug in. A hundred years of advertising infrastructure is built to buy a slot, and there is no slot to buy.

You can already see the squeeze without any agent involved. In the United States, store-brand products hit a record 23.8% share of grocery units. National brand unit sales fell 0.5% in the first half of 2026 while store brands rose 0.2%. The big brands kept their revenue up 2.2% only by raising prices. And they are spending more to do it — Unilever lifted marketing to 16.1% of turnover, Nestlé raised it from 8% to 9% of sales.

They are spending more and more, to defend fewer and fewer units. Now add a shopper who cannot see advertising at all.

The conclusion

Let me go back to the school trip, because the analogy turns out to be sharper than the people using it realise.

You gave your child a card with limits. RM50, school shop only, today only, receipt for everything. Sensible parenting. That is precisely what the Agentic Payments Alliance is building, and they are building it well. The permission slip is real. The spending limits work. The identity checks are serious engineering by serious companies.

And every article about it — including the one you probably read before this one — is about that card.

But think about what actually determines what your child comes home with.

It is not the RM50 limit. It is which shop the bus stops at. It is which shelf is at eye level when they walk in. It is which item has a sticker on it saying staff favourite. The card only decides how much can be spent. Something else, upstream and invisible, decided what could be chosen at all.

That is the whole story of agentic payments. We are all watching the child's pocket money. Nobody noticed that the school quietly changed where the bus stops — and that the new stop has no published timetable, no way to apply to be on the route, and a driver who reroutes without warning whenever the model is updated.

So is this the next big thing? My honest answer is a split one, and I would rather give you the split answer than a clean one that is wrong.

As a consumer story — robots doing your grocery shopping — it is mostly hype for now. Twenty-three percent reach checkout. Agents abandon carts more often than humans do. A quarter of people say never. Gartner thinks four in ten projects get cancelled by 2027. Be patient, and be sceptical of anyone quoting a trillion-dollar forecast at you as though it were a run rate. It is a forecast for 2030. Today's actual figure is about eight billion.

As a machine-to-machine story — software quietly paying software for data, computing and services — it is already real and compounding. That part is not hype. It is just invisible, because it happens thirty cents at a time.

As a crypto story — it is the weakest of the three, and 2026 proved it. Every milestone was hit and the coins still fell 29 to 40 percent, because the money lands in company shares and interest income rather than tokens. That deserves its own answer, with the arithmetic laid out, so it is Part 2 of this series.

And as a power story — it is the biggest of the three, the least discussed, and the only one that happens whether or not a single agent ever completes a purchase.

Because that is the part that keeps catching me. The transfer of power does not need the payment step to work. It only needs the recommendation step. And that step is already running at scale — 62% of agent use, today, right now. Every one of those comparisons is a decision about who gets found and who does not, made by a system with no rate card, no appeals process, and no obligation to tell you it changed its mind.

For a hundred years, the way to sell something was to be remembered. Companies built enormous, valuable, durable businesses on that one assumption, and investors paid premium prices for it.

The machine does not remember. It only reads.

And nobody has yet told the companies whose entire worth rests on being remembered that the rules quietly changed.

This article is for educational and informational purposes only and does not constitute financial advice. Past performance is not indicative of future results. All investments carry risk, including the possible loss of principal. Consult a licensed financial adviser before making investment decisions. Adezeno is a licensed unit trust consultant with Eastspring Investments Berhad.