READ MORE
The thought of automated software making purchases on behalf of consumers from online merchants thrills much of the tech and financial industry.
The chatter started in early 2025 and has surged since then, giving rise to the term "agentic commerce."
A few reality checks have called the idea into question. For example, Amazon recently blocked Meta's Muse bot, citing terms of use restrictions. This was after the e-commerce giant had previously blocked Perplexity's AI bot, and even won a preliminary injunction that was later overturned as the litigation dragged on.
But even if major platforms like Amazon haven't accepted that their e-commerce services dissolve into an API call for automated shopping assistants that seek out low prices, payment companies welcome automated shopping.
Mastercard on Wednesday published a report exploring what needs to happen to make agentic commerce workable.
"AI agents are learning to shop, pay, and act on behalf of consumers, businesses, and machines," the report says. "ICSC and McKinsey project that agentic commerce in US consumer retail could reach $1 trillion in revenue by 2030, and Gartner expects AI agents to intermediate more than $15 trillion in B2B spending by 2028."
The word "could" is doing a lot of work here. As the payment biz observes, every player in the commerce chain needs to get onboard with the idea of agentic commerce or it won't happen.
At The AI Conference in San Francisco on Wednesday, Lindsay Walker, product manager at Hedera AI Studio, said that will take some time.
"Agentic commerce is not a reality yet," said Walker during her presentation. "There are a lot of products out there claiming that they're creating agent wallets or agent cards. But when it comes down to it, it's not functional yet and hasn't truly emerged."
Within five years, she expects agentic commerce to be a real thing.
But this vulture is skeptical that there will be much demand for surrendering the power of the purse, based on several factors.
First, e-commerce sites already offer automated repeat purchasing, without AI agents.
Second, low-cost items are already trivially easy to buy from online retailers with only a few clicks. An AI agent is unlikely to improve that process.
When people purchase high-cost items, they take pains over their choices with the sort of close consideration many would be reluctant and/or remiss to delegate.
Third, trust matters in commercial transactions. AI companies have shown that their models cannot be trusted and there's no transparency into how agents would make purchases, or whether the vendors involved influence prices or purchasing decisions.
Personalized pricing via AI algorithms is already under scrutiny by US lawmakers, so further efforts to automate transactions may face regulatory friction (though perhaps not under the current regulate-yourself administration).
Points of friction
Walker defines agentic commerce broadly, as "more than just that moment of purchase. It includes discovery, it includes negotiation, it includes settlement, and it includes that post-purchase process and logistics. Discovering and comparing options is just as important as the payment moment, which is actually the only thing that people have been focusing on so far."
Within that continuum, Walker said, agentic commerce can be as simple as AI helping with product selection and the purchase process. That's viable.
It gets more complicated when users delegate authority to an AI agent that makes decisions on its own. That, she said, is where things break down currently.
Money is too important. Funds are too important. Value is too important. You need to have deterministic gates that can't be bypassed by a probabilistic model."
"The way that we build e-commerce right now works for humans, but it does not work for AI," she said.
To make AI transactions viable, she said, transactions cannot be run by probabilistic models. "Money is too important," she said. "Funds are too important. Value is too important. You need to have deterministic gates that can't be bypassed by a probabilistic model. It can't be left open to interpretation."
Rather, people need a way to provide input in a deterministic way that signals intent to purchase.
"Traditionally in commerce there's something called 'card present' where if a merchant doesn't want to be held liable for a purchase that a human claims they didn't make, they have to prove that the [payment] card is present," Walker explained. "Well, with an AI agent, cards are not present, ever."
To solve this quandary, she contends, the industry needs a mechanism for tethering a person's identity to an agent.
There are also simple functional problems with how e-commerce is set up today. Walker said there are products out there that companies describe as agentic wallets or agentic cards. But they don't work.
"This is an example checkout that I tried to do on the website," she recounted. "I was given a specific virtual card or account to make this purchase. However, there was no point in creating this because the AI agent couldn't make the purchase."
The AI agent couldn't log in to access Walker's account, which contained the relevant shipping address. It also couldn't access the DOM element on the web page where it was supposed to enter the credit card number. So Walker had to intervene and enter the information manually.
Another problem, she said, is that the site calculated taxes and shipping based on the customer's location. When agents operate from a remote server, she said, that may change the location data.
Some of these issues are being addressed by the x402 protocol for internet payments, but Walker said the specification is still evolving and only addresses the payment moment.
"The reality right now is we're simply working with agent-assisted shopping," said Walker. "There's no true agentic autonomy going on. And it's very different than autonomous agentic commerce. The human is still there at the moment of purchase. They're still making all the decisions."
For shopping bots to work, people have to be able to trust that agents will do what they're supposed to do, said Walker. And we still need suitable infrastructure.
"We need to change the whole e-commerce flow," she said. "We need to change all the surfaces so agents can interact, so they're imbued with the ability to do so. And we need to be able to do it at machine speed.
"Machine speed means that payments, agreements, and proofs have to move as fast as the agent itself between systems. Not siloed data. It needs to happen in fractions of a second, with fractions of a cent, to execute all of these transactions and exchanges."
But today's e-commerce companies are accustomed to commanding more than a fraction of a cent for their goods and services. Getting them to buy into this frictionless agentic future will not be an easy ask. ®
