Agentic Commerce Is Here, and Not Everyone Is Ready for It

For about two decades, the shape of online buying has stayed the same. You search, you compare, you click, you type in a card number, and you press “buy.” Every business with a website has been built around that sequence.

Now a different kind of customer is showing up. It does not browse. It does not get distracted by a banner. It is a piece of software acting on your behalf, and in the past few weeks it has been given the ability to spend your money.

This is what people mean by agentic commerce: AI agents that don’t just answer questions or recommend products, but actually complete the purchase or booking. Two recent Wall Street Journal reports show how fast this is moving and how divided the business world already is. One covers United Airlines banning AI bots from booking flights. The other covers banks racing to deal with AI bots using credit cards.

We read both closely. This post walks through what they say, what is still unknown, and what any organization with a website and a sales funnel should be thinking about now. It is part of our ongoing look at digital transformation and AI.

What Agentic Commerce Actually Means

Start with the difference between a chatbot and an agent, because the whole story hangs on it.

A chatbot passively answers questions. You ask, it replies. An AI agent can go and complete tasks online. It can plan a trip, which means juggling dates, flights, hotels, and reservations, and it can follow through on the booking itself. The WSJ’s United report points to trip planning as one of the most promising things agents can do, and notes that Meta CEO Mark Zuckerberg said last month that travel planning is among the top use cases for Meta’s Muse agent.

So the shift is not about smarter answers. It is about delegated action. You tell the agent what you want, set some limits, and it handles the clicking and paying.

The names in the reports are worth knowing. Meta’s Muse and Instinct are described as personal AI assistants that were empowered in recent weeks to make purchases on behalf of cardholders. OpenAI’s Dots is mentioned as another agent in this category. And OpenAI, according to the same reporting, bought a startup called Odessia this week, which is developing AI for booking travel. When companies this large start buying their way into AI travel booking, it tells you where they think the money is.

The Week the Pushback Started

On Wednesday, November 11, United Airlines said it will prevent personal AI assistants from accessing its systems to book tickets.

That is a clear line. Agents can still show customers United’s flight options and prices, but they will not be able to complete the transaction on United’s website. Looking is fine. Buying is not.

United gave several reasons, and they are more interesting than a simple “we don’t trust AI.”

• Data privacy. The airline said third-party agents could expose customers’ sensitive information or make unintended purchases. Anyone who has watched an AI tool confidently do the wrong thing will understand the second one.
• Regulation. United said automated AI actions could run into Federal Aviation Administration and Department of Transportation rules about the information airlines must present to customers. The example the report gives is almost darkly funny: an AI agent cannot attest, during check-in, that a passenger’s bags do not contain explosives. A person has to do that. A bot standing in for the person creates a real compliance problem.
• Control of the customer relationship. This is the one that explains the tone. United CEO Scott Kirby said at an event in Chicago that the company would likely take a cautious approach to agentic commerce, and put it bluntly: “I don’t want to create another middleman.”

That sentence carries a lot of history. Airlines have fought this fight before, spending years wrestling with online travel agencies over how tickets are displayed and sold. An AI agent that sits between the airline and the traveler looks, to an airline, like the same problem with better technology. The agent decides what the customer sees. The airline loses the chance to show its own offers, advertising and loyalty benefits.

It is worth being fair to Kirby here. The report says he has also described AI as a transformational tool and often talks up United’s technology. This is not an anti-AI company. It is a company that wants to decide how AI touches its customers.

Not Every Airline Is Saying No

If you only read the headline, you might think the travel industry has closed its doors. It has not. The same article lays out a patchwork of approaches:

• Southwest launched a “plug-in” in ChatGPT earlier this year so customers can search its inventory and find flights. Its CEO, Bob Jordan, described it as a way to test how customers might use AI.
• Delta said authenticated AI assistants can complete bookings, with extra safeguards for sensitive transactions such as trips for unaccompanied minors.
• American Airlines said some agents, including Muse, can buy tickets through certain partners, and customers may be able to direct agents to complete bookings on the airline’s own website.
• Other U.S. airlines have not taken a public stance against agents, according to the report.

Notice how different these are. One airline experiments with a search-only plug-in. Another allows bookings but adds guardrails for high-stakes cases. Another works through partners. And United draws a hard line. There is no industry standard for agentic bookings yet, and each company is writing its own rules in real time.

Why United Says No Now, Even Though Few Bots Are Booking

Here is a detail that gets lost. United said it is not yet seeing many bookings from AI agents. Its back-end systems can identify the source of a booking, whether that is an individual, a traditional travel agency, or an AI agent, so it has some visibility into this.

So United is not responding to a flood. It is setting policy before the flood arrives.

The wider numbers back up how early this is. Uber said that during Halloween week, its drivers made 320 million trips on Uber and Uber Eats, and AI agents made just 500 of the Uber Eats orders and four ride bookings. Four. Companies, as the report puts it, are not seeing an onslaught of agents making bookings yet.

That matters for how you read all of this. Agentic commerce today is small. The argument is about who sets the rules while it is still small, and that is exactly when the rules are cheapest to set.

The Payments Side: Banks Are Split

The second WSJ report, on AI bots using credit cards, shows the same tension from a different angle. The money layer is where the stakes get very concrete.

The core description: personal AI assistants have been empowered to make purchases on behalf of cardholders, a step toward a future where consumers no longer need to start a purchase or press “buy” at checkout. There is buzz about it, and there is worry about it, often inside the same company.

The worries, according to the reporting, include higher fraud, more returns, and new legal risks. Some payments executives say they are already seeing higher return rates from agentic shopping than from consumers shopping themselves. They also report elevated fraud when agents buy from fake online stores that never deliver. Another concern is AI supercharging points gaming, which could make premium cards more expensive for banks to run.

And then there is the question of who pays when something goes wrong. Credit cards come with fraud protection, but banks do not view an authorized charge the same way. According to the report, some executives say their banks plan to tell customers that authorizing an agent to shop with their card is equivalent to handing the card to a relative or friend. If you gave an agent permission, the bank may treat the purchase as yours. That is a significant shift in who carries the risk.

Stripe, JPMorgan and the Two Camps

The report describes early camps forming.

• The embrace camp. Stripe has embraced agentic purchasing and is seeking a central role in it, according to people familiar with the matter. Its Link wallet has more than 300 million users, and Stripe wants to use that base to let agentic shoppers complete purchases. A Stripe spokesman described agentic commerce as nascent but promising. The report adds that cards from Bank of America, Capital One and Citigroup are being used by agents when consumers enable them.
• The skeptic camp. JPMorgan Chase, the largest U.S. credit-card issuer by purchase volume, is described as among the more skeptical banks. The report says it has been blocking agents when consumers give the bots their card information, and has been slow to allow other solutions its rivals are trying.

A third view comes from executives at other large banks, who reportedly see spotting AI bots and shutting them off as a difficult, almost impossible challenge. And Michael Spelfogel, president of the card-industry tech company Cardless, offered a pointed prediction: banks can slow this down for a year or two, but they cannot stop people from using agents to spend their own money. That is one executive’s view, not a fact, but it captures the pressure banks feel. Another executive, Rain CEO Farooq Malik, called it a land grab.

How the Technology Tries to Make This Safe

Two mechanisms in the report are worth understanding, because they show where the plumbing is heading.

• Agentic tokens. Stripe’s Link uses them. They work a bit like the tokens that hide your card details in Apple Pay: the agent never sees your real card number. According to the report, these tokens come from Visa and Mastercard and include additional information, including what the consumer wants to purchase, the spending limits they set, and whether the AI bot is trusted. In plain terms, the token carries the permission slip along with the payment.
• One-time virtual cards. For issuers like JPMorgan that have been at odds with the token approach, Stripe has a separate mechanism. Its wallet creates a one-time virtual card equal to the purchase amount. Right after the transaction goes through, Stripe charges that amount to the consumer’s real card, which pays off the virtual one.

It is clever, and it creates a new problem. Merchants are complaining to Visa and Mastercard that the virtual card costs them more in fees, because it is classified as a commercial card rather than a consumer card.

That is a good reminder of something we see across digital transformation projects. A fix at one layer often shifts the cost to another. Nobody is acting in bad faith here. It is a very old system meeting a very new kind of customer. The report notes that today’s payment rules, fraud protections, federal laws, and network fees were not designed with robots in mind, and that moving one piece, even by a cent, can ripple into billions of dollars won and lost.

What Merchants Lose When the Customer Is a Bot

Beyond fees and returns, merchants raised a quieter concern: visibility. When an agent shops for someone, the merchant sees less about how that customer browsed, making it harder to collect data and recommend future purchases.

This is the same worry behind United’s “middleman” comment, seen through the lens of retail. A business’s relationship with its customer is built on data and attention. If an agent stands in the middle, both can disappear.

Different companies are making different choices. Amazon blocked Meta’s Muse from making purchases in September. Walmart, on the other hand, said it will allow consumers to authorize AI agents and will distinguish between purchases the consumer makes and those the agent initiates. That second choice is quietly smart. It lets Walmart learn what agent shopping looks like without treating all traffic the same.

Regulators Are Watching

Two more details from the bank report show this is not just a corporate debate.

Large banks, including JPMorgan, Citigroup, Bank of America, and Capital One, recently moved to restrict AI agents from accessing consumer deposit accounts, citing the safety of personal information. Economists and analysts, the report says, have raised the fear that agents could easily find higher-yielding deposit accounts elsewhere and move customers’ money there.

The U.S. Treasury Department is asking banks questions about frameworks for AI and payments, according to people familiar with the matter. This month, the Office of the Comptroller of the Currency gave its bank examiners a new framework for evaluating banks’ use of AI, which lists “autonomous payment agent[s]” in the highest risk tier.

When a regulator puts something in its highest risk tier, expect more documentation, more scrutiny, and slower adoption in regulated industries. That is useful context for anyone planning an AI roadmap.

What This Means for Digital Transformation

We want to be careful here. The reports do not tell us how this will play out, and anyone who claims certainty is guessing. But a few practical takeaways follow from what is on the record.

• Decide your agent policy on purpose. United decided. Walmart decided. Amazon decided. They chose differently, but they chose. The risk is not deciding and finding out later that your site, checkout, and terms of service quietly handled agents in whatever way your developers happened to build them.
• Know where your traffic comes from. United can tell whether a booking came from an individual, a travel agency, or an AI agent. Walmart plans to separate agent-initiated purchases from human ones. If your analytics cannot tell the difference, you will not be able to measure what is happening, let alone respond to it.
• Separate “show” from “do.” United’s approach, letting agents display options and prices while blocking transactions, is one model. Other businesses may choose the reverse or something in between. The point is that discovery and transaction are two different permissions, and you can set them differently.
• Treat authorization and liability as design questions, not legal footnotes. If a customer’s agent makes a purchase, who is responsible for a mistake? The banks are already drafting their answer. Your terms, refund policies, and customer service scripts should have an answer too.
• Expect messy edges. Returns, fees, fraud, and compliance are where this gets hard. The technology to let an agent click “buy” is the easy part.

What This Means for Education and Lead-Generation Websites

Here is where we shift from reporting to our own opinion, and we will label it as such.

Education is a different kind of decision. Choosing a course, a certificate, or a degree program is slow, personal, and requires high trust. We do not think a student will hand an agent a credit card to enroll somewhere on a whim.

But the front half of that journey, the research, is very likely to be influenced by AI assistants, and probably sooner than the buying part. A prospective student might ask an assistant to compare programs, find courses in a particular subject, or summarize what a program covers and what it costs. That makes this a search and content question before it is a payments question.

If that is right, the practical work looks familiar to anyone doing SEO, AEO, and AI optimization:

• Make program pages clear and complete. Duration, format, prerequisites, outcomes, and next steps should be stated plainly, not buried in a PDF or hidden behind a form.
• Use structured data. Schema markup helps machines understand what a page is about. For a site that publishes deep, specialized content, this is one of the strongest ways to be understood and cited by AI systems.
• Answer real questions directly. FAQ sections written in plain language help both human visitors and AI-generated answers.
• Keep facts consistent. If the program page says one thing and the blog says another, an agent will notice before a human does.
• Measure the right things. Track which pages bring in leads, where visitors come from, and whether AI-driven referrals begin to appear in analytics.

None of this requires predicting the future. It makes a website better for people today and better prepared for agents tomorrow.

So, Is Agentic Commerce a Threat or an Opportunity?

Probably both, depending on who you ask and what you sell.

For banks, it is a land grab and a liability puzzle at the same time. For airlines, it is a possible new sales channel and a possible new middleman. For merchants, it is more sales with less visibility into customers. For regulators, it is a risk category with a new name.

What stands out in these two reports is how little consensus there is. Delta allows authenticated bookings. United says no. Amazon blocks, Walmart permits and labels. JPMorgan resists while Stripe builds. When a market is still choosing its rules, organizations that think carefully and early tend to have more say than those that wait.

And a reminder worth repeating: it is early. Four ride bookings out of 320 million trips is not a revolution yet. But “not yet” is not “never,” and the companies making policy this week are betting it will grow.

Frequently Asked Questions