CX Dive reported on 3 November 2026 that the banking industry is increasingly on edge about AI shopping agents moving money inside customers’ accounts. Meta released its Muse shopping agent in September 2026, and Amazon and Walmart have since blocked or restricted its access. Executives from PNC, Wells Fargo, BMO, Stripe, Visa and Cross River Bank told the publication that nobody has settled who is liable when an agent buys the wrong thing.
The clearest example came from Cameron Fowler, the CEO of Early Warning Services, the company behind Zelle and Paze. His son used Muse to buy a red coat that did not fit. His question was simple: “who’s liable for that?” Bill Demchak, the CEO of PNC, went further and said there would have to be “a massive liability shift” onto the tech companies before banks go along. John Piazza, who runs Stripe’s card issuing product, put it plainly: “we don’t have case law on liability.” Wells Fargo already warns its customers that if they give an AI tool their username and password, they may be liable for any mistakes the tool makes.
The excitement assumes two gates that are not open
The claim around agentic commerce is that agents will soon shop and pay for us. I test every claim the same way. Is there a need, what actually changed, and what did not.
There is a need, for some purchases. Nobody wants to compare forty listings for a printer cartridge. What changed is that the agents exist now. Muse is live, and a teenager can buy a coat with it.
What did not change is who has to say yes before any of this works. Two parties hold the keys. The platform that holds the product, and the bank that holds the money. Amazon was the first gate, and it closed on Muse. The banks are the second gate, and this week they said out loud that they want liability settled before they open it.
The 50/50 is being decided by whoever holds the money
My position on agentic commerce has been that it is too early to call. It is 50/50. Platforms that feel their hold on shoppers slipping will block agents. Others will open up. Expect a split.
This story tells me where the split gets decided. Visa’s Kaitlan Moran wants a full audit trail before she calls agent commerce responsible. Cross River’s Andrew Lambert says the incentives in the existing payment model “don’t work anymore” once agent platforms sit outside the usual payment networks. That is the payment side saying the rails are not built yet.
The giants can afford this. Amazon, Meta, Visa and PNC can negotiate liability for years, and whatever they settle on will suit them. A shop in Singapore with a Shopify store and a WhatsApp number does not get a seat at that table. It gets the outcome.
So the practical question for an operator is what to build while the big players argue. The answer is to plan for a split, where agents research and recommend on some surfaces long before they are allowed to pay on any of them.
What this looks like for a shop that is not Amazon
Take a skincare store that sells on its own site and answers customers on WhatsApp and Instagram. The owner reads this news and wonders whether to rebuild checkout for agents.
Not yet. The checkout is the last thing that will open, because that is exactly where the banks sit. Wells Fargo’s warning means a customer who hands an agent their login carries the risk themselves, so most customers will hesitate too.
What is already decided for a business this size is the data side. Agents choose between products on attributes. If the store’s ingredient list, skin-type guidance, delivery times and returns policy are vague, no agent can recommend it, even on the surfaces that stay open. Wherever your information shows up, it needs more precision, more detail and an accurate reading of what you are. Winning an agent’s recommendation takes enough attributes that are materially, measurably different from the shop next door.
The other shift is the super long tail. Everyone asks their agent the same question slightly differently. To catch that, your information has to cover a much broader spread of questions than a product page used to. We do this ourselves at AeroChat. It publishes many pages, each aimed at one integration with one specific piece of software, so each page catches one long-tail question.
What to do now
- Keep the checkout you have. Humans still pay, and the bank side of agent payment is years from settled.
- Make every product and policy fact on your site specific, dated where it matters, and different from competitors in ways a machine can compare.
- Cover the spread of questions your customers actually ask, in the words they use, across pages, not on one page.
- Watch two signals before spending on agent checkout: which platforms open to agents, and whether the banks and card networks publish a liability framework.