If you sell anything online, you've probably seen the headlines about AI agents doing the shopping. Most of that coverage is a few years ahead of reality. But something real did happen underneath it, and it's worth understanding before somebody tries to sell you a solution for it.
Over the past eighteen months, the major card networks and processors built infrastructure for purchases made by AI on a customer's behalf:
I've been in payments long enough to watch a lot of technology show up before anyone figured out the rules for it, and this is another one. The purchase can happen just fine - what's missing is an answer to who's on the hook when the customer says they never asked for it.
The buy-inside-the-chat experience turned out to be harder than the early announcements suggested. OpenAI pulled back its Instant Checkout product in March 2026 and moved toward a model where the AI helps the customer find the product and the customer completes the purchase on the merchant's own website.
For merchants, that's the better outcome. The sale still happens on your site, which means you keep the account, the login, the email address, and whatever loyalty history goes with it. The AI acts more like a very good referral source than a middleman.
As for how big this gets, the honest answer is that nobody knows. Writing in The Green Sheet, Chargebacks911 founder Monica Eaton points to industry projections that agentic commerce could reach a quarter to nearly a third of global online purchases by 2030. Treat that as a directional guess rather than a forecast.
What's more useful is a Checkout.com study from June 2026 finding that about three quarters of merchants expect consumers to adopt AI-assisted shopping faster than merchants can prepare for it. Regardless of timeline, I think it’s spot on that consumers will be ready before merchants.
None of these are emergencies today. They're the mechanisms to understand now so you recognize them when they show up on your statement.
Fraud tools look for suspicious human behavior. Unusual device, odd purchase pattern, session activity that doesn't add up. An AI agent buying legitimately for a customer can trip the same rules a scraper does. Its clicks and typing don't look human, if there are any at all, and it checks out faster than any person could, so the fraud tool may read it as a bot.
Eaton makes the point that this shows up as declined orders rather than chargebacks, which means it doesn't appear on the reports most merchants actually look at.
Today the volume is small enough that few merchants would notice, but I'd rather you know how this works now than find out from a dip in your approval rate a year from now.
Card disputes have historically fallen into three buckets: real fraud, merchant error, and buyer's remorse. Each one has established evidence requirements and a clear path for who absorbs the loss.
An agent-initiated purchase can fall outside all three. The card wasn't stolen. The merchant didn't make a mistake. The agent may have followed its instructions exactly. And the customer still looks at the charge and says they didn't want it. Maybe a subscription renewed automatically, or the agent substituted a similar product, or it booked a flight that fit the calendar but not the preference.
The evidence that normally wins a chargeback, things like authentication results and delivery confirmation, doesn't speak to the question at the center of that dispute, which is what the customer gave the agent permission to do. The networks haven't settled who absorbs that loss. Until they do, I expect these to go the way most gray-area disputes go, with the benefit of the doubt going to the cardholder and the merchant absorbing the loss.
When an AI recommends a product, it's working from whatever structured product data it can find. That means your product pages and, if you have one, the product feed your ecommerce platform generates: item names, prices, sizes and options, whether something is actually in stock, shipping timelines, and your return policy. A person can figure most of that out from a photo and a paragraph. Software can't. If the data is thin, inconsistent, or out of date, the AI recommends a competitor whose data is cleaner.
Of the three, this is the one I'd actually spend time on today, because it isn't really about AI. If you've been putting it off, AI gives you one more reason to knock this one out.
Get your product information in order. Accurate pricing, real inventory counts, complete product details, a return policy written on your site in plain language rather than buried in a PDF. This pays for itself in search and customer service regardless of what happens with AI.
Look at your declines, not just your chargebacks. Most merchants review chargebacks monthly and never look at declined authorizations at all. That's a blind spot with or without AI in the picture. Ask your processor for decline reporting and find out what percentage of attempted sales you're turning away and why.
Ask your processor what they're seeing. Not whether they have an AI strategy, which everyone will say yes to. Ask whether their platform can identify an agent-initiated transaction, and whether they're seeing any yet in your vertical. The answer tells you how close this is to mattering for your business.
Make cancellation easy. If you bill on a recurring basis, a customer should be able to cancel without emailing support. That's good practice today, and it becomes more important as more subscriptions get set up by something other than a human.
You'll see advice telling merchants to build permission frameworks and audit trails that record exactly what a customer authorized an AI agent to do. That will likely matter someday. It doesn't matter yet, and building it now means guessing at standards the networks haven't published.
The reasonable position is to know that the requirement is coming, so that when your processor or platform rolls out support for it, you understand what it's for and turn it on. Don't go build it yourself.
If you operate in a restricted or high-risk vertical, expect AI shopping platforms to exclude you early, the same way ad platforms and mainstream gateways do. That doesn't change the fundamentals. Keep your dispute ratio well under threshold and work with a processor who'll advocate for you when a rule changes. Those two things have protected high-risk merchants through every platform shift so far, and they'll carry through this one.
My take after 16 years in this business is to pay attention but not spend money on this yet. The rails are being built by Visa, Mastercard, and the processors you already work with, so most of it will reach you as a feature your processor turns on. In the meantime, keep your product data clean and keep an eye on your declines. Both are worth doing no matter when agentic commerce takes off.
---
Sources: Monica Eaton, "Why agentic commerce is creating a new category of dispute," The Green Sheet, Issue 26:07:02 (read it here); Visa Trusted Agent Protocol announcement, October 2025; Mastercard Agent Pay; Checkout.com, "Agentic Commerce 2026: The State of Consumer Demand and Merchant Readiness," June 2026.
Want a second set of eyes on where your setup stands? We help merchants across standard and high-risk verticals get a clear read on their processing, fraud, and dispute exposure. Reach out to Polaris Payments and we'll walk through it with you.
