I've been thinking about how quietly the most important question in ecommerce changed this month.
For most of my career, the battle was for attention at the top of the funnel. Search, social, retail media, influencers. Whoever got you to the product page won. Checkout was plumbing. It was the part you optimized for speed and then stopped thinking about.
We started Rokt in 2012 because we thought that was wrong. The moment someone decides to buy is the moment they're paying the most attention, trust the brand the most, and have told you the most about themselves. It's where intent is confirmed, not assumed.
This month the rest of the industry noticed, because AI agents arrived at the checkout page.
In the space of a few weeks, Walmart opened its store to approved shopping agents with spending limits and its returns policy intact. Amazon blocked outside agents, including Meta's. Google began rolling out a single cart that holds products from many retailers. Meta told investors that shopping is the top use of its Muse assistant, and that demand was so high it had to ask Stripe to raise credit limits. Mastercard opened early access to payments made by machines. And the big US banks said they're finding unauthorized agents inside customer accounts and shutting them out. As the CEO of PNC put it, "there's going to have to be laws passed on liability shifting."
These aren't technology decisions. They're decisions about who is trusted at the moment money changes hands.
What changes when the shopper isn't human
An agent doesn't get tired at the third page of checkout. It doesn't respond to urgency banners. It reads what's actually there. So everything we've been getting away with on those pages (the irrelevant offers, the static messages nobody updates, the friction we told ourselves was harmless) is about to be read by something with no patience for it.
That's good news if you've been doing the work. Relevance has always been the thing that made the transaction moment valuable. Agents will reward it more harshly than people ever did.
Four questions I'd ask this quarter
I'm not going to pretend anyone has this figured out. But if I were sitting in an ecommerce company's leadership meeting next week, these are the questions I'd want answered.
How many of our orders already come from agents? And what is the week-on-week growth telling us about where this is going for us? Most companies I talk to can't answer the first part yet. You can't set a strategy for a customer you can't see.
Are we open, closed, or something in between, and did we choose that on purpose? Walmart and Amazon have made opposite calls. Both are defensible. Drifting into one by default isn't.
What does our checkout say to an agent? Is the information structured, accurate and relevant? Would an agent acting in your customer's interest choose you again?
How do we keep a relationship with the customer if we let the cart belong to someone else? (And does it matter?) If the cart belongs to Google or the conversation belongs to Meta, what do you still own? Returns, service, loyalty, the post-purchase moment. Someone replying to my LinkedIn post put it well this week: the first return may tell you more about that relationship than the first order. Decide which of those moments you'll fight for.
The part that doesn't change
Every one of the companies above is really making the same bet: that trust at the moment of purchase is worth more than anything that happens before it. Walmart is betting it can extend that trust to agents with clear rules. Amazon is betting it can't. Banks are betting the liability isn't worth it yet.
I think they're all right about the first part. The moment of transaction is where the relationship is won or lost. The difference now is that you'll have to earn it twice: once with the person, and once with whatever is acting on their behalf.
The good news is that the way you earn it hasn't changed. Be relevant. Be honest about what you're showing and why. And respect the moment.
Published 25 October 2026. A shorter version of this first appeared on LinkedIn on 24 October 2026.
