Agentic Commerce

For over twenty years, we’ve shopped online in exactly the same way: browse a grid, scan thumbnails, click into product detail pages, compare, repeat. It’s so ingrained we barely notice it anymore. The scroll-and-filter ritual feels like part of the internet itself.

But habits we take for granted can dissolve overnight. Remember how search used to work before ChatGPT & co? Ten blue links, page after page, clicking through, skimming, bouncing back. Then large language models (LLMs) arrived and quietly rewired the behaviour: you ask, they answer. Research compressed into a single response, with sources linked at the bottom (recently fewer sources are hallucinated).

I was reminded of this shift listening to a recent Kassenzone episode, a German eCommerce podcast, where “agentic commerce” came up as the next big disruption in online retail. The idea is simple: what if shopping is about to shed its learned behaviours, just like search did?

How would that change eCommerce?

From agents as research companions to agents as buyers

Millions of people already heavily use tools like ChatGPT or Claude for any kind of online research. Instead of collecting twenty tabs and scanning them manually, we ask the model to summarise, weigh, and explain. That same behaviour, extended just one step further, points towards a very different kind of eCommerce customer journey.

Instead of browsing endless product lists, we tell an agent what we’re after. It finds options, compares them, and proposes the best fit. Instead of merchants competing for every click and scroll, the journey collapses: intent stated → options shortlisted → one-click confirmation.

What feels today like a curiosity (“what if an AI could buy black socks for me?") could soon feel as normal as asking an AI to explain a complex topic. And once that shift happens, the familiar act of shopping may look strangely outdated.

The OpenAI move

This isn’t just speculation. With the launch of the Agentic Commerce Protocol (ACP) and Instant Checkout, OpenAI is providing an API / protocol for agents that don’t just advise, but transact — powered by Stripe.

According to OpenAI and Stripe, ChatGPT users can already buy products from etsy marketplace, with support for Shopify stores and their product catalogues coming soon.

The protocol gives merchants a way to expose product data in a structured, machine-readable form. Instant Checkout closes the loop by letting an agent place the order directly, using the merchant’s existing payment systems. The customer still confirms, but the agent does the work.

It marks a shift from theory to infrastructure.

Amazon’s “Buy for Me”

Interestingly, Amazon has been experimenting in the same direction. Its Buy for Me feature lets customers describe what they’re looking for in natural language, and the assistant then selects items and adds them to the basket.

It’s still early and limited, but the fact that the world’s largest online retailer is testing this kind of flow matters. It suggests the idea isn’t fringe. If both a platform like OpenAI and an incumbent like Amazon are building around agent-led buying, then the old assumption — that customers will always scroll through endless product grids — suddenly looks less secure.

What shifts in the customer journey

The Kassenzone presenter, Karo Junker de Neui, put it neatly: after twenty years of online shopping, we hardly even notice how strange the ritual is. Hours spent hopping between shops, scrolling product grids, comparing specs, trying to be sure we’ve found the “right one.”

OpenAI’s Agentic Commerce Protocol and Amazon’s Buy for Me both point to a different path: agents that do the searching and curating, leaving the customer only to confirm.

That shift removes dozens of micro-moments that brands and retailers have optimised for since the dawn of eCommerce. The familiar landscape of campaigns, banners, filters, A/B tested buttons, and carefully crafted PDPs suddenly matters less if the agent does the heavy lifting.

Or maybe, it’ll only matter for the fraction of purchases where customers still want to browse and linger, have a full brand experience, rather than delegate.

Once you’re used to working with LLMs, it’s a small behavioural leap — from asking an AI to summarise a set of web pages, to asking it to buy you a pair of trainers. But the impact on how we discover and select products could be profound.

Implications for merchants

If agents reshape the customer journey, the ripple effects for merchants are hard to ignore:

Commoditisation risk

If an agent’s decision is weighted by price, delivery speed, and return policies, then the painstaking design of your storefront UI may not matter as much. Product pages, brand imagery, and conversion tweaks lose ground to the raw inputs agents actually use.

Discovery and brand recognition

Where does branding live when the customer no longer scrolls through your shopfront? Does it show up in the agent’s explanation of why it picked your product? Does it depend on whether the customer remembers you well enough to name you in the prompt? The battleground of discovery may shift from pixels to memory.

Customer relationship shift

For years, merchants have sought a “direct” customer relationship. In an agentic world, the relationship might be mediated — even owned — by the agent. That changes how loyalty, communication, and trust are built. Do you court the customer, the agent, or both?

Marketplaces like Amazon and eBay already show how transactional these dynamics can become, where the platform owns the relationship and the merchant risks becoming interchangeable.

None of these are playbooks or prescriptions. They are open questions. But they cut to the heart of how value is created online.

The ritual may not last

For decades, we’ve accepted that online shopping means scrolling, filtering, comparing. We’ve optimised every step of that ritual, treating it as permanent.

Sellers don’t need to panic. But they should at least imagine the possibility that the scroll-browse-compare ritual — the backbone of eCommerce for more than twenty years — could already be on borrowed time.