Agentic Commerce Is Happening More Slowly—and More Deeply—Than the Hype Suggests

There are two easy mistakes to make about agentic commerce in 2026. The first is to take the demos literally and assume that autonomous software is about to replace ordinary online shopping. The second is to notice that most people still approve purchases themselves and conclude that little has changed.

The evidence supports neither extreme.

Forrester’s mid-2026 assessment is appropriately restrained: most experiences marketed as agentic are still conversational, and humans still drive the purchase decision and checkout in most cases. Gartner found that willingness to delegate the final purchase remains low. Deloitte’s European research similarly shows broad interest in AI assistance alongside persistent demand for transparency and control.

That is the consumer-adoption story. The infrastructure story is moving faster.

The rails are appearing before the habit

Over the past year, major commerce and payment companies have converged on the same missing pieces.

Google and Shopify have built UCP around product discovery, carts and checkout. OpenAI and Stripe have ACP. Google has AP2 for payment mandates. Visa and Mastercard are working on agent identity, authorization and payment acceptance. Stripe has launched a protocol for machine payments. Shopify has opened structured catalogue and commerce capabilities to any developer. Live authenticated agentic transactions have already been demonstrated in Europe.

Individually, every announcement can be dismissed as platform positioning. Collectively, they form the early outline of a stack: machine-readable products, live merchant state, portable identity, explicit authorization, payment rails and standardized actions.

Infrastructure transitions often look like this. The rails become credible before the dominant user behaviour changes.

Full autonomy is the wrong milestone

The industry also tends to define success too narrowly. An agent does not need to make the final purchase independently to change the economics of shopping.

If software compresses a forty-minute research task into five minutes, builds a correct basket and asks the user to approve it, a meaningful part of the transaction has already moved. Search, comparison, planning, basket construction, authorization, payment and post-purchase service can be delegated separately.

McKinsey’s “automation curve” is useful for precisely this reason: agentic commerce is a progression of delegated activities rather than a sudden handover from human to machine. Its projection that agents could mediate trillions of dollars of consumer commerce by 2030 is a forecast, not evidence of current adoption. The more reliable signal is that many of the companies that would have to make such a market function are now spending real engineering effort on the prerequisites.

What retailers should infer

The appropriate response is therefore neither panic nor passivity.

Retailers should improve product data because AI discovery already exists. They should make price, inventory and cart functions accessible through controlled services because transactional agents need live truth. They should measure AI-referred traffic and recommendation visibility. They should experiment first where shopping requires significant research or basket construction. And they should keep identity, authorization and payment boundaries explicit.

They should also avoid protocol religion. The standards will evolve. The durable investment is a merchant architecture that can expose the same commercial capabilities safely to several interfaces.

ASOS CTO Przemek Czarnecki offered perhaps the most useful warning in his June interview with McKinsey: the biggest mistake is investing in use cases that are not commercially meaningful. That is a better test than whether a project looks futuristic.

Agentic commerce is not an overnight revolution. Its interface is arriving unevenly. Underneath that interface, however, the assumptions of ecommerce are already being rewritten: who performs the search, where the shortlist is built, how products are represented, who is authorized to act and what counts as a customer.

The change is slower than the hype because consumer behaviour takes time. It is deeper than the hype because the infrastructure being rebuilt will outlast the current generation of assistants.

The rails are being built before the behaviour is mainstream. That is often how infrastructure transitions begin.

“Most ‘agentic’ experiences are still conversational.” — Forrester, May 2026