The Hidden Cost of Ecommerce

The great achievement of ecommerce was to make markets easier to inspect. A consumer could compare prices without visiting several shops, browse an assortment larger than any physical store, and buy at almost any hour. The internet compressed distance, lowered search costs and removed a remarkable amount of transactional inconvenience.

It did not make choosing proportionately easier.

That distinction is important because online retail tends to treat “friction” as though it were a single quantity. Page speed, form fields, payment methods and checkout abandonment are measured obsessively because they are observable. The cognitive work that precedes the cart is harder to see. A shopper may spend forty minutes deciding what belongs in a basket and forty seconds paying for it.

George Stigler’s classic 1961 paper, “The Economics of Information,” described search as an economic activity: information is valuable, but acquiring it costs time and effort. Digital markets changed that equation dramatically. Search costs collapsed. Yet the reduction in search costs often exposed a second problem. Once a buyer can see hundreds of alternatives, someone still has to decide which of them is worth choosing.

Herbert Simon’s idea of bounded rationality is useful here. People do not calculate an optimum across every available option. They work with limited attention, incomplete information and finite time. Ecommerce expanded the choice set far faster than it expanded human attention. Filters, reviews, rankings and recommendation engines are all, in different ways, mechanisms for repairing that mismatch.

The customer as system integrator

The problem is easiest to see when a purchase consists of several related products. A person preparing dinner for eight does not naturally think in tomatoes, pasta sheets, parmesan and stock-keeping units. They think about the dinner. Someone repainting a bathroom does not begin with primer, rollers, filler and masking tape. They begin with the room.

Retail databases are organized around products. Human intentions are organized around outcomes. Traditional ecommerce asks the customer to translate between the two.

That translation is work. The shopper identifies what components are required, infers quantities, checks compatibility, remembers what is already owned, weighs quality against price and notices missing dependencies. The site helps retrieve the pieces; the customer performs the integration.

This division of labour made sense when computers were excellent at databases and poor at judgment. It is becoming less inevitable. Modern AI systems can work with incomplete objectives, ask for constraints, compare trade-offs and revise a plan as new information appears. The economic importance of this capability is not that typing a sentence is easier than clicking a filter. It is that part of the reasoning can move from the buyer into software.

A different test for AI shopping

The value of that delegation should vary sharply by category. Reordering a known printer cartridge involves little decision cost. A weekly grocery shop, a home-improvement project or a complicated travel kit involves much more.

A useful way to think about adoption is to compare decision cost with delegation cost. The consumer must still explain the objective, inspect the result and decide how much authority to hand over. If verifying the agent’s work takes longer than doing the task manually, the interface has created theatre rather than efficiency.

That is why current consumer evidence is more nuanced than the rhetoric around autonomous shopping. Gartner reported in May 2026 that only 11% of surveyed US consumers were willing to let AI make purchase decisions even in relatively low-stakes categories. They were considerably more open to AI narrowing the field: 31% were willing to let it narrow choices for household supplies. The pattern is economically coherent. Delegating research carries less risk than delegating the transaction.

At the same time, Adobe’s retail data suggests that AI-assisted discovery is producing unusually high-intent traffic. Visitors referred from generative AI services have been converting better and engaging more deeply than conventional traffic. That does not prove that agents are replacing ecommerce. It suggests that software can perform useful work before the shopper reaches the store.

The next unit of friction

The industry’s next frontier may therefore be less glamorous than “autonomous commerce.” It is the systematic reduction of decision effort.

That has architectural consequences. Probabilistic systems are well suited to interpreting goals and comparing alternatives. Prices, inventory, customer identity, promotions and payments are not probabilistic facts. Production systems need a clean boundary between reasoning and commercial authority. IBM makes the same point in its work on enterprise agents: flexible reasoning is valuable in ambiguous tasks, while rigid controls remain appropriate where predictability and accountability matter.

For retailers, this suggests a different definition of a good shopping experience. The objective is not to maximize engagement or even minimize clicks in isolation. It is to reduce the amount of unnecessary thought between an intention and a satisfactory purchase.

Ecommerce largely solved the economics of access. The next generation of commerce is beginning to address the economics of judgment.

Retail databases are organized around products. Human intentions are organized around outcomes.

“Consumers are not looking to outsource shopping decisions to AI.” — Kate Muhl, Gartner