Agentic commerce is often described from its furthest endpoint. In that version of the future, a consumer tells an artificial-intelligence system what she wants and leaves the rest to the machine. The agent searches, compares, chooses, pays and arranges delivery.
Commerce has not reached that point. In 2026, most AI-assisted shopping still centers on discovery, recommendation and comparison, while consumers retain control over the final decision and, in most cases, the transaction itself.
That does not make the shift insignificant. An AI system does not need to complete a purchase to influence it. If it reduces thousands of possible products to a handful of acceptable choices, it has already intervened at one of the most valuable points in the customer journey.
Technology companies are now building ways for AI systems to interact more directly with merchants’ product catalogs, inventory, loyalty programs and transaction services. The details of those emerging standards matter less than the direction of travel: retail systems are beginning to communicate with software that may act on behalf of the customer.
The immediate implication is not that the retail website is disappearing. It is that the website is no longer the retailer’s only meaningful interface. A second storefront is emerging behind the first: one designed to persuade people, another designed to be understood by machines.
The storefront behind the page
For most of the ecommerce era, retailers competed within a familiar sequence. Consumers searched, clicked, browsed, compared and eventually checked out. Retail technology evolved around that journey. Search optimization improved discoverability, richer product pages improved conversion, recommendation systems expanded baskets and checkout redesign reduced friction.
AI-mediated discovery moves part of that process outside the retailer’s controlled environment. A customer looking for a waterproof running shoe for wide feet and a fixed budget can ask an AI system to interpret those requirements, compare products across merchants and produce a shortlist before any retailer homepage is opened.
At that point, the information beneath the storefront begins to matter as much as the storefront itself. Product attributes, current prices, availability, compatibility, delivery options and return conditions become inputs into discovery. A retailer can lose a sale before a customer has visited its site simply because the available data are incomplete, ambiguous or stale.
This changes the role of the intermediary. Search engines mainly directed consumers toward retailers. AI systems increasingly interpret what the consumer wants before deciding which products deserve consideration. The strategically important moment may therefore move further upstream, from checkout toward the formation of the shortlist.
Online, this is predominantly an information problem. Physical retail introduces a more difficult one.
The distance between inventory and the shelf
Consider a customer asking an AI service to assemble the ingredients for dinner from a nearby supermarket, stay within a fixed budget and ensure that everything is available for collection later that evening.
From the perspective of a database, the request appears straightforward. The catalog identifies the products, the pricing system calculates the basket and the inventory system reports that the goods are in stock.
The store may tell a different story. One item remains in the back room. Another has been placed on the wrong shelf. A delivery has been received into inventory but not yet moved onto the sales floor. The final unit is already in another shopper’s basket. The system says the product exists; from the customer’s point of view, it does not.
Retailers have dealt with such discrepancies for decades. What changes under agent-mediated commerce is their consequence. When a customer travels to a store and fails to find an item marked as available online, the result is a poor experience. When a machine is expected to make a reliable decision on the customer’s behalf, the same discrepancy becomes a failure in the chain of reasoning on which the purchase depends.
Online retail needs reliable data. Physical retail needs reliable data about a physical environment that is continuously changing.
Technologies developed long before the current enthusiasm for generative AI are already narrowing that gap. Computer vision can detect empty facings and misplaced products. RFID can improve visibility into the location of tagged merchandise. Electronic shelf labels are becoming essential infrastructure for many grocery retailers, allowing centrally managed prices and promotions to be reflected in stores with far less delay and manual intervention than paper-based systems permit.
These technologies were not introduced for AI shopping agents. Their justification has been operational. But together they reduce the distance between the physical store and the digital systems attempting to describe it.
The store is becoming more observable to software.
Retail’s reality gap
Large retailers already operate extensive digital systems. Point-of-sale platforms record transactions. ERP systems maintain inventory balances. Product-information platforms define assortments. Pricing engines distribute price changes. Order-management systems coordinate fulfillment.
The problem is not a shortage of data. It is that reality is fragmented across those systems.
A retailer can be highly digitized and still struggle to answer a basic commercial question with confidence: whether a particular customer can obtain a particular product from a particular location at a particular moment.
Inventory accuracy captures only part of this problem. The broader requirement is observability: the ability to reconcile what the retailer’s systems believe to be true with what is actually happening in stores, stockrooms, warehouses and fulfillment operations.
That distinction becomes more important as machines begin to participate in commercial decisions. An AI agent cannot rely solely on an inventory balance. It needs a credible account of what the retailer can actually promise: what the product is, what it costs, how much sellable stock exists, where it is and whether it can be picked, reserved, delivered or collected.
Those facts will continue to reside in different systems. The strategic requirement is not to combine them into one vast database. It is to reconcile them quickly enough to produce a commercially trustworthy answer when another machine asks for one.
That is a substantially harder problem than adding a generative-AI interface to an existing retail application.
From readable to actionable
Target’s Proxima project offers an indication of where this may lead. The company describes Proxima as a digital twin for inventory positioning, designed to simulate operational decisions before they are applied across the network. In a limited pilot involving 63 fresh-food items, Target said the system improved on-shelf availability by 2.5 percent. More significant than the pilot itself is the direction the company has described for the platform: AI-enabled decision support that can evaluate options, prioritize interventions and eventually automate some operational responses.
It would be premature to call this consumer-facing agentic commerce. The more important development is architectural: physical operations are increasingly being represented in forms detailed enough for software to reason over them.
A plausible path to agentic physical commerce is therefore less dramatic than the image of an external chatbot controlling store infrastructure. A customer may delegate an objective to an AI system. The AI asks the retailer what is available and under what conditions. The retailer provides reliable information about products, prices, local availability and fulfillment while retaining control over how any requested action is executed.
The customer’s AI does not need to control the store. It needs the store to provide a dependable account of what can be promised.
The next step is to move from information to action. Today, commerce systems are mostly built to answer questions about products, prices, availability and delivery. More advanced forms of agentic commerce will require them to accept requests to reserve, substitute, pick, return or redirect products.
At that point, the technical problem also becomes one of governance. Identity must be established, payment authority bounded and the limits of an agent’s discretion made explicit. Retailers will need to decide not only what information machines can access, but what actions they are allowed to take.
The competitive advantage may lie below the AI
Not every category will move at the same speed. Routine and repeatable purchases are naturally suited to delegation, while categories built around experience, judgment and physical interaction are likely to remain more dependent on human choice. Physical stores are therefore unlikely to become less relevant simply because software becomes better at shopping. Some will become more important as distributed fulfillment assets; others will derive greater value from experiences that structured data cannot reproduce.
The same underlying requirement applies to both. The store must remain compelling to people while becoming more intelligible to machines.
Much of retail’s current AI agenda remains focused on visible applications: chatbots, employee assistants, recommendation engines and generative features inside consumer apps. But if large retailers can access broadly comparable foundation models, the model itself becomes difficult to defend as a durable differentiator.
The more consequential distinction may lie in the quality of the systems underneath it. Sophisticated AI cannot compensate for product data that are incomplete, inventory that cannot be trusted or fulfillment systems that cannot support the promise the AI has made.
Preparation for agentic commerce therefore looks less like a model-selection exercise than an operational discipline. Product information must be precise enough for machines to interpret. Inventory records must correspond more closely to what is actually sellable. Fulfillment systems must support controlled actions. Identity, consent and authority must be explicit.
These investments retain value even if autonomous shopping develops more slowly than its advocates expect. Better inventory visibility reduces out-of-stocks. Better product information improves ecommerce. More accurate location data improves picking.
Agent readiness does not require belief in the most ambitious version of the future. It requires only the more modest assumption that software will mediate a larger share of product discovery, comparison and fulfillment.
The invisible storefront
For most of the digital era, retailers focused on making stores accessible through screens. AI reverses the problem. Increasingly, the retailer must make the underlying business intelligible to software it does not control.
That turns familiar operational weaknesses into competitive ones. Incomplete product data can mean exclusion from consideration. An inaccurate inventory signal becomes a broken promise. A gap between the database and the shelf can travel directly into another company’s interface.
The challenge, then, is not simply to make retail compatible with AI. It is to make the retailer’s own account of reality coherent enough for AI to trust.
For a generation, retailers worked to make the store accessible through the internet. The next phase runs in the opposite direction: the store itself must become legible to software.