Where are we in AI-powered shopping? What does Anthropic's latest move tell us?
We've been talking for a long time about how artificial intelligence will change the shopping experience. In the early stages, we saw this change mostly through personalized product recommendations, more successful search results, and customer-specific campaigns. With generative AI, a new era has begun where customers can describe what they want in natural language instead of searching for products one by one.
Now we're taking it a step further. Artificial intelligence is moving beyond being just an assistant that answers customer questions and is transforming into an agent that takes part in the entire shopping process.
Let's consider a simple example from food retail. A customer logging into a supermarket chain's mobile app might make the following request: "I want to plan my family shopping for four this week. I want two quick-to-prepare meals, school snacks for the kids, and a Sunday breakfast. I prefer lactose-free products and my total budget shouldn't exceed five thousand TL. I already have pasta and olive oil at home."
In today's e-commerce experience, customers need to solve their needs by searching across different categories, comparing products, and creating their shopping carts item by item. A new generation shopping agent, however, can understand the customer's needs, consider their past preferences, identify suitable products and alternatives, evaluate stock availability and promotions, and create the shopping cart. If a product is out of stock, they can also suggest an equivalent product that matches the customer's preferences.
The change here isn't just about a better user interface. The real change is that artificial intelligence is starting to get involved in the purchasing decision.
Anthropic's commerce agent blueprint, announced on September 2nd, is a noteworthy development in this regard. Anthropic has released an open reference architecture to enable companies to develop their own commerce agents. The structure includes two main agents: a "shopping agent" that serves the customer and a "merchant agent" that assists the retailer's own teams. The shopping agent covers functions such as product discovery, comparison, shopping list creation, adding items to the cart, and customer service; while the merchant agent focuses on sales analytics, inventory tracking, price recommendations, and campaign preparation. Interestingly, functions like campaign preparation are designed to require human approval.
I think the most important point to emphasize here is that artificial intelligence is starting to integrate not only into the customer side of shopping, but also into the retailer's decision-making processes.
While the agent on the customer side is trying to answer the question "What should I buy this week?", the agent on the retailer side is dealing with questions such as "Which products are at risk of stock shortages?", "Which campaigns didn't yield the expected results?", and "Which products should we take price or promotional action on?".
Therefore, in the future, evaluating the shopping experience solely through AI assistants offered to the customer will not be sufficient. The level of integration of the product, customer, stock, price, campaign, and supply data that fuels this experience, and how quickly this data can be translated into decisions, will become far more important.
Another aspect of Anthropic's approach that caught my attention is the emphasis placed on keeping the customer relationship with the retailer. The shopping agent works within the retailer's own website or mobile application. The customer can continue their product research and shopping within the retailer's own channel without moving it to a third-party AI platform. The situation is similar on the payment side: payment institutions like Visa and Mastercard accompany this structure through programs that provide tokenized identity and spending rules instead of giving the agent raw card information. In other words, the checkout and payment security layer is not in the hands of the model provider.
From my reading, Anthropic is positioning Claude here not as a store or payment platform, but as an intelligent layer that the retailer can integrate into their own systems.
This issue will become even more important in the coming period. Because on one hand, general-purpose AI platforms like ChatGPT and Claude are poised to become personal shopping assistants for consumers. On the other hand, retailers will want to develop their own shopping agents while protecting their customer relationships and data.
I believe this impact will be even stronger in food retail. This is because shopping is high-frequency, significantly repetitive, and customer preferences can be learned over time. When information such as the brands the customer uses, dietary preferences, consumption habits based on family structure, price sensitivity, previously accepted substitute products, and shopping frequency is used correctly, the AI agent can increasingly become a better shopping assistant.
However, technology alone will not be enough.
The real value lies in understanding customer intent and combining it with the retailer's available data and operational decision-making mechanisms. Recommending a product to a customer may become easier. However, ensuring that the product is in the right store, anticipating stock risks, making appropriate pricing and promotion decisions, and identifying alternative products when necessary remain fundamental decision-making challenges for retailers.
The agent learning the company's language.
There's an often overlooked but crucial point here. For a trade agent to truly enter a company's life, they need to connect to that company's language—that is, the layer of meaning in its data model.
Imagine asking a merchant agent, “Which campaign didn’t deliver the results we expected?” For the agent to answer this question, they first need to know what “results” means at that company. Is sales measured as gross or net? Are returns deducted? Are consignment items included? Which base period is chosen for comparison? Are the cannibalization and halo effects of the campaign on other products accounted for? None of these are things the model can automatically figure out; they are company-specific definitions. If the agent starts generating queries based on raw data, the resulting answer will be grammatically perfect but functionally wrong. Worse, the inaccuracy won’t be easily noticed because the answer will seem reasonable.
The situation is no different on the customer side. “Lactose-free” isn't just a label; it's a whole network of relationships. It requires knowing which product group it belongs to, which product it can replace, and which customer has previously accepted which substitute. If these relationships aren't defined, the agent starts to guess.
Therefore, the human approval requirement should be read not as a solution, but as an indicator. The weaker the layer of meaning behind the agent, the more decisions you have to send back to humans. If the corporate metrics are defined, business rules are established, and the relationships between the product, customer, and store are modeled, then a much wider decision-making scope can be delegated to the agent. In other words, the semantic layer and ontology are the main factors determining how autonomously artificial intelligence can operate.
This brings us to three very concrete points.
Making catalog data understandable to the agent. Today, most retailers have product data prepared to be read by the human eye. For an agent to fulfill a request such as "lactose-free," "quick-prepare," or "school snack," information on attributes, ingredients, diet, and allergens needs to be available in a structured form. Without this data, the quality of the model becomes irrelevant.
Defining substitute product relationships as data. The phrase "If it's out of stock, it suggests an alternative" sounds like it's the model's job; however, it's an infrastructure issue. Knowing which product replaces which, for which customer, and under what conditions requires modeling product relationships. Product infographic studies represent one of the most concrete areas of progress in this field today.
Taking store-based availability and stock accuracy seriously. If an agent sees incorrect stock at the store level, the experience is ruined on day one. A customer won't trust an assistant a second time if they recommend a product and then the assistant takes it back saying "unavailable."
Therefore, AI-powered shopping shouldn't be seen simply as "a smarter chatbot." Artificial intelligence is now not only influencing customers' purchasing decisions, but also becoming a part of retailers' decisions to meet those demands.
I think the competition going forward will take shape right here.
Retailers who will have an advantage are not those who use the best AI model, but those who best connect the customer, data, decision-making processes, and operations.
>> Read the article at Retail Turkey
Source: Anthropic, “Building commerce agents with Claude”, September 2, 2026 — claude.com/blog/claude-for-commerce-agents