When AI Shops Badly, Who Pays?

The article analyzes the emerging phenomenon of 'agentic commerce' – shopping through AI assistants (Google, Microsoft, OpenAI), where the entire purchasing process from selection to payment takes place inside the conversation without ever visiting an e-shop. It describes the competing protocols (UCP, ACP), market fragmentation, the loss of merchants' direct contact with customers, and fundamental legal gaps regarding liability for AI errors, price discrimination, and personal data protection. Although analysts forecast a market worth hundreds of billions of dollars by 2030, consumer trust remains low and regulation (GDPR, the EU AI Act, DMA) is so far failing to keep pace with the new model.

The race to checkout inside conversations is rewriting the rules of e-commerce — and regulation can't keep up

A user types into a chat window: "I need wireless headphones under 2,000 crowns, mainly good battery life." Within seconds they get three specific products with prices, reviews and a "Buy" button. They don't click any link, don't visit any e-shop. They pay inside the conversation. The package arrives in two days.

This is exactly how Google, Microsoft and OpenAI imagine shopping in 2026. Starting in the autumn of 2025, all three gradually launched infrastructure in which the entire shopping process — from the first query to payment — takes place inside an AI assistant. And alongside them, dozens of other companies, from Shopify to Visa to Mastercard, are building the protocols, payment gateways and security standards for a world in which the customer of an e-shop is not a human, but an algorithm acting on their behalf.

Agentic commerce — as the new commerce model is called — fundamentally changes who controls the customer relationship, where advertising takes place, and who bears liability when something goes wrong. Current legal frameworks have no answers to most of these questions.

Before we dive into the technical details and the strategies of the big players, it's worth asking a basic question. Not everyone shares the technology companies' enthusiasm. Forrester analyst Sucharita Kodali put it bluntly in an interview with GeekWire — e-commerce is not a problem that needs fixing. It's still unclear what added value AI-mediated purchases bring, if we set aside the disintermediation of Google.

According to a ChannelEngine report from January 2026, based on a survey of 4,500 consumers in the US and Europe, only 17 percent are willing to let AI complete a purchase on their behalf. The main barrier? Trust and transparency. Paradoxically, 58 percent of consumers already use AI tools to research products — but between research and payment there is a chasm.

According to an Adobe survey, traffic to retail websites from generative AI tools grew by 693 percent year over year during the 2025 holiday season. Interest is growing exponentially — but trust is not, at least not yet.

Even so, all the major platforms are betting that the gap will close. And they are building the infrastructure as if it already had.

All three players have chosen a different strategy. The differences are not merely technical — they reflect a fundamentally different relationship to both the consumer and the merchant.

Google is betting on an open standard. The Universal Commerce Protocol (UCP), announced on January 11, 2026 at the NRF conference, defines how an AI agent communicates with an e-shop: it browses the catalogue, negotiates checkout parameters, applies a discount and completes payment through a standardized API. UCP was developed by Google together with Shopify, Etsy, Wayfair, Target and Walmart. More than twenty additional partners — including Visa, Mastercard, American Express, Stripe and Adyen — have endorsed the protocol. A merchant publishes a profile on their website at the address /.well-known/ucp, where they declare what their system can do: basic checkout, coupons, loyalty points, order tracking. The agent reads the profile, negotiates the intersection of capabilities, and carries out the transaction.

UCP orchestrates three existing protocols: A2A (Agent2Agent, developed by Google, managed by the Linux Foundation) for communication between agents, MCP (Model Context Protocol, developed by Anthropic) for connecting to data and tools, and AP2 (Agent Payments Protocol) for cryptographically verified payments. To this Google adds advertising directly within the conversation of its AI Mode, which has more than 75 million daily users (figure from December 2025, Search Engine Journal / Nick Fox, Google).

Microsoft has chosen the path of trust. Copilot Checkout, launched on January 8, 2026, builds on existing enterprise relationships with retailers. Unlike Google and Amazon, Microsoft is not a seller — it doesn't compete with e-shops for customers. It is betting that merchants will entrust it with their data more readily. Its partners are PayPal, Shopify, Stripe and Etsy; among the first brands are Urban Outfitters, Anthropologie and Ashley Furniture. According to PayPal data from the launch press release, the probability of a purchase within thirty minutes of an interaction with Copilot is 53 percent higher than without it. When purchase intent is present, conversion rates are 194 percent higher (PayPal Newsroom, January 8, 2026).

OpenAI started first. It launched Instant Checkout in ChatGPT on September 29, 2025 in cooperation with Stripe. Through Etsy (immediately) and Shopify (a gradual rollout) it has access to more than a million merchants. With roughly 700 million weekly ChatGPT users (OpenAI figure from September 2025) it leads in reach. But conversions still lag — according to a report by The Information, some key partners are still waiting for integration details. At the same time, OpenAI developed its own open standard with Stripe — the Agentic Commerce Protocol (ACP), a direct competitor to Google's UCP.

Alongside them, other players are entering the game. Amazon is testing "Buy for Me" — an AI agent buys a product from an external e-shop without the customer leaving Amazon. The feature, however, sparked a wave of criticism, because some brands were included without consent and with inaccurate listings. Visa is developing the Trusted Agent Protocol for the secure identification of AI agents in the payment chain. Mastercard has launched Agent Pay. Criteo offers a recommendation engine optimized for AI shopping assistants.

The analyst firm Bain & Company, in its report "Agentic AI in Retail" (November 2025, report behind a paywall), estimates that agentic commerce could account for 15–25 percent of online retail in the US by 2030, a market worth 300 to 500 billion dollars a year. Morgan Stanley, in its own analysis, forecasts 10–20 percent (190–385 billion USD). Both figures are projections with a high degree of uncertainty — real adoption does not yet exist.

The traditional model worked for decades: a user enters a query into a search engine, gets a list of links, clicks through to the merchant's website, examines the product, adds it to the cart, pays. Every step generated data, advertising, optimization opportunities.

In the new model, this entire journey disappears. The user describes a need in a conversation. The AI agent interprets the intent, searches the offerings, compares parameters and recommends a specific product. The purchase takes place with a single click inside the chat interface. The e-shop ships the package, but never saw the customer on its website. Semrush data show that 92–94 percent of interactions in Google AI Mode end without a single click to an external website. That is roughly double the rate of the traditional SERP (search engine results page), where zero-click reaches 35–46 percent.

For merchants this means losing direct contact with the customer. Loyalty programs, cross-sell offers, retargeting — everything that depends on the customer visiting the website — loses effectiveness. The AI platform controls the interface, the data and the context. The merchant supplies the goods and handles complaints, but has no access to the full breadth of the customer's purchasing preferences.

The market has forced a new discipline into existence: AI Engine Optimization (AEO). Whereas SEO optimized web pages for search crawlers, AEO optimizes product data for AI agents. Structured attributes, descriptions matching real-world use scenarios, accurate availability and pricing information — all of this decides whether an AI agent recommends a product or skips it.

But even a merchant who adapts their feeds faces a further question: adapts them to whom? Each of the three platforms requires a different integration — and no common standard exists.

There are today at least four distinct paths into the AI conversational checkout. Google requires a UCP endpoint. OpenAI and Stripe require ACP integration — their own API, Shared Payment Tokens, a different architecture. Microsoft Copilot Checkout goes through existing partners: if an e-shop uses PayPal, Shopify or Stripe, it is "inside" with no further work. Amazon Buy for Me requires nothing at all — Amazon crawls the e-shop's website and buys like a user, but the merchant has no control over what the agent displays or orders.

UCP and ACP are both open-source, but they are competing ecosystems with different payment primitives. There is no interoperability between them. An e-shop that wants to be available through Google, ChatGPT and Copilot must either support both protocols, or bet on an intermediary that abstracts it away.

The only platform that today offers coverage of all three major players from a single integration is Shopify. It co-developed UCP with Google, for ACP a change reportedly on the order of a single line of code is enough according to Stripe, and Shopify merchants are automatically enrolled in Copilot Checkout. For an e-shop on a regional platform — whether that's the Czech Shoptet, the Polish Shoper, or any local alternative — the situation is radically different. None of them yet supports either UCP or ACP.

And even if an e-shop manages the technical integration, the operating costs remain. Each platform has a different liability model, a different chargeback resolution system, different display rules. The merchant must monitor whether their price in Google AI Mode matches their website, whether ChatGPT is not displaying outdated data from the feed, whether Copilot correctly interprets availability. For large retailers, manageable. For medium and small e-shops — the backbone of European e-commerce — it is a barrier that may de facto restrict agentic commerce to the big players and the Shopify ecosystem.

In AI Mode, Google is testing shopping ads that appear directly within the assistant's responses — labeled as sponsored, but visually integrated into the conversational flow. The Direct Offers format goes further: a merchant can offer a personalized discount or bundle to a customer whom the AI model deems ready to buy.

Vidhya Srinivasan, Vice President and General Manager of Google Ads & Commerce, summed up in a blog post on February 11, 2026 that Google is not merely adding advertising to the AI environment, but "reinventing what advertising is."

It is precisely the personalization of offers, however, that has sparked controversy. Lindsay Owens, executive director of the think-tank Groundwork Collaborative, labeled UCP as infrastructure for "surveillance pricing" — price discrimination based on conversational data, search history and retail information. Senator Elizabeth Warren subsequently, in early February 2026, sent a formal letter to CEO Sundar Pichai demanding an explanation of whether Google plans to use sensitive user data to manipulate consumers. She set February 17, 2026 as the deadline for a response.

Google responded with a ban: merchants may not set a higher price in AI Mode than on their own website. Direct Offers can only offer a lower price or added value (free shipping, loyalty points). Critics, however, object that the problem is not the price itself — it is that the AI decides, based on a profile, to whom it shows which offer in the first place. According to surveys by Morning Consult and Consumer Reports, a majority of Americans — between 59 and 66 percent depending on how the question is worded — are concerned about algorithmically driven price discrimination.

All the platforms agree on one principle: the merchant remains the so-called merchant of record — the seller listed on the receipt. The AI platform (Google, Microsoft, OpenAI) provides the interface and the intermediation, but is not a party to the transaction. The merchant bears full legal responsibility for the product — quality, price, delivery, VAT, consumer rights. The payment processor (Stripe, PayPal) handles the transaction technically, but is not liable for the goods.

The model is clear — as long as everything works. The problem arises the moment the AI agent makes a mistake.

The consumer says: "Order me running shoes for up to 3,500 crowns." The agent selects a model that doesn't fit. Who bears the responsibility? The current legal framework says the consumer — they delegated the decision to the agent, that is, they acted through it. The merchant fulfilled the order as it was received. But consumers hardly perceive it that way. The card schemes won't assume liability, the customers and their banks won't either, and the AI model technically earned no money. That leaves the merchant — even though the mistake was made by a third party's algorithm.

The problem is multiplied in the case of so-called friendly fraud — a situation in which a consumer disputes a legitimate transaction. According to Mastercard data cited in the Chargeflow State of Chargebacks Report, friendly fraud accounts for 75 percent of all chargebacks (the figure may relate to a specific segment). With AI agents the risk deepens: a consumer can claim that "the agent bought it on its own" or "it ordered something I didn't want." Existing chargeback resolution systems do not distinguish between fraud, an AI error, and an inaccurate user instruction.

The debate over liability covers the AI agent's errors, but there are also deliberate attacks. In a so-called prompt injection attack, a manipulated product description or web page can induce the AI agent to behave in undesirable ways — for example, to recommend a more expensive product or bypass the user's preferences. An AI agent can also "hallucinate" — recommend a product that doesn't exist, or assign parameters that do not match reality. In traditional e-commerce the customer examines the product themselves. In agentic commerce they delegate this check to the AI — and thereby accept a risk that did not previously exist.

Google UCP contains GDPR mechanisms at the technical level. The Agent Payments Protocol (AP2) requires cryptographic proof of the user's consent to each transaction. GDPR and CCPA signals are a mandatory part of the data messages. Authentication runs over OAuth 2.0.

But the problem lies deeper. If an AI agent decides, on the basis of a user profile — search history, conversational data, previous purchases — which product to show the user and at what price, this is profiling with an economic impact. Article 22 of the GDPR grants the consumer the right not to be subject to a decision based solely on automated processing, where it has legal or similarly significant effects. An AI agent's purchase recommendation approaches that threshold.

The EU AI Act complements the GDPR with a risk-based approach. AI shopping agents are not explicitly classified as high-risk systems. But if an agent actively influences a purchasing decision and mediates a payment, regulators may argue that it is a system with an impact on the consumer's economic rights.

The situation is complicated by active lobbying by technology companies. According to an analysis by the Corporate Europe Observatory, Google in an internal lobbying document asked the German government to introduce a "disproportionate effort" exemption for compliance with Articles 15–22 of the GDPR. Germany subsequently, in October 2025, submitted to the European Commission its own proposal for a broad simplification of the GDPR. The European Commission, as part of the Digital Omnibus proposal from November 2025, is considering softening the definition of personal data — pseudonymized data could fall out of the full scope of GDPR protection. The EDPB and EDPS sharply criticized this approach in February 2026, saying it would "significantly narrow the concept of personal data." If it nevertheless passes, it would substantially shift the balance of power between the platforms and the regulators.

The European Commission is, as of February 2026, conducting several parallel investigations into Google. A new antitrust proceeding from February 9, 2026 examines whether Google artificially inflates the prices of advertising auctions at the expense of advertisers. An investigation under Article 102 TFEU from December 2025 focuses on AI Overviews and AI Mode — whether Google uses publishers' content without proper consent or compensation. A proceeding under the DMA (Digital Markets Act) from November 2025 examines whether Google's "site reputation abuse" policy unfairly suppresses legitimate publishers in search results.

None of these proceedings, however, directly addresses agentic commerce. Regulation lags technology by months to years. Before the EU pronounces on the liability of AI shopping agents, millions of transactions will take place in a legal vacuum.

Further open questions are piling up. How will returns and complaints work when the customer never visited the merchant's website and communicated only with an AI agent? The EU Consumer Rights Directive grants a 14-day withdrawal period against the merchant — but the customer will expect a solution from the platform. How does one distinguish advertising from an organic recommendation in a seamless conversation, where the line between sponsored and unsponsored content disappears? And how does one prevent three platforms controlling a majority of purchases from creating a new form of price cartel?

The impacts of agentic commerce are not limited to the US. European e-commerce grew up on its own ecosystem: price comparison sites (PriceRunner in Scandinavia, Idealo in Germany, Ceneo in Poland, Heureka in Central Europe, Kelkoo in France), local marketplaces and specialized shopping portals. Many of them operate on a CPC model (cost per click): the e-shop pays for every customer who clicks a link on the platform and goes through to the merchant's website. If the customer buys directly inside the AI conversation of a global platform, no click occurs and the monetization model stops working.

The openness of UCP gives European players a theoretical opportunity to connect — e-shops that deploy an endpoint for Google's sake make their services available to other platforms as well. But openness does not mean neutrality. Google governs the protocol's governance, and in smaller markets — where AI Mode may not quickly reach a critical mass of users — e-shops may lack the motivation to implement it. Idealo, PriceRunner and other comparison sites are dealing with a similar problem, some of which have antitrust proceedings open against Google.

According to a Heureka Group survey from February 2026 of nearly six thousand respondents, 46.5 percent of Czechs use AI when shopping. Almost one in seven (14.8 percent) trusts an AI recommendation more than salespeople in a store (9.2 percent) or influencers (6.7 percent). Even so — the main trust still lies with other users' reviews (51.4 percent) and information from specialized e-shops (41 percent). The finding is consistent with ChannelEngine and Adobe data from other markets: consumers are curious about AI, but not yet ready to entrust it with their wallet. Local reviews, verified experience and knowledge of the local market remain key in purchasing decisions — and it is precisely in this data that the global platforms still have gaps.

The integration fragmentation described above amplifies this problem. A local e-shop without Shopify has no easy path into agentic commerce. Companies such as Lengow, ChannelEngine or Feedonomics (part of BigCommerce) are already responding in the market — their analysts unanimously identify demand for a middleware layer that automatically generates both a UCP and an ACP endpoint from an existing product feed and connects to a local payment gateway. It is an analogy to how feed management tools made Google Shopping accessible to smaller merchants too a decade ago.

The European regulatory framework, however, also creates opportunities. Data stored in EU data centers, GDPR-native processing without transfer outside the EU, and transparent AI decision-making — the user sees why the system recommended a particular product — become, in the context of agentic commerce, a differentiator, not merely a cost item.

The rules by which the game will be played are, as yet, written by no one. Not the regulator, not the legislator. Interest is growing exponentially — but trust is not. And in the gap between them, it is being decided who will control commerce for the coming decade.

Analytical report prepared as of February 16, 2026. Source data verified against primary sources: Google Developers Blog, Shopify Engineering Blog, Microsoft Source blog, OpenAI blog, PayPal Newsroom, Stripe Newsroom, Search Engine Journal (December 2025, Nick Fox / Google AI Mode), Bain & Company "Agentic AI in Retail" (November 2025, paywall), Morgan Stanley "Agentic Commerce Impact" (2025, paywall), ChannelEngine Marketplace Shopping Behavior Report 2026 (January 2026, n=4,500), Corporate Europe Observatory, Heureka Group survey (February 2026, n=5,945), Chargeflow State of Chargebacks Report (Mastercard data), Morning Consult / Consumer Reports (algorithmic pricing surveys), Senator Warren's letter to the office of Sundar Pichai (February 2026), UCP specification v2026-01-11 (ucp.dev), Semrush AI Mode study (September 2025), GeekWire (Forrester / Kodali citation). The situation is evolving on a timescale of weeks.

Methodological note

The concept, structure and editorial line of the article are the work of the author, who prepared the content outline, established the key theses and directed the entire creative process. Generative AI (Claude, Anthropic) was used as a technical tool for research, fact-checking and fleshing out the author's draft.

The author edited the outputs throughout, verified the key findings and approved the final wording. No part of the text was published without human review. All factual data were verified against the publicly available sources cited in the text.

The procedure complies with the requirements of Article 50 of EU Regulation 2024/1689 (the AI Act) on the transparency of AI-generated content. #poweredByAI