Summary
Google is testing a feature that lets Gemini call businesses on a user’s behalf, while new industry data suggests AI-assisted shopping has crossed into mainstream use in the U.S. Together, the developments point to a shift from AI as a recommendation tool toward AI as an active participant in commerce, communication and transaction workflows.
AI agents are moving from advice to execution
Artificial intelligence has been part of e-commerce for years, but much of its role has been advisory. Systems recommend products, generate listings, answer customer questions or highlight operational issues. The consumer or merchant still performs the important action: making the call, changing the appointment, checking availability or deciding whether a recommendation should be implemented.
Google’s latest experiment with Gemini pushes the technology further into execution. According to TechCrunch, the company is testing a feature called “Call for Me,” which allows Gemini to contact businesses on a user’s behalf. The initial rollout is aimed at Pixel 11 owners in the United States who pay for a Gemini subscription. Users also need the beta version of Google’s Phone app for Android, and the feature remains an experiment rather than a broadly established service.
Google says the agent can handle tasks such as asking whether a product is in stock, making a restaurant reservation, moving an appointment to another date or asking a store to place items on hold. It can introduce itself to a business, navigate automated phone menus, wait on hold and continue the conversation once a person answers.
That matters because telephone-based commerce remains difficult to automate through conventional web interfaces. Inventory questions, reservations and special requests are often handled by businesses that do not expose every function through an online API. An agent that can make a call and conduct a conversation is effectively extending software automation into a channel that has traditionally depended on human labor.
Supervision is becoming part of the product design
The proposed interaction model is not one in which the agent disappears completely after receiving a single instruction. Google says users can follow the call as it happens through a live transcript and take over at any time. The call will also be placed from the user’s own phone number.
Those design choices point to a supervised form of agentic commerce. The agent performs the repetitive work, but the user retains visibility and an emergency route back into the conversation. Google also says Gemini will be able to share personal information approved by the user as part of the call. This expands the range of tasks the agent can handle, but it also raises the importance of defining what has actually been authorized.
As an agent moves from finding information to making arrangements, the consequences of misunderstanding become more significant. Asking whether a product is available is different from agreeing to a purchase. Rescheduling an appointment is different from merely presenting alternative dates. A system therefore needs to distinguish between low-risk information requests and actions that create obligations, disclose sensitive details or could be difficult to reverse.
The available information does not show that “Call for Me” independently completes every kind of payment or purchase transaction. It does show, however, that Google is testing a broader model in which an AI system can represent a user in a real commercial conversation. That model creates practical questions for both sides of the call. Users need to know what the agent may say and do. Businesses need ways to recognize an automated caller and determine whether the request is genuinely authorized.
Google has experimented with AI-assisted calling before. Earlier demonstrations showed an assistant making a salon reservation, including attempts to make the conversation sound natural. Gemini later introduced “Ask for Me,” another feature that allowed the system to contact businesses. The newer experiment appears to emphasize not just conversational realism, but the completion of more complex, multi-step tasks using user-approved information.
Shopping behavior is already changing
The consumer side of the market is also moving in this direction. The Next Web, citing NIQ’s Agentic Commerce Tracker, reported that 51% of U.S. consumers used at least one AI tool to help them shop during the previous month. It was the first time the proportion had passed 50% in the tracker, according to the report.
Product recommendations were the most common use, reported by 20% of consumers, followed by personal shopping assistants at 16%. Consumers primarily use AI to compare options, evaluate value and narrow their choices. NIQ described this as AI “compressing decision-making” rather than replacing shopping altogether.
That distinction is important. A consumer may ask an AI system to compare products and still make the final selection personally. Yet even this earlier part of the purchase journey can significantly change how brands compete. If shoppers increasingly encounter products through AI-generated answers rather than a conventional search page, product descriptions need to be structured, accurate and understandable to machines as well as people.
NIQ said product content, discoverability by AI systems and structured product data now matter alongside shelf placement and search rankings. For brands and retailers, this means that the path to purchase is becoming less dependent on a single search interface. A product must be represented clearly enough for an agent to identify its attributes, compare it with alternatives and respond to a shopper’s specific request.
The survey should still be read within its stated limits. NIQ’s tracker is based on a monthly survey of roughly 500 U.S. consumers as part of its Quick Question research. The tracker also covers Canada, but the figures cited in the report were for the United States. The 51% result is therefore an indicator from a particular research methodology, not a universal measure of AI adoption across every market, demographic or shopping category.
The merchant side has a different problem
Consumer-facing agents are only one side of the transformation. The other is whether an AI system can operate an online store rather than simply assist a person who operates it.
The Next Web’s discussion with e-commerce industry veterans describes a persistent gap between the promise of AI and day-to-day operations. Agencies and large sellers may have many software products available, but those tools often remain disconnected. One application identifies a problem, another recommends a solution, and a human decides whether the suggestion makes sense, implements the change and checks the result.
This is the difference between a collection of copilots and an operational agent. A copilot can make a useful recommendation while leaving the workflow to a person. An operational agent would need to understand the business context, interact with multiple systems, execute a permitted change and verify that the change produced the intended result.
For e-commerce teams managing large product portfolios, the relevant systems may cover listings, inventory, advertising, orders, customer service and platform policies. If an agent cannot connect those pieces, it may add another dashboard rather than remove complexity. The next layer of progress will therefore depend not only on better language models, but also on infrastructure that can coordinate tools, preserve context and manage workflows.
That infrastructure must include controls. The system needs to know which actions are available, which require approval, how to record the decision, what to do when data conflicts and when to return the matter to a human. Automation is valuable only if it reduces repetitive work without creating new errors in inventory, customer communications or platform compliance.
Authorization and identity become core commerce infrastructure
Once an AI agent is allowed to call a merchant, change an appointment or make a request related to an order, identity and authorization stop being secondary technical details. The agent may use a customer’s phone number and approved personal information, but the merchant still has to decide whether the caller genuinely represents that customer and whether the requested action is binding.
For users, transparency will be just as important. A live transcript, clearly defined permissions, an accessible handoff mechanism and an auditable record can all help reduce the risk of unwanted actions. A broad instruction should not necessarily grant unlimited authority for every later step. Requests involving payment, sensitive data or irreversible commitments may require a more specific confirmation than a simple request to gather information.
The current source material does not establish that Google’s experiment resolves these questions. It does show that commercial agents are moving toward a model in which users delegate narrow, observable tasks rather than merely ask for information. That shift makes the boundaries of delegation a product requirement, not just a legal or security concern.
Taken together, Google’s phone-calling test, NIQ’s consumer data and the merchant-side discussion point to the same direction: AI is becoming an active participant in the commerce chain. It can help consumers discover and compare products, communicate with businesses and potentially complete selected service tasks. On the merchant side, it may eventually coordinate store operations across several systems.
The distance between a successful phone call and a reliable commercial workflow remains substantial. Agents must handle ambiguity, preserve authorization boundaries, expose their actions to users and give businesses practical ways to verify identity. Whether agentic commerce becomes durable infrastructure will depend less on how human the conversation sounds than on how clearly responsibility is defined when the agent acts on someone’s behalf.
Source: link