Summary
Google is testing a Gemini feature that can call businesses on a user’s behalf, while NIQ says 51% of U.S. consumers used at least one AI shopping tool in the past month. The shift suggests that commerce is moving from AI-assisted discovery toward supervised agent execution, bringing authorization and payment controls into sharper focus.
AI is moving from advice to execution
Artificial intelligence has been part of online shopping for years, but most consumer applications have remained on the advice side of the transaction. Chatbots can help people search for products, compare specifications, summarize reviews and narrow a list of options. The consumer, however, still has to open a website, complete a form, make a phone call or submit the final request.
Google’s latest experiment points toward a more active role for agents. According to TechCrunch, Gemini’s “Call for Me” feature can contact businesses on behalf of a user and handle tasks such as checking whether a product is in stock, making a restaurant reservation, moving an appointment to another date or asking a store to hold an item. The system can introduce itself, navigate automated phone menus, wait on hold and conduct a conversation with the business.
The significance is not simply that an AI system can place a phone call. It is that the agent is beginning to interact with the physical and operational layer of commerce, rather than only interpreting information already available online. Many businesses still handle inventory questions, appointments and service changes through telephone channels. If an agent can operate those channels with the user’s approval, the path to commerce no longer ends at a search result or an app interface.
Google’s design also shows that the industry is not yet treating this as fully autonomous commerce. The feature is initially being tested with U.S. Pixel 11 owners who pay for a Gemini subscription and use a beta version of Google’s Phone app. Users can approve the personal information that Gemini shares during a call, watch a live transcript and take over at any time. These controls frame the product as supervised delegation rather than an unrestricted digital representative.
AI shopping is becoming mainstream, but delegation is not complete
The consumer behavior needed to support agentic commerce is expanding. NIQ’s Agentic Commerce Tracker found that 51% of U.S. consumers used at least one AI tool to assist with shopping during the past month. It was the first time the measure exceeded half of the U.S. market in the tracker, which covers multiple stages of the purchase journey, from product discovery to the decision to buy.
Product recommendations were the most common use, followed by personal shopping assistants. Consumers primarily use AI to compare alternatives, evaluate value and reduce the number of options they need to consider. NIQ described the trend as AI compressing decision-making rather than replacing shopping altogether.
That distinction matters. A consumer who trusts AI to identify a suitable product may not be willing to let it accept a substitute, use a payment method, change a delivery address or agree to a merchant’s cancellation policy without additional confirmation. Recommendation and execution carry different levels of risk. A poor recommendation can be ignored; an unauthorized payment or a mistaken appointment change can create financial, contractual or customer-service consequences.
The likely near-term path is therefore incremental. Agents may first receive permission to perform bounded, relatively reversible tasks: checking inventory, collecting quotes, requesting available appointment slots or summarizing order status. More consequential actions, including confirming a purchase, changing payment details or accepting terms, are more likely to require explicit approval.
The missing layer is workflow infrastructure
The growth of consumer-facing AI does not mean that merchants and e-commerce operators have already built autonomous businesses. Reporting from The Next Web highlights a persistent problem in e-commerce operations: companies have accumulated many software tools and AI copilots, but daily work remains fragmented and manual.
One system may identify a problem with a product listing. Another may recommend a change. A human still has to decide whether the recommendation is appropriate, apply it in the relevant platform and verify that the change worked. Large sellers and agencies may manage operations across marketplaces, advertising systems, inventory tools, customer-service platforms and analytics dashboards. The presence of multiple AI features does not automatically connect those systems into a coherent operating process.
For an agent to move from assistant to operator, it needs more than a language interface. It needs a reliable infrastructure layer that can connect models with commerce systems while managing identity, permissions, tool access, state, exceptions and audit records. The agent must know not only what action would be useful, but also whether it is authorized to take that action and how to stop when an unexpected condition appears.
This is why adding a chat window to existing software is unlikely to be enough. An AI system that can generate a recommendation is not necessarily capable of changing inventory, modifying an appointment or submitting an order safely. Businesses will need explicit policies that define which tasks can be automated, which require a second confirmation and which must remain with a human operator.
The same issue appears on the consumer side. A user’s instruction is rarely a single, unlimited permission. “Find out whether the item is available” does not necessarily authorize the agent to buy a different item. “Move my appointment” may not authorize a higher price or a different service. Systems must preserve the boundaries of the original request instead of treating general intent as permission for every related action.
Authorization becomes the core transaction problem
As agents act for users, authorization becomes more complex than a login session or a one-time confirmation button. A robust system needs to distinguish between actions the user explicitly approved, actions the agent may decide within a defined range and actions that require a new request for permission.
Google’s decision to let users approve the personal information shared during a call, observe a live transcript and intervene reflects this challenge. Visibility and human takeover create important control points. They do not, however, solve every problem. Users may not continuously monitor a call, and a transcript may not fully explain a merchant’s policies, fees or operational constraints.
Agentic commerce will likely require more granular controls: task-specific authorization, spending and time limits, merchant or category restrictions, isolation of sensitive information and records for each tool call. The system should also make clear whether an agent is asking a question, making a commitment or initiating an irreversible action.
Those requirements are particularly important for payment accounts, corporate funds and institutional wallets. Whether an agent can initiate a payment, alter a beneficiary, change settlement instructions or accept a commercial term should not be determined solely by a model’s interpretation of natural language. Such actions need enforceable policies, verifiable permissions and, where appropriate, human approval. A wallet or payment system designed for agentic use would also need clear separation between the agent’s operating authority and the user’s underlying assets.
Auditability matters as much as prevention. When an agent makes a mistake, a merchant and a user need to determine what instruction was given, what data the agent relied on, which permissions were active and why a particular action was taken. Without that record, disputes become difficult to resolve and trust is likely to remain limited.
Merchants will have to become legible to agents
For brands and retailers, the transition may change how products are presented and discovered. NIQ’s findings suggest that product content, structured data and the ability to be accurately surfaced by AI could become as important as conventional search visibility and physical shelf placement.
Agents need more than attractive marketing copy. They need current product specifications, price and availability information, delivery coverage, return rules, restrictions and compatible alternatives in formats that can be interpreted consistently. If the data is incomplete or stale, an agent may give a misleading recommendation or take an action based on conditions that no longer exist.
Merchants will also need ways to recognize and interact with authorized agents. That does not necessarily mean treating every automated request as valid. Businesses will need to verify that an agent is acting for a real user, understand the scope of its authority and provide a reliable path to human support when a request falls outside automated rules.
This creates a two-sided infrastructure challenge. Consumers need controls that prevent agents from exceeding their mandate. Merchants need standards that make agent requests identifiable, accountable and operationally useful. Payments, identity, customer service and dispute processes will increasingly have to accommodate interactions in which the human principal is not directly present at every step.
The transition will be gradual, but the direction is clear
NIQ’s consumer data shows that AI-assisted shopping has moved beyond a narrow early-adopter audience. Google’s phone-calling experiment shows that platforms are testing how far that assistance can extend into real-world execution. At the same time, the operational experience of e-commerce companies suggests that the industry still lacks the system connections needed for agents to run complex workflows reliably.
The immediate future is therefore unlikely to be a world in which consumers hand over every purchase to autonomous software. It is more likely to involve supervised agents that perform bounded tasks, ask for confirmation at defined checkpoints and keep a detailed record of their actions. The boundary between recommendation and execution will become a product and policy decision, not merely a technical one.
For commerce platforms, payment providers and wallet infrastructure, the competitive question may shift from who offers the most persuasive AI interface to who can make delegated action safe, predictable and auditable. The winning systems will need to manage identity, authorization, data access, payment controls, exception handling and human intervention as a unified process.
AI is therefore entering commerce in stages. It first helped consumers find information, then helped them evaluate options, and is now beginning to communicate with businesses and carry out limited tasks. Whether that progress becomes genuinely autonomous commerce will depend less on the ability to generate fluent responses than on whether agents can operate within clearly defined permissions and remain accountable for every consequential action.
Source: link