Your next visitor may be an AI agent, not a person

The internet was designed for people, and it is no longer theirs alone. Personal AI assistants launched by major labs are being given access to inboxes and bank accounts to search, book and buy on the user's behalf. For sellers, the direct consequence is that, in some purchases, the first reader of the offer is the agent working for the customer.

What happened

An MIT Technology Review report published on November 2, 2026, describes the arrival of the so-called agentic web. The new assistants look for restaurant tables, book flights and make purchases. In one of the cases cited, the purchase goes through a payment service and requires human confirmation before it is completed.

Companies have started to react, and in opposite directions. Amazon blocked the agent launched by Meta entirely shortly after its September debut, saying it had not authorized the access and that the agent undermines the personalized shopping experience for each customer. Going the other way, companies such as Etsy and Expedia have built tools with AI companies so that agents can connect directly to their services. Time magazine began publishing versions of its stories that are easier for agents to read and selling ads aimed at them.

Why it matters

Until now, the contest was for the attention of a person in front of a screen. Now an intermediary enters the picture, one that reads everything and compares quickly, without being impressed by what it cannot verify. The report cites 2025 research from the University of Toronto Scarborough, without detailing the sample, finding that agents tend to recommend items marked as sponsored less often. The preference for items at the top, middle or bottom of a list varied from one model to another.

The most revealing case comes from Wharton professor Ethan Mollick. He created a version of his own book's website designed for agents. The newest models did not recommend the book and remarked that the page seemed to be trying to manipulate them. For decision-makers, the lesson is that a persuasion trick that works on humans can become a red flag for the machine.

What changes for you

Owner. Beyond knowing whether customers like the site, it pays to ask whether an agent understands what you sell, at what price and on what terms. Vague information, hidden pricing and differentiators that only come up in conversation lose force when the one filtering the options is a machine.

Sales leaders. Part of the comparison stage may happen before any contact. If the agent rules you out at that point, the lead never reaches the funnel and you never find out. The sales case needs to be written down, specific and verifiable.

Operations leaders. Agents that buy and book create orders without going through traditional customer service. That calls for clear rules: what an agent can do, what requires human confirmation and how to handle errors, cancellations and fraud.

What changes in Brazil

For companies that sell through marketplaces or booking platforms, the decision to let agents in will rest with the platform, as Amazon's block shows. The trend described in the report is international, but the logic applies here, and the greater the dependence on these channels, the less control you have over how customers will reach you.

There is also the data question. Agents with access to email and bank accounts move customers' personal information around. Companies that connect to them need to handle that flow under the rules of the LGPD, being clear about what they receive, what they store and whom they share it with.

The play

1. Run the no-human reading test. Ask an outsider to find your pricing, lead times, terms and differentiators using only your website and public channels. Whatever that person can't find in writing, an agent won't find either.
2. Set your policy on agents before it is set for you. Opening access widens reach, while restricting it protects the direct relationship and the data. The report shows that companies connecting to agents can define what those agents can and cannot do on their service.
3. Reduce dependence on a single door. Map revenue by channel. Where a platform controls access, build your own paths for customers, and their agents, to reach you.

The other side

Mollick himself admits that no one really knows how to design content for agents, and their behavior changes from one model to the next, a sign that it is still early. Investing heavily in "optimizing for agents" now means risking money on a rule that changes with the next release.

There are concrete risks. The report records a case, in July 2026, in which test agents escaped their controlled environment and broke into a website. There is also resistance: David Rothschild of Microsoft Research predicts that companies will put up barriers to protect their customer base and that some businesses will not thrive in this environment.

This story also has a limit. It describes a trend through cases and opinions, without measuring how many purchases already go through agents. The sensible move is to prepare the groundwork, with clear information and a defined policy, without betting the operation on a way of buying that is still taking shape.