For over thirty years, digital ordering has followed a relatively linear trajectory: first restaurant websites, then apps, and finally major delivery platforms. Each step added an interface between the venue and the customer, but consumers continued to choose where to browse the menu, which channel to open, and how to complete the purchase.
AI-powered agents introduce a different possibility: delegating even those choices. Not an assistant that points to a restaurant or summarizes reviews, but software that receives an instruction, searches the web, selects items, and guides the order all the way to checkout. In a test reported by Restaurant Business, Meta's agent, Muse, successfully placed an order at Pizza Hut using its built-in browser and navigating the chain's website just as a human would—clicking, scrolling through pages, and filling out the required steps.
It is still an early experiment, not proof that purchasing behavior has already shifted. But it signals a potentially significant transformation for the food service industry: the entry point to a restaurant might no longer be the website, the brand's app, or a platform like DoorDash and Uber Eats. It could become a conversation with a personal agent.
From PizzaNet to delegated purchasing
The history of online restaurant ordering helps capture the scale of this shift. In 1994, Pizza Hut tested PizzaNet in parts of Santa Cruz: a bare-bones web service that allowed users to browse a menu, place an order, and provide an address and phone number. The store would then contact the customer to verify the request, and payment took place upon delivery. It was an incomplete system and less efficient than a phone call, according to reports at the time, but it foreshadowed a direction that would soon solidify.
Today, restaurants' digital properties are genuine storefronts. Menus, images, loyalty programs, promotions, and recommendations serve to increase both order frequency and ticket size. On the other hand, third-party apps have turned delivery into an option accessible to a far broader range of establishments. In return, they have carved out a dominant position in restaurant discovery and customer relationships.
An agent changes the nature of intermediation. If the user simply types “pizza”, the software can gather the necessary preferences—location, items, any constraints—and handle the rest of the work. The promise is to remove friction: fewer menus to compare, fewer accounts to open, fewer steps to follow. But it also means that restaurants have fewer opportunities to guide the customer's navigation through their design, campaigns, and highlighted elements on the page.
The agent’s browser is the new playing field
A particularly significant element of the Pizza Hut case is that Muse did not need a dedicated integration with the chain. It used its own browser and interacted with the existing website. If this model spreads, restaurants would not necessarily need to build a separate connection for every agent to allow a customer to delegate an order. At the same time, a page designed for human reading and use would also have to be interpretable and usable by software acting on the user’s behalf.
For brands, the implications extend beyond technology. A digital presence could be evaluated on two fronts: its ability to directly persuade a human and its ability to allow an agent to execute a task effectively. Inconsistent information regarding menus, availability, hours, pricing, and customization options would become an operational hurdle, not just a flaw in user experience.
Delivery platforms, in turn, will not disappear simply due to the arrival of agents. They continue to offer a logistics network, a catalog, and payment and support tools. However, their role as a storefront could become less of a given if an assistant is tasked with searching and purchasing through the most suitable channel. The balance of power would hinge on what the agent is authorized to do, the sites it can navigate, and the preferences set by the consumer.
This dynamic could also affect proprietary apps. A brand that invested in bringing the customer into its own ecosystem might face a different kind of flow: the order still originates from the website or the app, but the customer has not necessarily seen the commercial promotions tailored for them. The value of owning a direct channel would remain, especially for data, loyalty, and margins, but the way that channel captures demand would change.
The primary obstacle is trust
The ability to take action is also what makes agents more delicate than conventional chatbots. To be genuinely useful, these tools may require access to personal information and sensitive services. In the reported case, Muse flagged an Amazon package delivery because the user had granted it permission to read their emails. Depending on the scenario, purchasing agents are designed to operate with personal accounts and financial data as well.
In the restaurant sector, a seemingly minor mistake can have concrete consequences: a wrong address, a duplicate order, an ignored customization, or an item selected that is incompatible with the customer's needs. The problem is not merely about model accuracy, but also about the ability to monitor what the agent has done, revoke permissions, and handle refunds or disputes once the purchase has been delegated.
For this reason, the barrier will not be solely technological. Many consumers might appreciate assistance in choosing items or filling out an order, yet remain far less willing to grant full autonomy over payments, email, and personal data. The benefit of saving a few minutes will need to be perceived as outweighing the risk of losing control. That threshold is higher in the food sector than in an informational search, because the operation triggers an immediate transaction that is often tied to a specific time of day.
A transition, not an automatic replacement
The agent landscape is moving fast. Until recently, dedicated shopping tools were largely a theoretical concept; over the past few months, products like Instinct, OpenAI's Dots, and Meta's Muse have arrived. However, their presence does not amount to mass adoption, nor does it prove that autonomous ordering will become the dominant behavior.
The phase now opening for the restaurant industry is therefore primarily one of observation and preparation. Businesses will need to understand whether agents generate additional orders, shift traffic between channels, and make it harder to present offers and high-margin products. They will also have to decide what information to make easily accessible and what rules to impose on automated interactions.
PizzaNet showed that the web could become a sales channel, even when the process was still cumbersome. Agents represent the next step: the web remains the infrastructure, but the one navigating it might not always be a person. For restaurants, chains, and delivery apps, the challenge will be maintaining visibility and commercial control over a purchase where the first point of contact could be software chosen by the customer.
