Your next customer is an AI agent
When customers hand restaurant orders to an AI agent, photography and upsells stop working. What still decides the order is identity, reliability and evidence.
What happens to restaurant technology when customers hand their purchasing decisions to AI?
Picture telling your assistant: “Order from the burger place. Everyone’s usual, but get me something different. Have it here by 6:30.”
You never open an app. You never see a photograph, a promotion or an upsell. The agent knows what your family eats, picks something new for you and places the order.
For two decades, restaurant technology has assumed a person on the other side of the screen. We cut clicks, tuned conversion funnels and hired designers to make ordering feel effortless. Agents break that assumption. The customer is still there, but an intermediary that knows their preferences now stands between them and our software, and it ignores most of what we designed.

The interface loses its grip
Off-premise ordering takes the hit first. Restaurants have built digital journeys to shape behaviour: photography, product placement, upsells, promotions. The industry is excellent at this.
An agent sees none of it. The restaurant shows a photograph of its new burger, and the agent skips it. The restaurant suggests fries, and the agent knows the customer doesn’t want fries, so it declines without passing the offer along.
The restaurant keeps its product and loses its influence over the purchase.
The brand survives. Customers still care which restaurant cooks their dinner, and an agent acting on their preferences will respect that. On-premise holds up too. Kiosks moved ordering from people to machines years ago, and brand experience on a kiosk matters as much as ever.
Restaurants now need two front doors: a branded experience for customers who order for themselves, and a machine-readable one for those who delegate.
Who the agent chooses
Consider three requests:
- “Order Five Guys through Uber Eats.”
- “Order Five Guys.”
- “Get me a good cheeseburger.”
The first names the channel. The second leaves the channel to the agent. The third leaves the restaurant and the channel to the agent.
Discretion is different from impartiality. When the agent picks, it will pick a marketplace, and the reason is state.
Uber Eats and DoorDash already hold everything the agent needs: the customer’s account, saved payment method, delivery address and every past order. “Everyone’s usual” lives in that order history. Ordering direct means the agent creates an account, enters payment details and starts the history from zero. Each step is a chance to fail.
Familiarity compounds the advantage. The marketplaces dominate the web and the training data behind today’s models, so agents reach for them by default, the way coding agents reach for Supabase. Documentation, examples and repetition shape which tools an agent trusts.
The competition for direct orders is a competition for identity, payment and order history. A restaurant that wants direct agent orders has to hold those, or connect to wherever the customer already keeps them.
Reliability beats efficiency
A structured ordering API beats an agent scraping a website, reading the DOM and clicking buttons. It costs less, but the saving is small and goes to the wrong party.
Suppose an agent burns 50,000 tokens navigating a website and 2,000 through an API. At a blended $2 per million tokens, the API saves about 10 cents per order. On a $40 basket that is a quarter of a percent, and the saving goes to the agent provider. These figures are illustrative.
The advantage that counts is task completion: the odds that the agent gets the order right the first time. A documented, predictable integration removes ambiguity, and agents will favour whichever channel fails least.
The standards for this exist. Anthropic’s Model Context Protocol (MCP) gives agents a common way to call external systems. OpenAI and Stripe’s Agentic Commerce Protocol (ACP) and Google’s Agent Payments Protocol (AP2) cover checkout and payment. A restaurant that exposes its menu and ordering through these standards becomes something an agent can use without guesswork. The agent still has to discover and choose it, though, and a better direct channel loses to a familiar marketplace the agent never questions.
Agents could dismantle the marketplaces. They could just as easily entrench them, or create new intermediaries that sit above both.
Who owns the customer
The lost upsell is the smaller cost. The larger one is the relationship.
When an agent orders through a marketplace, the marketplace owns the data. When it orders direct, the restaurant still faces open questions. Whose loyalty account earns the points? Does the agent sign in as the customer or as itself? Who gets the receipt, the feedback request and the next offer?
Loyalty programmes exist to give restaurants a direct line to their customers, and the agent now sits on that line. Restaurants need loyalty and identity that an agent can carry on the customer’s behalf, so every agent order lands in the restaurant’s own record. Without that, agent ordering becomes one more channel where someone else owns the customer.
Merchandising to machines
An agent needs the same information a customer does: products, prices, modifiers, ingredients and availability. It gets and uses that information differently.
Ask for a photo of the burger and it should show one. Ask for your usual without onions and it should call an endpoint and skip the pages.
The commercial problem is harder. How does a restaurant promote a new product to an agent, or show that a bundle is good value?
Agents respond to claims they can check:
- A bundle states its saving against the à la carte price.
- A new item carries structured attributes the agent can match to its user: spicy, vegetarian, similar to past orders.
- A limited-time offer carries its terms and expiry in a form the agent can evaluate.
The agent decides what reaches its user, and it passes along the offers it can justify.
Restaurants will merchandise to two audiences. One responds to imagery, story and experience. The other responds to evidence matched against a customer’s preferences. Same products, different persuasion.
The other side of the counter
Agents will act for operators too.
Take a surge in demand. Prep times stretch, couriers arrive early and wait, and customers get the wrong delivery estimates. Today a manager spots this late and fixes it across three or four systems. An agent with access to order data, kitchen timings and courier dispatch could catch it early. It could extend prep estimates, hold dispatch and propose pulling slow items from the menu until the rush clears.
Operators will adopt this slower than customers. A wrong lunch costs one person a meal. A wrong price change across a few hundred restaurants costs real money. Agents will start by diagnosing and recommending, then take on routine execution as they earn trust. Permissions, audit trails and rollback are the preconditions.
Capabilities over screens
Restaurant software is organised around applications: menu management, order management, delivery management and reporting. The surge cuts across all four. An agent ignores those boundaries and goes for the capabilities underneath.
A menu system controls what a restaurant sells. An order system manages execution. A delivery system orchestrates fulfilment. Screens are one way to reach those capabilities, agents are another, and the software that serves both will win.
The restaurant still makes the cheeseburger, and the customer still has to want it. Everything between those two facts is up for redesign, and the restaurants that open it to agents will take the orders.