A person can start an order by typing what they want to eat. For the restaurant, that conversation becomes a concrete promise: the right price, availability, dietary constraints, delivery and timing. The interface may be new; the operation behind it is still real.
On July 1, 2026, Square announced a ChatGPT app and Claude plugin to help eligible sellers appear and receive orders in those experiences. For US food and beverage sellers with an activated Square Online Ordering profile, customers can discover businesses, browse menus and place orders through Order by Cash App. Orders route into the seller’s existing Square Online Ordering setup, including its POS and Kitchen Display System. Square says these orders carry no additional marketplace commission. This is a specific integration for eligible US sellers; it does not establish availability for restaurants in Colombia.
Conversation is another entry point into the system
The announcement is often read as an interface change. The architectural change is more useful: an agent becomes another client of the restaurant’s systems. To answer well and create a valid order, it needs information that should already live somewhere reliable: menu, prices, hours, availability, delivery zones, allowed modifications, payment methods and rules for exceptions.
If each channel keeps its own copy of that information, the business ends up operating several versions of itself. An update may reach the website but miss the conversational channel. The result is a confident answer that offers an unavailable dish, an old price or a delivery promise the kitchen cannot keep.
That is why an AI ordering project starts with a trusted operational source and a reliable integration. The POS needs to receive items and modifications in a structured form; inventory and hours need to reflect reality; and failed payments, out-of-range addresses or sold-out items need defined paths.
“Direct” describes one part of the journey
When an order reaches a restaurant’s POS, it no longer needs to follow the same flow as a traditional marketplace order. That can change the fees and the operation. Still, the word “direct” does not answer who processes payment, retains customer data, provides support or can contact the customer again.
The economics also need careful comparison. On its US pricing page, DoorDash lists delivery commissions of 15%, 25% and 30% for its Basic, Plus and Premier plans; pickup carries a 6% commission, while DoorDash Online Ordering carries no commission but does charge payment processing. DoorDash says marketplace commissions cover listing and marketing, logistics, customer support and card processing. Those rates apply to its US plans; they are not universal fees.
Square, in turn, says eligible orders from ChatGPT and Claude have no additional marketplace commission. That does not mean the whole order is free or that the restaurant stops paying for payment processing, preparation, delivery and service. The useful question is how much margin each route leaves after covering what it actually provides.
The POS matters more than a convincing demo
A conversational experience can feel smooth and still fail at the points that determine trust: a missing modifier, a duplicate order after a retry, an allergy note that never reaches the kitchen or a refund with no clear owner. The system should handle those cases before the channel reaches more customers.
The architecture needs to establish what the agent can confirm automatically and when a person must step in. It also needs to define how modifications reach the kitchen, how order status is updated, what happens when the integration fails and which customer data is stored with permission.
How to evaluate the channel before scaling it
- Unify commercial information. Maintain a trusted source for products, prices, ingredients, availability and hours.
- Test the full flow. Exercise orders with modifications, sold-out items, failed payments, address changes, cancellations and retries.
- Set boundaries and owners. Define what the agent can resolve, what needs human approval and who receives each exception.
- Calculate cost by route. Include commission, payment, preparation, logistics, support, errors and repeat orders; volume alone does not determine profitability.
- Define the customer relationship. Clarify what information reaches the restaurant, what consent permits its use and which channel supports the next interaction.
A marketplace can still be valuable for discovery and coverage. A direct channel can make sense for repeat customers and a more controlled brand experience. A conversational integration can open another door. The decision depends on the costs and capabilities the full system can sustain.
ChatGPT orders are no longer a distant hypothesis for some US sellers. The lesson for any business evaluating AI agents is operational: conversation creates value only when it can read current information, execute a valid transaction and make clear who is responsible for the service after the order is confirmed.
SOURCES & REFERENCES