Case study · Restaurant commerce platform · October 2026
Ordering from a restaurant inside ChatGPT
In three days, a team at a restaurant commerce platform went from a planning document to a working ChatGPT app where a guest orders from a restaurant in plain language. In testing on one brand's sandbox menu, an AI model built 20 of 20 orders correctly and never invented a menu item.
The platform wanted to know whether a guest could order from one of its restaurant brands inside ChatGPT, on the ordering platform it already runs. The brief: one build for many brands, each with its own words and rules; real menus, prices and baskets; and no changes to any other team's code.
- From a planning document to a working app
- 3 days
- Test orders built correctly
- 20 of 20
- Target: 95% or more
- Invented items or options
- 0
- Target: 0
- Median tool response time
- 0.46 s
- Target: 2 to 5 s
Tools
The solution
The team built one app per brand, all from a single build, between ChatGPT and the platform's ordering API.
The conversation.
Agent layer
One app per brand.
Ordering API
Menus, baskets, tax.
Checkout page
Opened by a checkout link from the agent layer. It calls the ordering API, and ends before payment.
ChatGPT holds the conversation. The agent layer does the exact work: search, matching "scrambled" to a real menu option, required choices and prices. It refuses anything that isn't on the menu. Each brand is a profile of its own words, store and rules, and the screens follow ChatGPT's own style.
Results
The app built every test order correctly. An AI model (Claude, standing in for ChatGPT's own) played the assistant in 33 scripted conversations on one brand's live sandbox menu, covering required choices, vague requests, off-menu items and alcohol.
| Measure | Target | Result |
|---|---|---|
| Orders built correctly | 95% or more | 100% (20 of 20) |
| Invented items or options | 0 | 0 |
| Conversations passing every check | 90% or more | 97% (32 of 33) |
| Median tool response time | 2 to 5 s | 0.46 s |
The one miss: asked the price of an item with an add-on, the model added the two prices itself. Full replies took a median of 15 seconds in the test setup, mostly the model's own time. Speed inside ChatGPT was not measured.
What it showed
The platform's ordering API already does almost everything an AI assistant needs: the build required no change to it.
It also surfaced five platform gaps, each written up with a proposed fix. The main missing piece for launch is a checkout that can accept a basket started in an assistant.