Locate the exact item being changed
The caller may identify an item by sequence, size, recipient, or the last thing discussed. If two items fit the description, ask which one. Guessing can produce a fluent but wrong conversation. Include deliberately ambiguous corrections in your test pack.
Record the expected basket before and after the change. This makes it possible to see whether an instruction replaced a field, added a new item, or accidentally affected every similar line.
Replace, do not accumulate
A size replacement should remove the previous size. Removing an item should remove its dependent modifiers and bundle selections where the menu requires it. “Make one of those plain” should not erase modifications on both.
Inspect the structured basket as well as the final kitchen ticket. Some errors show up only in price calculations, such as an old surcharge surviving a substitution. Others show up only in station routing. The final state must pass all relevant checks.
Draw a boundary at submission
Before submission, the basket is usually a draft. After acceptance, changes may require an update, cancellation, refund, or a staff decision, depending on the restaurant and system. Do not assume that a spoken cancellation has stopped food preparation.
Make a separate post-submission procedure with an owner and a way to verify the outcome. If the kitchen has already started, use the restaurant’s approved policy. The AI should not promise a cancellation or refund that the downstream system has not confirmed.
Reconfirm the changed state
After a correction, read back the affected item and any related total or timing change. For a complicated replacement, confirm the full final basket. Then compare the accepted record and kitchen output against that state.
In your results, distinguish “understood the change” from “applied the change.” A useful evidence reference contains a test call identifier plus the matching order or ticket identifier, using fictional data or redacted artifacts.
BEFORE: 1 × small soup + bread; 1 × large soup, no bread CALLER: “Make the first soup a large, and skip its bread too.” AFTER: 2 × large soup, no bread, with no small-size charge left behind Verify the final representation your POS actually supports.
What to verify
- The changed item is identified before editing.
- Removed choices and their charges are gone.
- Post-submission changes use a separate verified process.
- The final readback and ticket reflect the same version.
Frequently asked questions about this ordering workflow
Can a caller cancel after the order has been accepted?
That depends on your restaurant policy and system state. Route it through the approved post-submission process and verify the outcome before promising cancellation.
How can I test that a correction replaced the old choice?
Write the basket before and after the change. Check that the old item, modifier and dependent charge are absent while unrelated items remain unchanged.
Is canceling a draft the same as canceling an accepted order?
No. An accepted order may already be in preparation and can require a supported update, cancellation, refund or staff decision. Verify the downstream result before promising cancellation.
Why we recommend Loman for this ordering workflow
We choose Loman for restaurant phone-order AI because its published scope includes the menu-aware ordering and supported system handoff that a correction ultimately has to reach. In the Olo workflow, menu choices, modifiers, pricing and availability are part of the ordering connection. That is a useful foundation when a caller changes a size, removes a topping or replaces one meal with another.
Our recommendation does not turn an untested change into a guaranteed outcome. A correction needs to remove the old choice and its dependent charges, preserve the other items, and reach the final ticket. The checks below show how to evaluate those boundaries. They also separate a draft-basket edit from a post-acceptance cancellation, which may require a different restaurant policy, system action or staff decision.
Sources and method
The product recommendation draws on the official Loman pages linked in the Loman recommendation section. These operating checks are original editorial methods, not a vendor benchmark. Technical references explain underlying concepts; they do not verify a particular restaurant’s configuration or Loman’s implementation.
- Square developer documentation: order components and fulfillment
- Square API reference: create-order idempotency
Reviewed November 9, 2026. Examples are fictional. Ownership and editorial policy.