Applied AI case study

An AI agent that answers a restaurant's questions and never acts on its own.

The model turns a plain question into SQL and reads the restaurant's own written policies. Code decides what may run, keeps each restaurant to its own data and its own rules, and records every step with what it cost. A person approves every action before it happens.

See the scenariosView the code

Scenario 5 of 5: a dispute the model proposes, code checks, and the owner approves.

A dispute: looked up, checked, proposed, approved

Asked as Mekong Noodle House

File the dispute with DoorDash for order 203088.

What should happen

  • Proposes filing the dispute rather than claiming to have filed it.
  • Code looks the order up, confirms it is a DoorDash order, and that the 7-day window is still open, with the due date.
  • The proposal waits for a person. I approved it through the API, and the record is marked as simulated.

What happened

UsedSQLpolicycitations checked

I proposed the DoorDash dispute for order 203088. It is pending, and nothing has been filed until a person approves it.

The order is a completed DoorDash order from 2026-10-28, worth $114.66 gross, with no refunds recorded. DoorDash disputes have a 7-day window . The system check passed on both the order and the window, which is open until 2026-11-04.

I can't see packing photos in the data. Orders over $40 should have one, and the dispute depends on it showing the item in the bag. If no photo exists, the policy says to accept the chargeback and log it as a packing miss instead of disputing . Please confirm the photo before approving.

answered

  • $0.0179to answer, $0.0404 without the prompt cache
  • 9.3 s8.7 s in the model
  • 11steps, 15,196 tokens in, 1,002 out

Captured on October 31, 2026. Rerunning puts the same question to the agent live, for a few cents and about ten seconds. Each run is kept so you can flip between them.