AI agent for importing cars from the US

Where this started

Buying at US auctions and running the customs and logistics after it. A person assembled the document pack for every car, which took three to four hours, a mistake in a declaration cost a fine or a delay, and the whole process ran on which experienced person was handling it.

Who we built it for

A company buying cars at US auctions and running the customs and logistics itself, where the process depended on which experienced person was handling it.

What fought back

1. 
   01

   Filled forms, not written text

   The output is a customs declaration and a set of figures, not a paragraph. The agent fills predefined fields under a strict schema and computes the duties, so there is no free text anywhere it could drift.

2. 
   02

   The cost of being wrong

   A mistake in a declaration is a fine, a delay or an overpayment, so a model that invents is not an inconvenience here, it is the whole risk. Critical fields are validated, and a person confirms before anything is filed.

3. 
   03

   One car, many stages

   From the bid to the client collecting the car, every stage has its own data, its own documents and its own status. The deal carries all of it with a date and an author on every change, which is what makes the state auditable rather than remembered.

What we actually built

The agent does the broker routine rather than assisting with it: it pulls the lot data, prices the whole thing out so the total is broken down into car, shipping, duty and fee, fills declarations through structured output against a strict schema, generates the pack and carries the deal through its stages. A manager confirms before anything is filed, and every change is stored with its date and its author. Telegram and a web panel are the interface.

How the work went

1. Week 1

   The process taken apart into stages, and what data and which documents each one needs.

2. Week 2

   A prototype agent: lot data, a preliminary all-in cost, a draft declaration.

3. Week 3

   Document generation finished, field validation, the status model.

4. Week 4

   Integrations, notifications, the manager interface, tested on real deals.

5. Week 5

   Live, then the corrections that only real edge cases ask for.

What came of it

The document pack went from three or four hours to twelve or fifteen minutes per car, and errors in declarations all but disappeared. Built in four to five weeks.