# Design the review queue before you train the model

Ask a team what their automation project needs and they talk about the model. Watch the project go live and the hard work turns out to be somewhere else.

Every working example we have seen ends in a fork. Clean cases flow through on their own. Doubtful ones go to a person. That person needs a screen, and that screen decides whether anyone uses the system.

Take refund handling. A model reads the request and the evidence, and a rule checks the original transaction and the refund policy. Requests inside policy get paid. Anything outside it goes to a person with the reason attached.

If the reason is vague, or the reviewer has to open three other tools, the queue backs up and the team goes back to doing it by hand.

So the questions to answer first are about people:

1. Who sees the doubtful case?
2. What do they see on one screen?
3. What does one click do?
4. What gets recorded, and under whose name?

Then measure. Count a week of real cases before the build, with minutes per case and the share of exceptions. Count the same week again after launch.

We cover seven examples, from invoice capture to inspection reporting, in [Intelligent Automation Use Cases](https://dev-station.tech/intelligent-automation-use-cases/).

Dev Station builds the integration and the review layer for operators in the US and UK. Send us one process and a week of its cases, and we will tell you which part is automatable: [dev-station.tech/contact](https://dev-station.tech/contact/).
