Your first AI project should solve one boring, frequent problem.

Look for repetition, not novelty

Choose a task that happens often enough for improvements to matter. Ask the person doing it to walk through several real examples. Record the normal path and the exceptions. If every case is fundamentally different, the first project may be too broad.

Choose an output someone can check

A draft quote can be compared with a price list. A completed intake record can be checked against the original enquiry. Define acceptance criteria before selecting a model. A persuasive paragraph is not the same as a correct result.

Check access before promising automation

Identify which systems hold the information and whether reliable integrations are available. Confirm who can authorize access. A process that depends on a changing third-party website needs different expectations from one supported by a stable API.

Write down the human decision

Separate preparing work from committing it. Name the person who approves messages, prices or changes. Test the boundary in the application as well as the prompt. Decide what happens when that person is unavailable.

Measure the whole job

Record the existing time per task, volume and correction rate. After a pilot, include the time spent reviewing outputs and fixing mistakes. An automation is useful when the complete workflow improves, not simply because one step became faster.