Write the task, the check and the next action.

Describe a person completing a task

Write down what they provide, what they need to understand and what they do afterward. “Add an AI assistant” is less specific than “prepare a draft summary that a reviewer can compare with the source.”

State what counts as useful

Agree on the important fields, source support or format. Include information that should remain absent when the source does not provide it. Identify requests the product should decline or hand to a person.

Prepare permitted examples

Use synthetic or appropriately protected material with expected outcomes. Cover ordinary tasks, difficult variations and missing information. Keep evaluation examples separate from those used to develop prompts or adaptations.

Keep the user's correction path visible

A fictional product might suggest a summary of an invented maintenance request. The reviewer can compare it with the source, correct it and approve a saved draft. An unavailable model leaves the original request usable. This is an illustrative plan, not a delivered customer project.

Separate suggestions from actions

Define what may be displayed, saved, sent or changed and the approval required at each step. Application permissions should limit the action independently of the text the AI produces.

Include the operating plan

Estimate usage ranges and cost, identify support ownership and define behavior when a provider fails. Keep the tested configuration, evaluation results and release recovery together.

A pilot should end with a review decision: expand, revise, use a simpler method or stop. Review the development process or discuss the first task.