1. Define the task and limits
Describe the user, input, useful output and action that follows. Record permitted data and decisions that must remain with a person.
From prototype to useful application
Make each stage answer a product question.
The process should show whether the AI feature helps the intended person complete the task.
2. Establish a comparison
Use the current method or a straightforward non-AI approach as context. Prepare permitted examples and agreed checks for correct, incomplete and inappropriate results.
3. Build a bounded working version
Connect the model step to a usable interface. Include review, correction and failure states rather than only reproducing a successful prompt.
4. Test the complete result
Check application permissions, relevant saved records and external actions using controlled tests. Compare quality, time and cost under the intended pilot conditions.
5. Release and review
Agree on the first audience, usage limits, support and recovery. Review actual problems and repeat the relevant checks before expanding the feature or changing its model.
Keep the first release focused.
Use the planning guide to prepare a concrete brief.