Do not begin with “which model do you want to use?” Understand the work first, then decide where AI belongs.
User and trigger
- Who performs this work every day?
- What event triggers the task?
- What format is the input, and where does it come from?
- Who receives the result, and what later decision does it affect?
Current workflow and cost
- How long does one complete cycle take today?
- Which step is slowest, most error-prone, or most dependent on experience?
- What are the normal and peak volumes?
- What is the real impact of a wrong result?
Data and systems
- May the data enter a model, and what must be redacted?
- Which internal systems and tools must be connected?
- Which actions may run automatically, and which require approval?
- What logging, audit, permission, and retention rules apply?
Success criteria
- Which three metrics must improve for the business team to call the project valuable?
- Who accepts the result, and who decides whether usage expands?
- How quickly can the team collect the first real feedback?
After the interview, write a one-page project judgment before writing code: target user, current workflow, critical constraints, minimum loop, success metrics, and stop condition.