Define the job
Describe the task, the user, and the expected output. Identify the current process so there is a comparison. Be specific about the portion AI should assist. If success cannot be described, clarify the problem before choosing an implementation.
Map the information flow
Identify input sources, the services they pass through, and who sees the output. Identify confidential, personal, or licensed information before selecting providers. Review applicable terms, data-handling settings, and contractual arrangements. Do not assume an external model is suitable for every dataset.
Set authority and oversight
Decide whether the system suggests content or can take actions. Define approvals, permissions, and escalation. Make human review expectations clear. Design for missing context, incorrect output, and requests outside the task. A plausible response may still require verification.
Evaluate and review
Use representative examples and record the results, including failure cases and review burden. Decide who approves release and what evidence is required. Retain a way to stop or roll back. Changes to models, prompts, data, or tools should trigger review of relevant evaluations.