Start with one decision
Do the work that makes an AI experiment understandable.
The sequence is intentionally small. Keep the current human process available until the evidence and accountable owners support a different decision.
Frame the mission question
Name the person, program, decision, or operating burden that could improve. Describe the current workflow and what dignity, access, trust, quality, and accountability must remain true.
Map information and authority
Write allowed and prohibited information, account and service boundaries, human decision authority, mandatory review, escalation, and the non-AI fallback before comparing features.
Prepare representative cases
Use safe synthetic, public, or approved minimum information. Include normal, difficult, ambiguous, incomplete, accessibility, multilingual, and stop-path cases with a review rubric.
Run the smallest useful pilot
Compare the full current and proposed workflow, including preparation, prompting, review, correction, documentation, escalation, support, and the result people receive.
Decide from evidence
At the written gate date, choose stop, revise, continue at the same scope, or scale under a separately reviewed operating plan. Record the owner and rationale.
Keep change and exit visible
Set retest triggers for provider, model, plan, data, population, consequence, integration, policy, and owner changes. Keep a usable path to suspend, remove access, clean up, notify, and fall back.