Nonprofit AI field guide
Reviewed July 21, 2026Start with the mission question, not the model
Turn a broad interest in AI into one bounded question that matters to people and programs.
Name the result before the technology
Write the result in terms of a person, program, decision, or operating burden. 'Reduce the time needed to prepare a first draft of the weekly volunteer update' is testable. 'Use generative AI in communications' is not.
Describe the current path, its friction, and the human owner. Include what must remain true for dignity, access, safety, trust, and mission quality. A faster output that weakens those conditions is not an improvement.
Choose work that can tolerate learning
Prefer reversible, reviewable work with low information sensitivity and a clear fallback. Drafting an internal outline is usually easier to bound than deciding eligibility, assessing risk, giving clinical guidance, or communicating with someone in crisis.
State the exclusions before the test. Name the people, records, decisions, and situations the pilot will not touch. A useful boundary is specific enough that a staff member can recognize when to stop.
Set the decision the pilot must support
A pilot should end with a choice: stop, revise, continue, or scale. Agree on the evidence needed for that choice, who will review it, and when the decision occurs.
Keep the first test small enough to observe. A narrow use case with ten representative examples can teach more than a broad rollout whose effects cannot be separated from training, policy, process, and data changes.