← All guides

Human review of AI output: a checklist for a small nonprofit

November 20, 2026 · 4 min read

A staff member pastes a paragraph from an AI tool into the newsletter. It reads well, it's fluent, and it says your food bank served a number of families that nobody on staff recognizes. Human review of AI output is the step that catches that, and at a small nonprofit nobody may have said whose job it is.

This is for an executive director, board member, or staff lead who has been told "we always check it" and wants to know what checking means.

What donors expect and what nonprofits say they do

Candid's summary of a Blackbaud Institute report, AI at nonprofits: impact and trust, compared what donors say they want with what nonprofit staff say their organizations do. Requiring human review of AI outputs was something 61% of donors wanted, and 42% of nonprofits reported doing. Assessing AI systems for bias, accuracy, or security risks showed a similar gap, at 60% of donors and 28% of nonprofits.

Those are survey figures from one report, and the article is a summary of it. They say nothing about your donors specifically. They do suggest that a plain statement of who reviews what is something donors would want to see.

The same summary notes that 85% of respondents said they use AI at work, while only 33% said their organization uses it very effectively. Using it and having a way to check it are separate things.

What human review of AI output has to catch

"Review" fails when it means "glance at it." Give the reviewer a short list of things to look for.

  1. Facts and figures. Every number, date, name, and quotation traced to a source you can open. If you can't find the source, cut the claim.
  2. Anything about a person. A story, a quote, or a description of a client should come from that person or from your records, never from the tool.
  3. Promises. Does the text commit the organization to something, such as a deadline, a service, or a result? A person with authority should agree to it.
  4. Tone. Would you say this to a donor face to face? Fluent text can still sound like nobody in particular wrote it.
  5. Private information. Did anything go into the tool that shouldn't have? The donor data article covers what never belongs there.

Say a staff member uses AI to draft a thank-you letter to a major donor. The reviewer checks the gift amount against the database, confirms the program named is the one the donor funded, and rewrites the first line in the director's voice. That takes five minutes, and it's the difference between a letter and a risk.

Decide who signs off, by kind of use

Not every use needs the same review. Sort your uses into three groups and write down who checks each.

  • Internal drafts nobody outside sees, such as meeting agendas or a first outline. The person who wrote the prompt reviews it.
  • Anything a donor, funder, client, or the public will read. A second person reads it before it goes out, and that person is named in your policy.
  • Anything about a specific person, a legal question, or money. Don't use AI for the first pass. A person starts the work.

Put the rule in one sentence in your AI policy, for example: "Anything we publish or send outside the organization is read by a second person before it goes out." A rule that short can be followed on a busy day.

Ask whether staff already do this

The checklist only matters if staff use it. Ask a few of them how they check AI output now, without making it a test. You'll hear a range, from careful to none.

If you'd like that picture for the whole organization, the AI readiness assessment has board and staff answer the same questions in private, and shows where the two groups differ on what's written down and where AI belongs. It takes about seven minutes each, and it never argues for adopting anything. It's free, and there's no paid version of it.

Start with one sentence and one named reviewer. You can add the rest after you've seen how it works for a month.

AI readiness assessment

Find out where your board and staff stand on AI

Both groups answer the same questions in private. The report shows where they agree, where they don't, and a few next steps, including holding off. About seven minutes each, and free.

Related guides