Which nonprofit task should you try AI on first?

The short answer

I would start with a recurring task that has clear source material, a recognizable good result, and someone available to review it. Drafting a short event announcement from approved details is one possibility. Allowing software to decide who receives a service is a very different undertaking. Choose a first experiment you can inspect, correct, and stop.

A small organization does not need an impressive demonstration. It needs a useful change that still makes sense on a busy Tuesday.

Name the work before choosing the tool

“Use AI for communications” is too broad to evaluate. “Turn approved event details into a first draft of our weekly announcement” gives us something concrete. We can see the input, name the person receiving the draft, and describe what needs to be correct.

I would ask the person doing that work to walk through a recent example. Where did they have to search? What did they copy twice? Which decision required their experience? A repetitive task can contain several different problems, and only one may call for AI.

If the work is copying the same fields between two systems, an ordinary integration may be enough. If people cannot find an approved document, a better shared folder may help. Anthropic's 2024 engineering guide also recommends beginning with a simple approach and increasing complexity when the task requires it. That is practical design advice, not evidence of a particular nonprofit's results. Read Building effective agents.

Compare a few candidates

I would put three recurring tasks beside one another and ask the same questions of each:

• Can we recognize a good result? “Correct date, location, registration link, and audience” is easier to check than “make our communications better.”
• Can a reviewer trace it to a source? Approved notes and current documents give someone a basis for checking the output.
• What happens if it is wrong? A draft held for review has different consequences from a message sent directly to a family.
• How much review will it need? A quick draft is of limited use if checking it takes longer than writing it.
• Can we try it with appropriate information? For an initial experiment, I would use public material or fictional examples before introducing private records.

The strongest first candidate is often modest: frequent enough to matter, narrow enough to understand, and easy to compare with the current process.

Three possible starting points

These are proposed examples, not client case studies.

Event communications. Give the tool approved event facts and ask for a draft announcement. A staff member checks every factual detail, adjusts the tone, and publishes it. Keep a reusable template alongside the AI version so you can see which approach actually helps.

Finding answers in a volunteer handbook. Try a small set of common questions against the current handbook. Require each suggested answer to identify its source passage. If the handbook does not answer a question, the useful response is to refer it to a person. Keep the trial internal while you learn where the answers fail.

Routine meeting follow-up. Ask for proposed action items from approved meeting notes. The meeting owner checks the person responsible, the deadline, and whether the group actually made that commitment. A plausible assignment is still wrong if nobody agreed to it.

Give the trial an owner and an ending

Before starting, I would write a short agreement: the exact task, the material the tool may use, the actions it may take, the reviewer, and the result that would justify continuing. The first version might prepare drafts while leaving sending, publishing, and changing shared records with a person.

Try ordinary examples and awkward ones. Include a missing date, conflicting instructions, or a question the source cannot answer. Notice whether uncertainty stays visible or disappears into a confident sentence.

Then record the total effort: preparation, review, corrections, and any work needed to keep the tool running. Ask the person using it whether the process is clearer. A short trial can reveal problems and guide the next decision; it does not establish reliability across every situation.

I would continue when the work becomes easier to complete and responsibility remains clear. If the trial creates more checking than relief, simplify it or choose another task. Learning where a tool does not belong is useful work, too.