We are good at adopting language before we adopt discipline. “AI” now sits in our proposals, our strategies, and our board decks, often before anyone has asked the harder question. Does this give a frontline team more hours with the people we serve, or does it simply automate a process we should have fixed years ago? Both can carry the same label. Only one is worth the spend.
So this piece is not about whether the technology is impressive. It is about a single test: does the tool return time to human judgment, or does it just help us do the wrong thing faster.
Where it earns trust
The honest wins are unglamorous. AI is genuinely useful where the work is high-volume, low-judgment, and already stretching our people thin. Translating across the many languages a single response touches. Drafting a first pass of a report so a coordinator edits instead of starting from a blank page. Structuring messy field data so an analyst can spend the afternoon thinking rather than reformatting. In these places the technology returns the one resource we never have enough of, which is attention.
A program officer who reclaims a real share of the week is that much closer to the people the work is for. That is the measure. Not how advanced the model is, but how much human judgment it frees for the decisions only humans should make.
It also helps where we have long struggled to listen at scale. Sorting large volumes of community feedback to surface what people are actually telling us is work the sector has historically done unevenly, and sometimes not at all. Used with care, and with a person reading the hard cases by hand, this becomes a real gain in accountability rather than a shortcut around it.
Where it only automates dysfunction
The trap is using automation to industrialize a broken process instead of repairing it. If our reporting burden is heavy because many funders each want a slightly different format, a tool that produces all those reports faster has not solved the problem. It has made the underlying duplication cheaper to sustain, which is not the same as making it right. If a needs assessment is already disconnected from what a community is asking for, feeding it through a model returns a confident, well-formatted version of the same blind spot.
A useful warning sign: be suspicious whenever a new tool lets us skip a conversation we needed to have anyway. Speed applied to the wrong process does not fix the process. It entrenches it.
The risks we own
Two risks sit with us, not with any vendor.
The first is accountability laundering. When a system shapes a call about who is included and who is left out, the responsibility is still ours. “The model decided” can never be an answer to a person whose family was missed.
The second is data dignity. The communities we serve did not agree to become training data, and a household in crisis has the least power to refuse and the most to lose if we are careless with what they share. If we cannot explain a decision to the person it affects, or safeguard the data of someone who had no real choice in giving it, the efficiency is not worth it.
The takeaway
Adopt AI where it returns time to human judgment and strengthens our accountability to the communities we serve. Decline it wherever it only helps us move faster in the wrong direction. The technology is not the test. What it frees us to do, and what it tempts us to avoid, is the test.