Two surveys came out last week, four days apart. They belong next to each other.
BDO asked 250 nonprofit leaders about technology for its tenth annual benchmarking report. Ninety-three percent said their organization uses AI in some way. Seventy percent have formally adopted it somewhere.
Four days earlier, a survey of 1,020 donors asked how they feel about that. Ninety-three percent said they want to be told when an organization uses AI. Seventy-seven percent said it matters a lot.
Nobody told them.
Donors Said Where They're Fine With It. It's the Part Nobody Built.
Here is the number everyone is quoting from the donor survey. Forty percent are uncomfortable with AI analyzing donor databases to find likely givers.
Here is the bigger one nobody is quoting. Seventy percent are worried about AI bots being passed off as human.
Thirty points apart. The sector picked the smaller one.
What people are actually upset about is being fooled. You call an organization with a question. Something picks up that sounds like a person, talks like a person, and never mentions that it isn't one. You find out later. That is the complaint, and it is a fair one.
Being analyzed is a different thing, and people mind it less. Forty percent uncomfortable means most people aren't. The uses donors said they are most comfortable with are the number-crunching ones — measuring impact, handling finances, catching fraud. Nobody is pretending to be a person in any of those.
Which is the part worth sitting with. Donors are fine with AI measuring impact. Measuring impact is exactly what 49 percent of the organizations BDO surveyed say they have no real system for. Fifty-nine percent say the technology isn't good enough, in the same year that 93 percent say they have it.
Donors gave permission for the hard thing. The sector spent its effort on the easy thing.
One note before going further. The survey asked about AI, and most segmenting isn't AI at all. It is a spreadsheet and somebody who knows the file. Those two got lumped together this year and they shouldn't have been.
Buying It Took a Week. Using It Was Never Scheduled.
In most organizations, bringing in a new tool looks like this. A license, a login, a training session, a line in the next board report. That is a purchase. It has a date, an owner, and a price, and when it's finished there is something to show.
Getting something out of it is a different job. Somebody has to decide what the organization actually wants to know. Somebody has to build what feeds it, change how the work moves around it, and still own the result six months later. That job has no date and no owner. It needs a few uninterrupted weeks from the person with the least time in the building. And nothing bad happens on any particular Tuesday if it doesn't get done.
So it doesn't. You can see where it ends up in the survey. Fifty-eight percent of these organizations put AI oversight in IT instead of giving it a real owner or a committee. BDO points out the risk that creates on compliance and finance. There is a quieter one too. Nobody whose job touches the actual decisions is on the hook for whether the thing ever does anything.
The tool showed up. The work stayed the same. So did the results.
That isn't anybody being careless. It is what happens when you adopt something in a busy year and never get a quarter to rearrange around it.
"Human in the Loop" Has Stopped Meaning Anything
Let's be clear about what this isn't. Using AI to help write things is fine. It is useful. It is not what the donors in that survey are upset about. Nobody in that group is bothered that a first draft had help.
There are two different jobs here and they keep getting called the same thing.
One is making it — turning a decision into a paragraph, a subject line, a layout. The other is deciding. What is this email for, what does it say, who gets it, and should it go out at all. The first can have help. The second is the job, and it belongs to a person who can answer for it afterward.
A phone system that lets someone think they're talking to a person, and an appeal that nobody really decided the content of, are the same problem in two places. Both times the donor assumed a human was involved somewhere. Both times there wasn't one. The first is obvious because it happens out loud. The second is far more common, because it happens quietly and nobody ever finds out.
It isn't the writing they mind. It's that nobody decided.
Which is why the phrase has to go. A human in the loop has come to mean somebody clicking approve on a list they didn't build and couldn't judge if they tried. A human at the helm is steering. One is a signature. The other is a decision.
The system wrote it. The system picked who got it. The first person to read it was the donor.
The Work Is the Deciding, Not the Sending
Deciding what people should hear is the work. It is also the work nobody has had two free weeks for.
On a donor file that means figuring out the segments, deciding what each group should hear and when, setting up a way to tell whether it worked, and writing the emails that carry it. Every one of those is a call somebody has to make and stand behind.
The platform can already do the segmenting. The information worth segmenting on is already in there — who opened what, who gave, who gave again, who went quiet and when. What never happened is the work that turns it into decisions. So one appeal goes to the whole list, the open rate comes back, and that one number tells you almost nothing you can use.
The how isn't complicated. The reason more organizations don't do it is time, not money.
It is also the part none of the new tools touch. A model can write the email, lay out the page, give you six subject lines. It can't tell you what the appeal is for, which donors need it, or whether to send it.
What makes a file worth measuring is asking people what they care about and writing the answer down. Our welcome sequences for new contacts have collected more than 12,000 survey responses — twelve thousand people saying in their own words what they want to hear about — and we build segments against more than 400 available targeting attributes. No vanity metrics. Just the numbers that help you make better decisions.
Five Weeks Is Enough to Settle Four Things
GivingTuesday is December 1 and year-end runs right behind it. Five weeks isn't enough to build a measurement system. It is enough for four things.
Write down which calls stay with a person. Not which tools are banned — which decisions are owned. What each appeal says, what it shows, who gets it, and whether it goes. Everything after those can have help. None of those four should quietly answer itself.
Tell people you're using it, in plain language, before the appeals go out instead of after somebody asks. The survey's own advice is to say it twice. Once when you first reach someone, again when they give.
Segment on what donors have told you or actually responded to, not just on what they gave. Gift size tells you what somebody can afford. What they open tells you what they care about.
And decide now what you want to be able to report in January. Measurement gets built before the campaign, not reconstructed after it.
Every one of those is something the system you already pay for can do. None of them is a technology problem.
You Already Have the Tools
Ninety-three percent is the good news, and it's being read like a finish line. It's a starting line.
The other survey is more encouraging than the headlines made it sound. Half of donors now say they are very familiar with AI, up from 32 percent three years ago, and 54 percent think it will make nonprofits better at their work. People aren't behind on this. They're waiting to be told what's going on.
If the part that's been sitting on your list since spring is the setup, that's worth a conversation.