Episode 3 Navigating the mAIZe: Donor Segmentation

Building a Donor Segmentation Model, Step by Step.

A few episodes back I mentioned a segmentation model I built for a crisis campaign — 1,100+ donors, five tiers, different outreach for each. This episode walks through exactly how, in five steps you could start on your own donor list this week.
No black box. No “trust the algorithm.” Just the actual process, including where AI helps and where it absolutely shouldn’t be making the call.

Transcript

SPEAKER_00 0:06

Last episode, I mentioned a donor segmentation model that I built for a crisis response campaign. It was 1,100 donors sorted into five tiers with each tier getting different outreach based on affinity and capacity. I said the difference between that and one nice email blasted to everyone was the whole story of who actually gets value from AI in this sector. And I got more questions about that one example than almost anything else in this show so far. So today I'm not going to describe it in the abstract. I'm going to walk you through how to actually build one step by step in plain language so you could go and start this with your own donor list this week. This is navigating the maze. Let's build something. So before I get into the steps, I want to be clear about what segmentation actually is because I think that this word gets thrown around really loosely in our sector. Segmentation is not sorting donors by giving amounts. That's the version that most CRMs already do for you. And it's not wrong, it's just a shallow approach to segmentation. A donor who gave $5,000 once three years ago after a single emotional appeal is not the same as a donor who's given $500 every year for a decade. It's the same rough total lifetime value, but a completely different relationship. It's a completely different next move. Real segmentation will ask: what does this specific donor's history tell us about what they're likely to respond to right now for this specific ask? That's a much harder question than sorting a spreadsheet by dollar amount. And it's exactly the kind of question that AI is genuinely good at helping with. Because it can hold a lot more variables in mind at once than a person manually reviewing a donor list can. The goal isn't to be clever for its own sake. The goal is that your best, most capable donors don't get a generic form email, and your smaller, more occasional donors don't get ignored just because nobody had time to write them something personal. Slow down at each step. Here's the actual process broken down into five steps. Step one get your data in one place and be honest about its quality. Before AI touches anything, you need your donor data pulled together. Giving history, dates, campaign or appeal source, any notes that you have from stewardship calls. If they have attended any events, if you if you track that. Most CRMs can export this as a spreadsheet. And here's the part that people skip. Look at how messy it actually is. Duplicate records, inconsistent date formats, donors merged incorrectly. AI can help you clean this up faster than doing it by hand, but it can't fix data that it doesn't understand. So this step is worth your time and attention before you move on. Step two, decide what actually predicts affinity for this specific ask. This is the step that people get wrong most often. They assume that the biggest past donors you have are automatically the top tier for every future ask. Not true. For the crisis campaign that I mentioned, past response to urgent time-sensitive appeals mattered more than total lifetime giving. A steady annual donor who's never responded to an emergency appeal might actually belong in a different tier than their total giving history alone would suggest. So before you build anything, name two or three factors that actually matter for this ask. Recency, frequency, past responses to similar campaigns, engagement beyond just giving. Step three, let AI do the sorting, not the deciding. This is the actual point where AI enters the process. You feed in the clean data and the factors that you named in step two, and you ask it to group donors into tiers based on those factors. Not to reinvent your criteria, but to imply them consistently across a list that is too large to sort by hand. For that 1100 donor list, this is the step that would have taken a person days or even weeks for manual review. AI did it in less than an afternoon. And this part matters. It did it consistently, with no fatigue, no donor 200 getting less careful attention than donor number 12. Step four, you review the tiers before anything goes out. This is the step that separates responsible use from reckless use. AI sorting is a draft, not a final answer. You or someone on your team who actually knows these donors needs to spot check the tiers. Does this feel right? Is there a donor who is in tier two who you know from a relationship the data can't capture actually belongs in tier one? Move them. The model is a starting point that saves you time, not a replacement for judgment and oversight. Step five. Write different messages for different tiers and let the tiering inform tone, not just channel. This is where a lot of people undersell what segmentation can do. It's not just top donors get a phone call, everyone else gets an email. The actual content and tone should shift too. A donor with a long history of quiet, steady giving might respond to a message about reliability and impact over time. A donor who's only given once in a response to a single emotional moment might respond better to being brought back into that same emotional urgency. AI can help draft first passes of each tier's messaging once you tell it what you know about that segment's likely motivation. But the actual sending, the actual voice, that still needs a human pass before it goes to a real donor. So don't overlook and don't rely simply on AI to push this out and be correct and accurate. I want to be straight about the trade-offs here because I don't think this show is useful if I only tell you the upside. This process takes real setup time the first time that you do it. It will probably take you several hours, not five minutes. And it requires someone on your team who understands your donors well enough to do that four-step review and do it properly. If nobody on your team has that relationship knowledge, the AI sorting is just a guess with better formatting. And it will occasionally get a donor wrong. Put someone in a tier that doesn't quite fit based on the data that doesn't capture the full relationship. That's exactly why step four exists. What it gives you back, donors who feel like they were actually seen and heard at a scale that one person who's reading a spreadsheet manually could never achieve, especially for time-sensitive campaigns. That 1100 donor list would have taken one person the better part of a week to sort and personalize manually, if not even more. This took less than an afternoon of setup and a few hours of review. That's the actual process. Five steps, nothing exotic, nothing that requires a data science degree. Get your data honest, decide what actually predicts affinity, let AI sort at scale and review it before anything goes out, and let the tiers shape the tone, not just the channel. If you try this on your own list, I genuinely like to hear how it goes. What surprised you, where the model got a donor wrong. That's exactly the kind of thing I want to bring back onto this show. So please reach out. Next week, I want to shift gears a bit, away from donors and into your own team. I'm gonna be talking about using AI in hiring and walk through a real search that I ran for a client. I'm Sean. This has been Navigating the Maze. See you next week.