AI Glossary for Nonprofit FolksNo computer science degree required.

Artificial Intelligence (AI)

This is the umbrella term, and it's stretched pretty thin these days. It covers any software that does something we used to think only people could do: writing, sorting, predicting, recognizing a face in a photo. Your email spam filter is AI. So is the autocorrect on your phone that keeps changing a donor's last name to something embarrassing. So is ChatGPT. That's why the word has started to mean everything and nothing at the same time.

Generative AI

This is the kind of AI that makes something new, like a draft letter, a summary of a 40-page report, or an image. When most people say “AI” today, this is what they mean.

Large Language Model (LLM)

The engine under the hood of ChatGPT, Claude, Gemini and the rest. It was trained on an enormous amount of written text, learned how words and ideas tend to fit together, and uses those patterns to write back to you. I think of it as a language engine, and in most products it's one piece working alongside search, databases and other tools. It's very, very good at sounding smart, which is a feature and a warning label all at once. When a colleague tells you they “ran it through AI,” they almost certainly mean one of these.

AGI (Artificial General Intelligence)

A hypothetical future AI that could outperform people at most kinds of work. The experts can't agree on what it means or when (or whether) it shows up. Great fodder for a board retreat over dinner. Just don't let a vendor sell it to you. Nothing you can buy this year is AGI, whatever the keynote speaker says.

Prompt*

The instruction or question you type in. The quality of what you get back depends heavily on what you put in. Ask it to “write a thank-you note” and you'll get something that sounds like a greeting card. Ask for “a warm, two-paragraph thank-you to a first-time $500 donor to our literacy program, mentioning that her gift covers books for a full classroom” and now you've got something you can work with. If you don't know the right question to ask, don't expect a great answer.

Training

How an AI tool learned what it knows in the first place, by chewing through huge amounts of data. It doesn't keep a filing cabinet of everything it read that it can go back and look through. What it learned gets baked in as patterns, which is closer to habit than memory. And like anyone's knowledge, it has a cutoff date and can be out of date.

Hallucination

When an AI tool makes something up and says it with a straight face. A statistic that doesn't exist, a citation to an article nobody wrote, a donor's spouse with the wrong name. If you remember one term from this whole document, make it this one. Anything AI produces is a starting point, and somebody with a pulse needs to check the names, numbers and facts before it goes out the door. I promise the donor will notice if you get her late husband's name wrong. The tool won't flag it for you.

Grounding*

Pointing the AI tool at sources you trust, like your CRM, your annual report or your website, so it works from those instead of from whatever it absorbed in training. The most common way to do this is called “RAG,” which stands for retrieval-augmented generation, and which nobody outside of a tech company should have to say out loud. Here's the difference in practice: a tool that drafts a stewardship report from your actual gift records is grounded. A tool that writes a lovely report from general knowledge about scholarships is guessing. Grounding cuts down on hallucinations, but it doesn't get rid of them, so still check pledge balances, gift restrictions and citations against your own records.

Context Window*

How much information the tool can work with in one conversation, counting your instructions, any files, the back-and-forth so far and the draft it's writing. Feed it more than it can hold, or bury an important instruction in the middle of hundreds of pages, and it starts losing the thread. That's why it might handle a 10-page grant report beautifully and then miss details halfway through your 300-page employee handbook.

Token

The little chunks of text these tools read and write. In plain English, a token is roughly three-quarters of a word. Numbers, tables, code and other languages break up differently and usually use more tokens, so a big spreadsheet export from your CRM can cost more than you'd expect. You'll mostly run into tokens when comparing vendor pricing or when a tool tells you you've hit your limit for the day.

Inference

The technical word for the AI actually doing the work, meaning answering your question or writing your draft. It mostly shows up in pricing sheets and contracts, so if you see it in a proposal, that's what they're charging you for.

Reasoning / Chain of Thought

When a tool works through a problem step by step before it answers, instead of blurting out the first thing that comes to mind. It's slower and often costs more, and it's usually (though not always) more accurate on anything complicated, like reconciling a budget variance or comparing the requirements of three different grant applications. For a simple question, it's overkill. And a neat list of steps on the screen doesn't guarantee the answer at the bottom is right.

Deep Research*

A mode in tools like ChatGPT, Gemini, Claude and Perplexity where you hand over a question and the tool spends several minutes searching the web, reading dozens of sources and writing up a report with citations. Think of it as a first pass at prospect or foundation research. Perplexity's version is worth a look for our crowd because it's quick and it makes checking the sources easy, which matters when a gift officer is about to walk into a meeting with what it found. There are also tools built to go further, like Alchemy Agentic’s Pig Knuckle, which runs a research job across several AI models and compiles the citation list from the full documents it actually read. Whichever tool you use, read the report and click through the citations that matter before you act on anything.

Fine-tuning

Extra training that makes a general AI tool better at a specific job. When a vendor tells you their product is “built for fundraising,” they may have fine-tuned a general model on fundraising material. Just as often, they've taken a general model and wrapped it in fundraising instructions and a library of fundraising content. Ask them which one, and ask what material they used. It's a fair question, and the answer will tell you a lot.

AI Agent

A tool that carries out a series of steps on its own instead of answering one question at a time. Let's say you ask it to pull every donor in your portfolio who hasn't been contacted in 18 months, research each one, and draft a short summary for each. An agent can run that whole list without you stepping through each one. That's powerful for research and back-office work. It's also exactly where you want a human reviewing the output, because nobody is looking over its shoulder while it runs. Before you turn one loose, decide what data it can reach, which actions need a person's approval, and how you'll see what it did.

API

The connection point that lets two pieces of software talk to each other. When a vendor says their tool “integrates with” Salesforce or Blackbaud, there's usually an API doing the handshake behind the scenes.

Connector / MCP (Model Context Protocol)

A standard way of plugging an AI tool into your other systems, like email, calendars, shared drives or databases, without writing custom code for every single pairing. Anthropic introduced MCP in 2024, and it's now an open standard overseen by the Linux Foundation. A connector only reaches the files and functions someone sets it up to reach, and getting the permissions and security right still takes your IT person or a consultant. Don't let staff install connectors to shared drives or the CRM on their own. This is where your data access conversation should start.

Open Source

Software whose inner workings are public, so anyone can look under the hood, use it or modify it within the terms of its license. There's a wrinkle with AI: “open” often means you can download and run the model, but the license may limit what you do with it, and what it was trained on usually stays private. It can save money and give you more control, but somebody has to maintain it. If you don't have that person on staff, you're paying a consultant instead.

AI Policy vs. AI Philosophy*

An AI policy is the rulebook: what staff can and can't do, what data stays out of which tools, who signs off. Every organization needs one. An AI philosophy comes first and goes deeper. It's what your organization believes about how this technology should serve your mission, your donors and your people. I think of it as a stool. The legs are functionality (does it work?), business (does it pay?), ethics (should we?) and policy, law and regulation (are we allowed?). Sitting on top is the person the whole thing is supposed to serve, whether that's a donor, a student or a member of your staff. Without a philosophy, organizations tend to buy software first and figure out what they believe later. My advice: don't let your policy become your ceiling. Let it be your floor, and build from there.

Data Privacy / Model Training Setting*

This is the one leadership and boards should lose sleep over. Plenty of AI tools will use what you type to keep training their models unless someone turns that setting off. Business and enterprise plans usually leave training off by default. Free and personal accounts usually don't, and paying $20 a month for a personal plan buys you features, not privacy. Those personal accounts are exactly what a well-meaning staff member signs up for on their own. Think about what flows through a development office: giving histories, estate plans, family situations, things a donor shared in confidence.

The first thing I do with any new tool is find the training switch and turn it off, but that switch is only step one. Even with training off, companies keep what you type for a period of time for safety review, anywhere from weeks to years depending on the tool and the plan. The setting also only protects the account it's set on. For donor and board information, use an account your organization manages, with a signed data agreement that spells out training, retention and who can see what. And keep the most sensitive details out of the tools altogether. I treat my own information like a fingerprint, and your donors deserve the same from you.

Human in the Loop*

If you take one idea from this entire glossary, take this one. A person reviews, edits or approves the AI's work before it goes anywhere that matters: donor letters, grant applications, anything public, and any decision about people, like hiring, program eligibility or who gets assigned to which gift officer. How the human is involved is the most important part of the whole setup, more important than which tool you pick.

Let's say a tool drafts a thank-you note from the donor file and sends it over a gift officer's signature without her ever seeing it. Months later the donor thanks her for the lovely note, and she has no idea what he's talking about. Nobody broke a law, but the relationship took a hit. Review also only works if the reviewer actually reads. Clicking approve on a stack of drafts that look fine is how this safeguard usually fails. Put a name on every sign-off, so when something goes sideways you can trace where it happened. And never let an automated score be the only basis for a decision about a person.

Predictive Analytics / Wealth Screening*

The older cousin of generative AI. Instead of writing anything, it scores or ranks your records based on patterns, like likelihood to give or capacity to give. Keep in mind that capacity tells you what someone could give, not whether they want to. Plenty of development offices have been using this for years without ever calling it AI. You've probably heard of some of the companies in this space, like DonorSearch, iWave (now part of Kindsight), Dataro and Hatch.ai. Buying one of these is the most common route, but it isn't the only one. An organization with good data and someone who knows what to do with it can build its own model, or hire a consultant to build one around the way that organization actually raises money. When you buy, you get someone else's idea of what a good prospect looks like. When you build, you decide that for yourself. Either way, it's worth separating predictive tools from generative ones when you talk to your board, because one has a track record in your office and the other is brand new.

Bias*

When an AI tool picks up skewed patterns from the data it learned from and hands them back to you as results. It matters most anywhere a tool is sorting people, whether that's donors, job applicants or program participants. If the history in your data was lopsided, the tool will be too, and it'll be very confident about it. Bias can also sneak in through stand-ins, like zip code, and through what you've told the tool to look for. Before you rely on a score, check whether it's wrong more often for some groups than others.