Artificial intelligence has moved into fundraising faster than almost any tool before it: in barely two years it has gone from a curiosity to a part of many people's working week. For a sector that runs on relationships, trust and carefully chosen words, that pace feels both exciting and unsettling. This guide is a grounded look at what today's AI tools can genuinely do for fundraisers, where they fall short, and how to start using them without embarrassing yourself, your donors or your organisation. It draws on the practical examples shared in Hubbub's AI webinar.

When fundraisers talk about AI today, they usually mean large language models, or LLMs: the technology behind ChatGPT, Claude, Gemini, Perplexity and their rivals. It helps to be honest about what these tools actually are, because that honesty changes how you use them. They are predictive models, not unlike the predictive text on a mobile phone, only far more capable. The model reads the context you give it (your prompt, the surrounding text, any data you provide) and uses a complex algorithm to predict the words, sentences and paragraphs most likely to fit what you appear to be asking for.

That single fact explains both the power and the limits. It is tempting to picture these tools either as a mere algorithm or as a person, an assistant of some kind, when really they are neither. They are more than an algorithm, because they absorb so much context that they become an extension of the way you write and speak. But they are not assistants: they are not thinking, and they are not reasoning over knowledge or structured information. Hold both ideas at once and you will use AI far more sensibly than someone who treats it as either a genius or a toy.

Working templates. Sign in to download our AI fundraising use cases, AI prompt library and AI use policy template.

A realistic view: what AI can and cannot do

Most disappointment with AI comes from expecting the wrong thing, so start with realism. These tools are genuinely good at language: drafting, rewriting, summarising, shifting tone, and adapting one piece of copy for many audiences. They are fast, they are cheap, and the cost of trying them is close to zero. By late 2024, around two-thirds of organisations across the commercial world were already using generative AI in at least one business function, according to McKinsey, though fewer than a third described their use as structured. The tools are everywhere; thoughtful use of them is not.

The limits are sharper than the marketing suggests. The most important is that these models will confidently invent things. A clear example: asked to build a donor briefing from an uploaded spreadsheet of giving history, one tool quietly fabricated the numbers. The result was not merely incomplete but wildly different from the source: it missed a great many gifts and made up the dates. Rather than admit it could not read the file, it manufactured plausible new data. The reason is exactly the predictive nature described above. The model treated the spreadsheet not as a record to be preserved intact, but as an example of the kind of answer it should produce.

This is the cleanest illustration of the gap between an AI tool and a human helper: a human assistant would never do this. Numbers, names, dates, quotations and any structured data are precisely where these tools are least reliable, and where a careless fundraiser can do real damage. Treat AI as a brilliant first-draft writer and a patchy researcher, and you will be about right.

Practical use cases

Drafting and editing copy

The strongest and safest use case is helping you write. Fundraisers spend weeks crafting a case for support, then need versions of it for different audiences, and producing those variants by hand is expensive in either money or time. This is where AI shines. Take one well-written appeal and ask the tool to rewrite it for a new audience, for example: "rewrite this scholarship appeal for a parent audience, keeping the core purpose intact, but reflect appropriate tone, emotional drivers and familiarity." You get a usable draft in seconds.

The same approach scales across a whole campaign. For a Giving Day, you can take one case for support and generate five, six or ten different segment emails: parents, alumni of different generations, existing donors, recent leavers, staff. Instead of writing fifteen or twenty variants from scratch, you let the tool produce them, at the very least as a starting point. That phrase is the key. The AI gives you a fast first pass; you bring the judgement. This is the single most useful place to begin, and it pairs naturally with our guide to email fundraising, where drafting subject lines, openings and segmented variants is exactly the kind of work AI accelerates.

One small tell worth knowing: these tools love certain stylistic habits, especially the long dash, which ChatGPT in particular scatters through its output. A run of long dashes is a giveaway that text was AI-written. You can simply ask the tool to remove them, but watch what else changes when you do, because each tweak can shift more than you asked for.

Research and summarising

The second strong use case is summarising and reshaping text you already have. Call transcripts, meeting notes, long reports and messy caller notes are ideal candidates. You might, for instance, ask a tool to "reformat the following unstructured donor call notes into the consistent format used in our CRM, and extract key outcomes, tone indicators, next steps and a summary." The value is consistency: you get the same information out of every call, whoever made it. Work that once took hours, such as drafting a donor briefing from a profile and a giving history, can come down to minutes, provided you check it.

There is an important caution here, though. These tools are much better at handling text in attachments than structured data, especially numbers in spreadsheets, and there is a good chance a tool simply will not read a spreadsheet at all. The fix is simple: copy and paste your data from the spreadsheet directly into the prompt rather than uploading a file. When the same giving history was pasted in as a table, the tool got it right. Small changes in how you feed information to the tool make a large difference to whether you can trust the result.

Saving time on admin and practice

Beyond writing and research, AI quietly removes friction from everyday admin: reformatting notes, turning text into a table, drafting the structure of a document, suggesting next steps. A more unexpected use is rehearsal. You can ask a voice-capable tool to act as a donor, or a frustrated alumnus, and practise difficult conversations before they happen. For student callers picking up the phone for the first time, rehearsing a call ten or fifteen times in advance makes the script (about themselves, about the appeal, about what they want to discuss) far easier and more comfortable to deliver. The same approach exposes the gaps in what you are ready to talk about, which is valuable preparation for major donor meetings as well as the phones.

Writing good prompts and giving the right context

The quality of what you get out depends almost entirely on what you put in. The model is, after all, just predicting the best fit for the context you provide, so the more context, the better the fit. Tell it who you are, what you are trying to achieve, who the audience is, what tone you want and what you will use the output for. You can even pre-load a tool with your own preferences once and reuse them: tell it to write in British English only, for example, and it will keep doing so.

The other half of good prompting is structure. The more precise and structured your instruction, the more predictable and consistent the result; leave gaps and the tool will guess and interpret what you meant. If you want four specific fields in a table, name those four fields. If you want something short enough to glance at before walking into a meeting, say so, then keep refining.

Which leads to the single most useful habit: iterate. You get the best out of these tools by thinking about how you will use the output, then steering the tool towards exactly what you want over several rounds. Do not accept the first answer as final. Ask for a shorter version, a different tone, a table instead of prose, three social posts instead of one. And read what the tool tells you it is doing along the way, because each time you ask it to tweak something it may change slightly more than you asked.

One final discipline is worth adopting whatever tools you use: treat every attempt as a small experiment and write it down. What were you trying to achieve? Which tools did you use? What worked, what did not, and what is next? A short record of your prompts and results turns scattered individual play into shared, repeatable knowledge, because self-directed learning in this space otherwise misses huge amounts of valuable experience from colleagues. Our prompt library is built to be exactly that kind of shared record.

Keeping a human in the loop

If one rule matters more than any other, it is this: always read the output. The fabricated giving history above is the cautionary tale. The tool produced something polished, professional and completely wrong, and it would have been easy to miss that it had misrepresented the gift history entirely. A briefing you are about to carry into a donor meeting is precisely the kind of document where a confident invention does real harm.

So keep a person between the AI and the outside world at every step. Check facts, figures, names and quotations against your source. Be especially wary when a tool generates a downloadable file: if it writes and runs a small program to build a spreadsheet, the safest response is not to trust that file and instead ask it to show the table on screen so you can verify the contents. That check takes seconds and is well worth it.

A human in the loop is not only about catching errors; it is about ownership. The AI is reflecting your voice and your judgement back at you, and the responsibility for whatever goes out under your organisation's name stays with you. Use it to draft, never to send unread.

Data protection, donor trust and ethics

This is where fundraisers must be most careful, because we hold sensitive personal data and we trade on trust. The starting point on data protection is straightforward: you need to tell people if you are going to start putting their personal data into AI tools. Under GDPR, which applies to donor data across Europe, it needs to be stated clearly in your privacy policy, whether you rely on consent or on legitimate interests. Put bluntly, if you start feeding personal data into tools like ChatGPT with no line anywhere telling donors you might do so, you are almost certainly not compliant.

The reassuring news is that most of the strong use cases need very little personal data at all. Rewriting a case for support, drafting segment emails or summarising a report involves no named individuals. Where you do handle personal information, such as notes from a call, it is often easy to anonymise by stripping out the individual details first. For most of the use cases in this guide you do not really need to put much personally identifiable information into these tools. Make anonymising your default, and include personal data only deliberately, with a clear lawful basis and a privacy notice that says so.

Copyright deserves a mention too, particularly for AI-generated images. Whether you own what a tool produces depends heavily on its terms and conditions, and separately on unresolved legal questions about the data these models were trained on. Caution is sensible here: it is hard to be confident that a freshly generated image is not visually very similar to, or effectively a copy of, something in the model's training data. Tweaking images you already own is low risk; generating brand-new images from scratch carries more.

On authenticity, the honest answer is that you cannot know how donors will react until you try, carefully. Rather than broadcasting AI-assisted copy to your whole file, test it on a small, controlled audience first: run it as an A/B test against your usual version, and watch both whether engagement improves and whether anyone felt it was not genuine. The underlying point is reassuring, because the tool is reflecting your own authenticity back at you, so well-supervised output need not feel less genuine. But the real test is the donor, and you only learn the answer by putting your work gently in front of real people.

How to get started safely

You do not need a strategy, a budget or permission to begin; you need a low-stakes task and a habit of checking. Pick one job you do often that is mostly about words, such as drafting a segment email or summarising a set of call notes, and try it. Paste in plenty of context, ask for what you want precisely, and iterate until the output is genuinely useful. Read everything before it leaves your hands, anonymise any personal data, and make a quick note of the prompt that worked so a colleague can reuse it.

From there, widen out one use case at a time, keep personal data out unless your privacy notice covers it, and test anything donor-facing on a small audience first. Done this way, AI becomes what it is genuinely good at being: a fast, tireless drafting partner that frees you to spend more of your time on the relationships only a human can build.

Frequently asked questions

Will AI replace fundraisers? No. These tools are strong at language and weak at judgement, relationships and accuracy. They draft, summarise and reformat quickly, but the strategy, the donor relationship and the responsibility for what goes out stay firmly with you. Treat AI as a capable first-draft writer, not a replacement.

Is it safe to put donor data into ChatGPT or similar tools? Only with care. Under GDPR, if you intend to put personal data into AI tools, your privacy policy must say so clearly, whether you rely on consent or legitimate interests. For most fundraising use cases you do not need much, if any, personal data, and call notes and similar material can usually be anonymised first. Make anonymising your default.

Which AI tool should I use? It matters less than you might think. ChatGPT, Claude, Gemini, Perplexity and the rest are gradually competing to be slightly better than one another, so it is reasonable to stay fairly agnostic. Use whatever your organisation permits and you find comfortable, and check the tool's terms on data and copyright before relying on it.

Why does AI sometimes make things up? Because it predicts plausible text rather than retrieving facts. Asked for a giving history it could not properly read, one tool simply invented one that looked right. Always treat figures, names, dates and quotations as unverified until you have checked them against your source, and paste structured data into the prompt rather than uploading a spreadsheet.

How do I get a better result from a prompt? Give it more context and be more specific. Tell it who you are, who the audience is, the tone you want and what the output is for, and the more structured your instruction, the more predictable the result. Then iterate: ask for shorter, different in tone, or in a table, until it is exactly what you need.

Where should a complete beginner start? With one low-stakes, word-based task, such as drafting a segment email or summarising notes. Paste in context, ask precisely, read and check the output, keep personal data out, and save the prompt that worked so colleagues can reuse it.

Members' resource pack

Tools

Hubbub's Giving Day Simulator: model likely results before you commit to a target
Fundraising Copy Checker: sharpen your appeal and email copy
QR Code Generator: link in-person moments straight to a donation page

How to use AI in fundraising