AI for charities: how to find quick wins and avoid money pits

Starting with a product is starting in the wrong place

The usual route into AI is a demonstration. Someone sees a tool that writes fundraising appeals, or a chat assistant that answers questions about a document library, and the conversation immediately becomes about buying it. The tool may be good. The problem is that nobody has yet described the work it is supposed to improve.

Starting with a product means you inherit its assumptions about how your charity operates. You then spend the next few months trying to bend your processes and data to fit, and the value quietly drains away. Starting with the work is slower for a fortnight and considerably faster after that.

There is a second reason to resist the product-first instinct. Budgets in most charities are small enough that one badly chosen project can damage confidence in later work. Getting the first choice roughly right matters more than getting it quickly.

Gather ideas from the people doing the work

Useful AI ideas often come from the people closest to the work, alongside strategic direction from leadership and input from technical specialists. They come from the fundraiser who rewrites the same thank-you letter again and again, the finance officer who reconciles two exports by eye, and the service manager who cannot find the policy she wrote eighteen months ago.

So ask them. A single facilitated session with a mixed group will surface more genuine opportunity than a strategy document. Ask people to describe the parts of their week they find tedious, repetitive or frustrating—not the parts they think AI could fix. They understand the friction and the context better than anyone; technical specialists can then help test whether AI is actually an appropriate response.

What to look for

Across sessions, four patterns come up again and again, and they are a good place to point your attention:

• Repetitive work—the same summarising, reformatting, tagging or drafting done many times a week with small variations.
• Information bottlenecks—one person holds knowledge that everyone else has to queue for.
• Delayed decisions—work sits waiting because nobody has had time to read, sort or check it.
• Knowledge that is hard to find—guidance, policies and past answers exist somewhere but no one can reliably locate them.

Write each idea down as a sentence describing the job, the person doing it and how often it happens. “Summarise incoming enquiries so the duty officer can triage them each morning” is an assessable idea. “Use AI in supporter care” is not.

Assess every idea against five things

Once you have a useful backlog of ideas, score them consistently. We use five dimensions, and the last three are the ones that most organisations skip:

1. 01Impact—if this worked, what measurably changes? Hours returned, response times, error rates, income, or the quality of a decision.
2. 02Effort—the honest cost of building, integrating, testing and maintaining it, including the internal time nobody usually counts.
3. 03Data readiness—does the information the idea depends on exist, in one place, accurately, with the right permissions?
4. 04Risk—what happens when the output is wrong? Consider confidentiality, safeguarding, personal data, regulatory duties and reputation.
5. 05Ease of adoption—will this fit into how people already work, or does it require a new habit at a busy moment in the day?

Plotting impact against effort

Score impact and effort, plot each idea, and the list sorts itself into four groups. The value of the exercise is not the diagram—it is the conversation that happens while people argue about where each idea sits.

• Quick wins—high impact, low effort. Good candidates to consider first, provided data readiness, risk and adoption are all acceptable. They build confidence and produce evidence for the next decision.
• Useful extras—low impact, low effort. Cheap improvements worth doing when someone has capacity, but they should never become the plan.
• Strategic bets—high impact, high effort. Genuinely worthwhile, but they need sponsorship, budget, a data foundation and a realistic timeline. Do not attempt one first.
• Money pits—low impact, high effort. The exciting demonstrations often land here. Name them out loud and stop.

Treat the matrix as a first filter rather than the whole decision. An apparent quick win may need to be delayed or dropped because the data is not ready, the risk is material, or adoption would be difficult in practice.

Why some exciting ideas are worth very little

An idea can be impressive and still have almost no measurable value. A chat assistant on your website that answers questions nobody actually asks. An AI-written report that a manager then has to check line by line, because the consequences of an error are serious. A predictive model built on two years of inconsistent data.

The test is simple and slightly deflating: if this worked perfectly, what would we stop doing, or do better? If the answer is vague, the idea is not ready. Sometimes the honest conclusion is that a form, a template, a report or a clearer process would deliver the same benefit for a fraction of the cost—and that is a good outcome from an AI workshop, not a failure of one.

Pilot one thing, with a success measure agreed first

Pick a single quick win. Give it a named owner, a small group of users and a defined period. For a contained, low-risk use case, six to eight weeks can be a sensible starting window; integrated or higher-risk work may need longer. Before you begin, write down what success looks like in a form you can actually check: minutes saved per case, proportion of drafts used without heavy editing, reduction in time-to-response.

Agreeing the measure first protects you from the most common ending, where everyone agrees the pilot was interesting and nobody can say whether it helped.

Proportionate human review, and knowing what is already in use

Match the level of review to the consequence. Where AI assists or drafts—summarising, suggesting, reformatting—a person checking the output before it is used is usually enough. Where something consequential is being decided about a supporter, a beneficiary, a funder or a set of accounts, a named person remains responsible for that decision, and that is the operating model rather than a temporary safeguard.

Alongside it, keep a simple inventory of AI use across the organisation: the tool, the purpose, who uses it, what data goes into it and who signed it off. Most charities are surprised by how much is already happening informally. An inventory is not a crackdown—it is how you find your best internal expertise and your genuine risks in the same afternoon.

Treat the backlog as a living document

The ideas you parked are not rejected. Tooling improves, costs fall, your data gets tidier and your team gets more capable. An idea that was a money pit last year can become a quick win once the underlying information is in one place.

Revisit the list every six months. Rescore it. Move things. That habit—rather than any single tool—is what makes a charity good at this over time.