I ended up sitting next to a woman called Ruth at a fundraising thing last spring, the sort of evening where everyone has a lanyard and the wine is warm. She runs a homelessness charity in the Midlands. Eleven staff. Somewhere north of forty thousand case notes going back nineteen years, most of them typed into a system that was modern when Blair was in office.
She asked me, quite directly, whether "all this AI" was something she should be paying attention to or whether it was for organisations with more money than hers. And I found I couldn't answer with any of my usual material, because everything I normally say assumes a company that can afford to be wrong once.
Charities can't. Every pound they spend on a tool that doesn't work is a pound somebody donated believing it would go somewhere else. That changes the maths, and it changes what a responsible recommendation looks like.
Where it actually helps
I went away and did some homework, and then I went back to Ruth. What I found is that the useful stuff in the third sector is almost embarrassingly mundane, which is exactly why it doesn't get written about.
Grant applications. A small charity's fundraising is often one person writing similar-but-not-identical applications over and over, each with its own word limits and its own idea of what an "outcome" is. The source material — what the charity does, who it helps, what happened last year — barely changes. That's a genuinely good fit: drafting from material you already have, with a human who knows the funder doing the final pass. Ruth's fundraiser reckoned it gave her back most of a day a week, and a day a week is a meaningful fraction of an eleven-person organisation.
Reporting. Funders want impact reports. Impact reports are usually a person pulling numbers out of a case management system and writing the same twelve paragraphs in a slightly different order. Same logic applies.
And then the case notes. Nineteen years of them. Ruth's team had a real and rather sad problem: somebody would come through the door for the third time in six years, and unless the worker on duty happened to have been there for all three, that history was effectively invisible. Not lost — sitting right there in the database — just not findable by anyone in the ninety seconds you have before someone stops trusting you. Being able to ask a question of nineteen years of records in plain English is, for that work, not a productivity gain. It's the difference between treating someone as a stranger and treating them as a person you've met before.
Where I'd be careful, and I mean properly careful
That last one is also the one that worries me most, and I told her so.
Case notes about vulnerable people are about as sensitive as data gets. Names, addresses, mental health, immigration status, children, abuse. If any of that ends up in a system where it's being used to train someone's model, or sitting on a server in a jurisdiction nobody checked, the charity has done real harm to people who had very little bargaining power to begin with. That's not a compliance risk, it's a betrayal, and it lands on people who came in asking for help.
So my honest advice to Ruth was: do the grant writing and the reporting first, because the material is public-facing and the downside of getting it wrong is an awkward paragraph. Do the case notes second, slowly, with someone who understands data protection actually in the room, and probably with something that runs on infrastructure you control rather than a free tier that pays for itself with your data.
The other thing I'd flag, gently, is that the sector is being circled at the moment. Charities are being sold AI transformation by consultancies who've noticed that "digital transformation" money exists in the grant landscape. Some of that is fine. Some of it is a discovery phase and a report, funded by a grant, that produces nothing anyone uses. If a proposal doesn't end with a specific person doing a specific task differently on a specific Tuesday, it isn't a project, it's a document.
The bit that stayed with me
What struck me most about the conversation wasn't the technology. It was that Ruth's instinct was already right and she just didn't trust it, because everything written about AI is pitched at organisations fifty times her size and written in a register designed to make her feel behind.
She isn't behind. She has a clear-eyed view of what her staff spend their time on, which is more than most FTSE boards can say. That's the whole prerequisite. The tooling question is downstream of it and much less interesting.
If you're running something small and worthy and you're being told you need an AI strategy — you probably don't. You might need one afternoon, a list of the three things that eat your team's week, and someone honest enough to tell you when the answer is no.
We do that conversation for free for charities, as it happens. Ruth's still going. The case notes project is on hold until they can afford to do it properly, which I think was the right call.
