How to use AI for grant funding without it backfiring

When a charity tells me they already use ChatGPT for funders, it usually means free ChatGPT and one line: "find me some funders for my charity." Maybe a sentence about what the charity does. A list comes back, it looks fine, and nobody can tell which bits are true.

Most of the time it backfires because of the question, and that's easy to fix.

Specific in, specific out

AI runs on a system. Give it a broad question and you get a broad answer. Give it a specific one and you get a specific answer.

So if you tell it almost nothing about your charity, it does a generic search of the internet and hands you generic funders. If you tell it who you are, who you serve, what the money's for and who's funded you before, it can pull a list that actually fits.

Write down what the money's for, before you ask

This is the first thing I'd tell any charity. Before you open a chatbot, write one document. It should cover:

• What grants go towards in your charity, or could go towards.
• What's been funded before, and by whom.
• Which funders have already said yes to you.
• Plans that aren't running yet.

That last one matters more than people think. Say you're an older people's charity and you'd love to run bus trips. You'd need a driver. You'd need a bus. None of it exists yet. Write it down anyway, and the chatbot can find funders for it.

Then paste the document in with your question. You'll see the difference in the first answer.

The same habit helps the application later. Whatever you ask for, show what it keeps running or makes possible. A funder reading "we'd like a grant" with nothing behind it has nothing to say yes to.

Only trust a funder fact you can see on the funder's own page

Say a chatbot tells you a fund closes on 31 March and gives up to 50,000 GBP. It'll say it with total confidence. It might be right. It might be describing a programme that's since closed.

Our rule is simple. Every fact about a funder carries a link to where it came from, and the date it was read. Use a directory to find a funder, then read the funder's own page for the facts.

You can try the same thing with a free chatbot. After it gives you a list, ask it:

"For each funder, give me the link to the page on the funder's own website where you found the deadline, the income rule and the area they fund."

Then open the links. If there's no link, or the page doesn't say what the chatbot said, treat it as a guess. If the page doesn't mention a deadline at all, write "not stated".

Check it, then check the check

AI has a habit of being wrong with confidence. It'll bridge a gap with something that sounds right, and it says it so well that you believe it.

So I decided nothing reaches a charity as confirmed on one read. In our system, a second reader goes back to the funder's own page and has to quote it for each key fact. If the page doesn't back it up, it gets flagged, and a person settles it.

A two-person team can't build that. You can still do the everyday version:

• Run the answer back through the chatbot: "Check every fact you just gave me against the funder's own website, and tell me which ones you can't confirm."
• Then open the funder's page yourself for anything you're going to rely on.

Yes, it takes time. It takes a lot less time than writing an application to a fund that's already shut.

What never goes into a free chatbot

Keep it to what's already public. Most of what's on your website is fair game, because the public can already see it. The exception is anything about a named person you support, even if it's published.

Keep everything else out. Nothing confidential, and nothing about the people you help.

It's worth knowing why. By default, ChatGPT can use conversations on its consumer plans to train future models. You can switch that off for new chats: Settings, Data controls, then turn off "Improve the model for everyone". OpenAI says its Business and Enterprise plans aren't used for training by default. Funders give the same advice from their side: Kent Community Foundation's own guidance tells applicants not to put confidential or personal details into AI tools.

Let AI draft, then make it sound like you

AI can write a good first draft once it knows your charity properly. Your voice, your projects, your numbers, the way you talk about the people you help.

What it writes from a one-line prompt is generic. The National Lottery Community Fund, the UK's largest community funder, is clear that using AI won't get your application rejected. What it warns about is the generic part: AI text on its own won't tell the unique story of your community, and being too generic may disadvantage your application.

So if you're drafting with a chatbot, read what it gives you as if you're the funder. Fill the gaps, because there'll be some. Add the detail only your charity has. Then send it.

When we draft for a charity, it's written from their own case for support, in their voice. The charity reads it, approves it and submits it.

What funders actually say about AI

When our checks read funders' own pages in September and October 2026, more than nine in ten said nothing about AI at all. A few said you can use it, and one asked you to say if you had. None banned it.

The Fundraising Regulator's guidance, published 08/12/2025, asks charities to agree an AI policy before they start using AI tools, and to have a person check anything AI produces before it's used. It doesn't mention grant applications specifically, but it covers fundraising broadly, and the same rule works for grants.

Three rules to start this week

Give it everything public. Where you're based, when you started, who founded you, your mission, what you're planning next. The more it knows, the better the list.

Check, check and check. Ask it to check itself. Then check it against the funder's own page yourself.

Let it draft, never let it send. Read it as the funder would and add what only you know before anything goes out.

What I believe now

In my experience, charities are under-using AI for the boring, repetitive jobs. Mostly it's used at the chatbot level: ask once, take the answer.

What's missing is a system that does the repetitive research, checks it against the source and puts funders you can actually apply to in front of you. I'm building that.

If you'd rather spend your hours on the work your charity exists to do, book a call. Before we speak, we'll build you a shortlist of 5 to 10 funders you could apply to this week, and it's yours to keep either way.