Every organisation we speak to is somewhere on the same journey. Someone asks whether staff can use ChatGPT. A conversation happens. No decision is reached. Meanwhile several people have been using it for months to tidy up reports and draft funding bids, because it saves them an hour and nobody told them not to.
That gap — between the unmade decision and the actual behaviour — is the risk. Not the technology.
Why banning it doesn't work
A ban feels like the safe option. In practice it usually produces the worst outcome: staff use these tools anyway, on personal accounts, on personal devices, where you have no visibility at all.
This isn't defiance. It's a fundraiser under deadline pressure who has found something that helps, and who doesn't think of pasting a draft into a website as a data decision.
The more useful move is to name a small number of approved tools, be clear about what can and can't go into them, and make it easy enough that nobody needs a workaround.
What's actually at stake
Three things worth being clear about, because they're often muddled together.
Where the data goes
Free consumer AI services may use what you submit to improve their models. Paid business tiers generally don't, and enterprise agreements usually offer stronger contractual commitments. This distinction matters enormously and most staff have no idea it exists.
For a charity, the practical implication is straightforward: beneficiary information, health or safeguarding details, staff records and donor data should not go into a free consumer AI tool. Not because something dramatic will definitely happen, but because you can't demonstrate to anyone — a funder, the ICO, a beneficiary — where that information went or what happened to it.
Accuracy
These tools produce confident, fluent text that is sometimes wrong. In a funding bid, an invented statistic is a serious problem. In advice to a beneficiary, it could be worse.
The rule that works: AI can help you draft, but a person is accountable for what goes out. Nothing leaves the organisation unread.
Copilot has a different risk profile
Microsoft Copilot doesn't have the same data-leaving-the-organisation problem — it works within your tenant under your existing agreement. Its risk is the opposite direction: it surfaces content the user already has permission to see.
That sounds harmless until you remember how permissions actually accumulate in SharePoint. If a folder of HR files was shared too broadly in 2022 and nobody noticed, Copilot will cheerfully find it when someone asks a related question. It doesn't create the exposure — it makes existing exposure discoverable.
What good guardrails look like
For a small organisation, this should be one page, not a policy document nobody reads. Four things:
- Which tools are approved. Name them. "AI tools" is too vague to follow.
- What must never go in. Be specific to your organisation — beneficiary records, safeguarding notes, health information, donor data, anything under NDA.
- What's fine. Say this explicitly. Rewriting a public blog post, summarising a published report, drafting generic text, helping with code or spreadsheet formulas. People follow rules they understand the shape of.
- Who's accountable. The person whose name is on the output, always.
Where trustees fit in
Boards increasingly ask about AI, and often aren't sure what they're asking. Three questions are enough for reasonable oversight:
- Do we have guidance, and do staff know it exists?
- Has anyone checked what's actually being used?
- Is anything decided about a person — a beneficiary, an applicant, an employee — being influenced by an AI tool without a human reviewing it?
That last one matters most and gets asked least. Using AI to draft a newsletter is a very different proposition from using it to help sort applications for support.
Start here
Three steps, none of which need budget:
- Find out what's actually in use. Ask, without making it a disciplinary matter. People will tell you if the question is genuine.
- Write the one-pager. Approved tools, what never goes in, what's fine, who's accountable.
- Say it out loud. Ten minutes at an all-staff meeting achieves more than a policy uploaded to the intranet.
Then revisit it in six months, because this is moving quickly enough that anything written today will need a look before long.
If it would help
Working out where the line should be?
We help charities write AI guidance that staff will actually follow, and check whether your Microsoft 365 permissions are in a fit state before Copilot goes anywhere near them.
Tell us what's going on