Before you let AI search your files, check who can already see them

One of the most useful things AI can do for a small business is answer questions from your own files. "What did we agree with that supplier in 2023?" "Where's the latest price list?" It works, and it saves hours of digging.

But there's a job to do first, and it isn't an AI job. It's checking who can open what.

The plain answer

These tools don't get special access. They see what the person asking can see. Microsoft says Copilot "only surfaces organizational data" that the user already has permission to view, in its privacy notes. Google says much the same in its Workspace privacy hub: "The AI follows your existing Google Workspace permissions."

That sounds reassuring, and it is, as long as your permissions are right.

In most small businesses they've grown up over years. A folder shared with "everyone" because it was quicker. A link sent to a freelancer in 2021 that still works. The HR folder that sits inside the general shared drive because that's where someone made it.

Until now that mess was mostly hidden by the fact that nobody knew where to look. Search inside a shared drive is usually poor, so a payroll spreadsheet three folders down was safe by obscurity. Ask an AI assistant "what does the office manager earn?" and obscurity stops being much protection. If the person asking can open the file, the assistant can find it and read it out.

Nothing has been broken. The AI is doing exactly what it says. It's your sharing that was always wider than you thought.

What to check before you switch it on

You don't need a big project for this. An afternoon with whoever looks after your Microsoft 365 or Google Workspace will cover most small teams.

1. List the sensitive stuff. Payroll, HR records, contracts, anything with client personal details, board papers, the owner's own folders. Usually it's a short list.
2. Check who can open each one. Look for "everyone in the organisation" style sharing and links that work for anyone who has them. In Microsoft 365 the group to look for is "Everyone except external users".
3. Kill the old links. Sharing links to people who've left, freelancers, or the old accountant. If nobody knows why a link exists, remove it.
4. Move rather than hide. If the HR folder sits inside a drive the whole company can see, move it to its own space with its own members. Fiddling with exceptions on subfolders is how this mess started.
5. Test it as a normal user. Once it's on, sign in as an ordinary member of staff and ask it for things they shouldn't see. Salaries, a named person's sick leave, the bank details. If it answers, fix the sharing, not the AI.

If you're on Microsoft 365 with Copilot, Microsoft has written a guide to tackling oversharing before roll-out. Its SharePoint Advanced Management tools include reports to find overshared sites and a setting called Restricted Content Discovery, which keeps chosen sites out of Copilot's answers. Microsoft says that comes with a Copilot licence. It's a useful stopgap, but I'd still fix the sharing underneath it.

It's the same if you run your own

This isn't only a Microsoft or Google thing. If you set up a private AI assistant on your own hardware, which I've written about before, you hit the same question from the other side.

Open WebUI, the chat screen many people put on top of local models, lets you load documents into a "knowledge" collection. Its docs say these are private by default, and the creator chooses which people or groups can use them. So the decision is yours, which is good. But the easy shortcut of loading the whole shared drive into one collection and sharing it with everybody gives you exactly the problem above, just on your own server.

The approach I'd take is several smaller collections that mirror who should see what. Sales documents for sales, HR policies for everyone, actual HR records for nobody but the person who handles them.

Does this mean don't bother?

No. Asking questions of your own documents is one of the AI jobs I think is genuinely worth doing, far more than another chatbot on the website. It just means the first step is boring housekeeping rather than buying something.

The happy side effect is that you end up with a file system you'd have wanted anyway. When someone leaves, you know what they could see. When a client asks who has access to their data, you can answer. That's worth having with or without AI.

If you're also wondering what your staff should paste into ChatGPT in the meantime, I covered that in whose account it is.

If you want AI working across your own files and systems, the private AI page has how I'd set it up. Where it needs to plug into software you already run, the team at Dev Partners can build the joins.