The most common Copilot story we hear isn't a security incident. It's a quiet moment in a pilot when somebody asks a harmless question and gets an answer drawn from a salary review, a redundancy plan or a folder from a project that closed three years ago. Nothing was breached. The permissions were always like that. Nobody had ever had a reason to look.
That's the honest summary of Copilot readiness. It's an access review with a deadline attached. The organisations that have a smooth rollout are the ones that did that work first, and the good news is the work is useful whether you buy Copilot or not.
Start with who can reach what
Most Microsoft 365 tenants have accumulated sharing over years: sites created for a project and never closed, "anyone with the link" links handed out to save a conversation, folders shared with the whole organisation because it was faster than picking five names. None of it caused a problem while finding things was hard. Copilot makes finding things easy.
- Find anything shared with everyone in the organisation, and check that's still deliberate.
- Review "anyone with the link" sharing, especially links created more than a year ago.
- Look at HR, finance and board content first. That's where the awkward answers come from.
- Check guest access. External people from finished projects are often still in the tenant.
A useful test before you buy anything: pick three sensitive documents and work out, from the permissions alone, who in your organisation could open them today. If the answer surprises you, that's your starting point.
Then deal with the duplicates and the leftovers
The second problem isn't confidentiality, it's confidence. If the same policy exists in four versions across three sites, Copilot will happily quote whichever one it finds. People stop trusting the answers, then stop using the tool, and the licences carry on being paid for.
You don't need a full information architecture project. You need a decision about where the current version of important things lives, old copies archived out of the way, and sites belonging to finished work closed properly rather than left drifting.
Retention is a business decision, not an IT one
Keeping everything forever feels safe until it isn't. Old content is more likely to be wrong, more likely to be sensitive and more likely to surface in an answer without its context. Agreeing how long different kinds of content should be kept is a conversation for the people who own the information, and it's worth having before Copilot makes the back catalogue searchable in a sentence.
Pick a small pilot with a real job to do
Rolling Copilot out to everyone on day one produces two outcomes: a spike in questions your support desk can't answer, and a lot of people who tried it once. A pilot of ten to twenty people, chosen because their work genuinely involves writing, summarising or trawling documents, tells you far more.
- Choose people who'll tell you honestly when it isn't helping.
- Agree two or three tasks to try, rather than leaving people to discover it in their own time.
- Give them somewhere obvious to report odd answers, and act on the ones that point at permissions.
- Set a review date. If the value isn't visible after a couple of months, that's useful information too.
What about the licence cost?
Copilot is a per-user cost on top of your existing Microsoft 365 licences, and it isn't small. That's an argument for buying it for the people whose work it changes rather than for everyone with a mailbox. It's also a reason to look at your current licensing at the same time - most organisations we review are paying for licences nobody's using, and that money can fund the pilot without touching the budget.
The honest version
Copilot is genuinely useful for some roles and largely ornamental for others. We'd rather tell you that before you buy it for four hundred people. Do the access review, tidy the obvious mess, pilot it with a group who'll be straight with you, and decide from evidence. The preparation improves your Microsoft 365 setup either way, which makes it one of the few pieces of AI readiness work that's worth doing regardless of the answer.