Which Data Problems Hold Up Your Copilot Rollout?

When middle-market leaders explain what slows their AI deployments, they most often point to two data-related reasons. RSM’s survey of 1,030 of them ranked data quality and availability first at 34%, with security and privacy close behind at 30%.

Microsoft 365 Copilot brings both concerns to the surface the day it goes live. When someone asks a question, it draws on the files, emails, and chats that person is already allowed to open, so it creates no new access and every exposure it reveals was already there.

The difference is how easily that access is used now. A folder shared with the whole company years ago for one project went unnoticed because nobody went looking for it. Copilot does that looking now, so its contents can turn up in anyone’s answer.

That makes your Copilot rollout a data decision, made site by site. Two problems are serious enough to hold a rollout on your most sensitive sites, namely access that nobody has reviewed and sites that nobody owns.

You can fix the other two, outdated files and unlabeled content, while people already use Copilot, but you also need to change the sharing defaults that let the exposure build up.

1. Access to sensitive sites that nobody has reviewed

Check the sites holding HR, finance, legal, and board material first, because a loose permission there can put salaries or deal terms in front of people who were never meant to see them.

What the research found

A readiness assessment included with Copilot licenses flags the sites most likely to be overexposed, such as those open to nearly everyone or holding files with different permissions from the site itself.

The UK government’s Copilot trial with 20,000 staff flagged the same risk, warning that searching with each user’s access may expose cases where people can open files they shouldn’t.

How to decide

Microsoft lets you take a sensitive site out of Copilot and company-wide search while you review it, using Restricted Content Discovery.

That hides the site from company-wide answers but changes no permissions, so anyone who can open its files, including people who were never meant to, can still ask Copilot to summarize them. Restriction buys time, but a finance site’s review still has to finish before its users get licenses.

2. Sites with no owner to make the call

Reviewing that access ends in decisions only the business can make, such as whether a pricing folder should stay shared. IT can produce the reports but can’t make those calls, so ownerless sites stay stuck.

What the research found

When the UK Information Commissioner’s Office assessed its own Copilot deployment in a February 2025 impact assessment, its plan relied on site owners reviewing access to their own sites. Its data protection officer pushed back, doubting that owners did this regularly and noting the plan simply assumed they did.

How to decide

Name owners for sensitive sites before rollout and for the rest during it, and have each confirm their site’s access before its users get licenses. Keep any site still unclaimed when reviews close out of Copilot until someone takes responsibility for it.

3. Outdated files that Copilot treats as current

Some of what Copilot finds is simply out of date. It can’t tell that a 2021 price list was replaced, so an old file produces a confident wrong answer. People can catch and correct a wrong answer like that once they are using Copilot, so stale files rarely justify delaying a rollout. An old file that also holds something sensitive is the exception, and it belongs with the sites in section 1.

What the research found

Microsoft’s guidance recommends deleting or archiving obsolete files to improve Copilot’s answers. A new archiving option in Microsoft Purview, generally available from mid-November, removes archived files from what Copilot draws on once an admin switches it on.

How to decide

Take a sales team whose shared folders hold years-old proposals and price lists. The worst realistic outcome is a rep quoting an outdated rate. Give the team Copilot now, ask reps to flag any answer built on an old proposal, since each answer shows the files it drew on, and archive those folders as they come up.

4. Sensitive files without labels

Sensitivity labels, the tags that mark a file as confidential, let you set rules that keep Copilot from using a file in its answers. Those rules apply only to files with a label, so an unlabeled board presentation can still feed Copilot’s answers for anyone who can open it.

What the research found

Microsoft’s data loss prevention rules for Copilot can keep the contents of labeled files and emails out of its answers, but only after someone decides what those labels mean. The ICO’s data protection officer flagged this gap in its own plan, which appeared to rely on labels for some risks without saying which ones or how.

How to decide

Start with your most sensitive category, knowing Microsoft labels at most 500,000 files per tenant per day automatically, about two days per million files. Where no rules are planned, fixing permissions protects the same files sooner.

Defaults that keep recreating the exposure

Cleanup fixes what is there today. If people can still create sites and share them with the whole company by default, the same exposure builds up again and the next review starts from scratch. Microsoft recommends tenant-wide defaults so new sites and sharing links don’t reintroduce oversharing.

Give one person ownership of those defaults, and revisit them whenever an AI agent is connected to the same data, because an agent inherits that access and can act on it.

You can start both halves of the work this week. Run the sharing-links report on your five most sensitive sites to see who can reach them today, then check whether your tenant still lets new sites be shared with everyone by default.

Before the next batch of Copilot licenses goes out, see how Simform prepares Microsoft 365 environments for AI.