An HR site open to all and an agent nobody owns: is the tenant ready for the next Copilot wave?

The problem

Copilot answers from everything a person can already open, so a rollout that went well in the pilot can go wrong in the next wave for reasons no single console shows. Two weeks before a wave of 400 people, the HR employee relations site has been shared with Everyone except external users since a migration, and 1,650 people could get it back in an answer; the recruitment site has Anyone links and no label; a closed litigation site has no owner and is still searchable; a quarter of Finance's files carry a sensitivity label and the DLP policy for Copilot is still in simulation; an agent published to the whole company was built by someone who has left and reads from a Dropbox connector nobody approved; an HR agent helps screen candidates and is not in the AI register; 34 licences have not been used in a month while the next wave is 14 licences short; and the wave itself is 70% trained, with 14 people not registered for MFA and 31 without a compliant device. This is not the tenant governance review I published a couple of weeks ago, which looks at admin roles, Conditional Access and tenant settings: it is the review before each Copilot wave and every month after it, about what Copilot can reach and who is using it.

What I built

• Licences: licensed users with no Copilot activity in 30 days or never, licences on disabled accounts, their monthly cost, and whether the pool covers the next wave once they are reclaimed.
• Waves: training, MFA registration, a compliant device and a licence for every person of a wave before its go-live, and a champion in each department.
• Oversharing: SharePoint sites shared with Everyone except external users or with Anyone links, sites with no owner or only one, no sensitivity label, and inactive sites Copilot can still search, with Restricted Content Discovery as the stopgap while they are fixed.
• Labels and DLP: label coverage per department and whether a DLP policy is enforced on the Microsoft 365 Copilot location; agents with no owner, published to everyone, using connectors not on the allowed list or unused, and agents that support decisions about people with no entry in the AI register; and adoption by department and app with its trend and the survey.
• A verdict per wave (ready, attention or not ready), report.html, readiness-pack.md for the Copilot steering group with a decision column per wave, CSVs for licences to reclaim, oversharing, agents, adoption, waves and actions, and a read-only PowerShell exporter for Microsoft Graph and SharePoint Online.

How it is built

Python with the standard library only. Every issue says which waves it reaches: the waves with people in its department, or every wave when it is critical oversharing or an agent open to the whole company, because any Copilot user who can open the content can get it back in an answer. A wave is not ready when a critical issue reaches it. The usage report is the Microsoft Graph report per user with the last activity date overall and per app; the site columns are the Get-SPOSite properties plus the counts of the SharePoint data access governance report; the agents come from the Power Platform inventory. Licence costs are estimates at a price set in the settings, and the AI register column is a record, not a legal assessment. The demo is a made-up energy company with 1,800 people in three waves, reviewed on 5 November two weeks before wave 2, with the problems placed on purpose: 37 licences to reclaim, 1,110 euros a month, four critical sites and every wave not ready until they are fixed. 14 tests, a runbook with the steering group agenda, go or no go per wave and the ADKAR steps, and a GitHub Actions workflow that runs the tests and rebuilds the demo data and the screenshot. The findings can go as evidence into the controls register I published earlier in the series.

What I learned

From deploying Microsoft 365 Copilot and AI tools and driving their adoption, and administering Microsoft 365 agents and Copilot Studio; from being Global Administrator of Microsoft 365, with PIM, Conditional Access and app integration; from Microsoft Purview, with DLP and sensitivity labels, from SharePoint for many years and from Power Platform; from change management with Prosci, ADKAR and Kotter; from the EU AI Act and responsible AI; and from business cases with measurable benefits. The lessons behind the tool are that Copilot does not create new access, it makes the old oversharing easy to find, so the permissions have to be fixed before the wave and not after; that Restricted Content Discovery buys time but changes no permission; that a label coverage of a quarter leaves DLP with little to act on; that an agent without an owner is nobody's problem until it answers the wrong question; that an agent that helps decide about people needs its entry in the register before it is used, not after; that a licence nobody uses is the cheapest licence for the next wave; and that training and champions decide the adoption more than the licence does.

Next steps

• Run the review three weeks and one week before each wave, and every month after.
• Read the agents from the Power Platform inventory API and the label coverage from the Purview content explorer export.
• Add the agents usage report and keep a year of adoption to set it next to the time people say they save.

Want to bring it to your environment or adapt it to your case? Get in touch and we can look at it together.