Permission & oversharing clean-up before Copilot / AI assistant rollout

AI assistants such as Microsoft 365 Copilot or Gemini only show people what they can already open. In most SMBs that is far more than intended: old "everyone" links, open Teams sites, forgotten guests, an unprotected salary list. A clean-up before rollout is a one-off project of days to a few weeks, using tools the tenant already has.

Problem & value

• Today: permissions grew over years. Sites were public by default, links were shared "to everyone in the company", guests were never removed. It stayed hidden because finding a file required knowing where it was.
• Why AI changes this: Copilot respects the permissions of the user [1], so overshared content now surfaces in plain-language answers.
• Value: lower leak risk for HR, finance and management files, a defensible answer for DPO and works council, and fewer stale files. No sourced benefit figures; measure your own baseline.

Variants

• Microsoft 365 (SharePoint, OneDrive, Teams): built-in reports plus a stopgap that hides sites from AI search [2][4].
• Google Workspace: audit link sharing and shared drives via admin console and Drive logs before enabling Gemini [3].
• File server / Nextcloud: script-based review of "Everyone" and "Domain Users" entries and shares.

How it works

inventory of sites/drives/shares
  -> risk reports (broad groups, "everyone" links, guests, unlabelled sensitive content)
  -> rank by sensitivity and exposure
  -> site owners confirm or remove access
  -> fix: links, groups, labels, hide from AI search
  -> re-check, then pilot rollout

• Discovery: the SharePoint admin center offers data access governance reports; snapshot reports need SharePoint Advanced Management (SAM), some activity reports come with E5. Licensing changes, so verify with your reseller [2].
• Site access review: flagged sites go to their owners, who confirm or correct permissions [2].
• Stopgap: Restricted Content Discovery hides chosen sites from organisation-wide search and Copilot without changing permissions [4].
• Optional AI help: a small LLM or Purview classifiers can sample files in suspicious sites and flag likely HR or contract content for review. Triage only, never automatic permission changes.
• Human-in-the-loop: every change is confirmed by a site owner or IT.

Data preparation

Which data: the permission structure (sites, teams, shared drives, links, groups, guests) plus a rough map of sensitive content.

Where it lives: SharePoint/Teams admin centers, Entra ID, Google Admin console and Drive logs, Windows file servers, Nextcloud.

Typical gaps: ownerless sites, public M365 groups, "Everyone except external users" in permissions, links that never expire, guests from finished projects, sensitive files in old sites, no labels.

Steps: 1. Export sites and shared drives with owner, last activity and sharing setting (PREP-003, PREP-009). 2. Run the sharing reports; sort by sensitivity, then exposure. The top 20 sites usually carry most of the risk. 3. Assign owners; archive dead sites. 4. Remove or expire open links and stale guests; use named groups. 5. Label sensitive areas (container labels on sites and Teams are the quickest start) and agree rules in PREP-024. 6. Hide still-unclean sites from AI search [4].

Tools: SharePoint admin center and SAM, Purview, Entra access reviews, PnP PowerShell or Graph scripts, Google Admin console, Excel; optional Syskit, AvePoint, ShareGate, Varonis.

Effort share: high. Roughly 1-3 days of IT work for tenant-wide quick fixes in a company up to ~100 users, plus 2-6 weeks elapsed for owner reviews (own estimate). Effort scales with the number of sites. See GUIDE-003.

Implementation paths

Which path for whom

• SMB - A: tools are built into Microsoft 365 or Google Workspace and the process is standard; no custom AI needed.
• Enterprise - B: thousands of sites need Graph-scripted inventories, ticketing and recurring reviews.
• Developer - C: PnP PowerShell, Graph SDKs and a local small LLM keep file samples in-house.
• Public - C: staff and citizen data plus procurement rules favour self-operated or sovereign setups (Nextcloud, openDesk).

Path A - Buy / use what you have

• Microsoft 365: SAM (access governance, site access reviews, Restricted Content Discovery), Entra access reviews for guests (licence-dependent), Purview labels. EU tenants benefit from the EU Data Boundary.
• Google Workspace: restrict "anyone with the link" and external sharing in the admin console [3]; log-based investigation tools depend on the edition.
• Specialist tools: Syskit, AvePoint, ShareGate (Belgium, EU), Varonis. A Microsoft partner often sells a fixed-price Copilot-readiness check.

Path B - Build on cloud APIs

Microsoft Graph or Google Admin SDK exports into Excel or a database; Power Automate or n8n sends owner reminders; optional LLM (Azure OpenAI EU region, Mistral) classifies file samples. Choose B for recurring automated reviews across many sites.

Path C - Self-host (open source)

Python/PnP scripts, self-hosted n8n, Ollama with a 7-8B model for sampling (HW1 CPU server, HW2 for speed). Low ops effort, needs scripting skills.

Path D - Custom training

Not sensible; rules plus an off-the-shelf classifier suffice.

State of the art (2026)

• Microsoft publishes an oversharing blueprint for Copilot deployments built on the tools above [1].
• Consultants agree on the order: fix permissions, then labels and DLP, then switch on the assistant. Labels without encryption may not stop an assistant reading content (consultant claim, verify in your tenant) [5].
• Limits: reports show configuration, not business sensitivity; they flag intentional intranet sites and miss sensitive files in rightly open sites. Hiding a site from AI search does not stop users opening it directly. Audit logs show history, not current state.

Shared resources & combinations

Same building blocks as most everyday automation (office-suite-ai, workflow-automation). Combine with PREP-009, PREP-010, PREP-024, PREP-025, PREP-013, PREP-003; SEC-017 (DLP) and SEC-020 (identity governance) are the longer-term programmes.

Effort, cost & KPIs

• A: 1-3 days IT plus 2-6 weeks elapsed; mostly internal time. SAM features required for Copilot deployment are reportedly included with Copilot licences; scope varies, verify the feature table [2].
• B: 2-6 weeks. C: 1-3 weeks plus small ops.
• KPIs: sites with organisation-wide or anonymous access, "anyone" links, inactive guests, ownerless sites, share of sensitive sites labelled, owner review completion, time to remediate.

People & change

Involve: management (scope, risk appetite), IT and IT security (execution), DPO (HR and customer areas), department leads as data owners (they decide who needs access), end users (informed). GUIDE-003, GUIDE-005.

Works council: depends. Where one exists, Copilot-type assistants and the logs behind the review are objectively suited to monitor behaviour: §87(1) no. 6 BetrVG (BAG: objective suitability suffices). §90(1) no. 3 requires timely information about AI-supported work processes. Public bodies involve the Personalrat. Run the clean-up jointly; it protects employee data, and individual activity reports stay out of scope (GUIDE-004).

DPIA/DPO: a DPIA (Art. 35 GDPR) is likely for the assistant rollout if HR or special-category data is processed at scale; the clean-up is a key mitigation. AVV (Art. 28 GDPR) with Microsoft or Google is required.

AI literacy: Art. 4 AI Act requires measures that support staff's AI literacy; brief site owners on what assistants see (GUIDE-006).

Communication and pilot: announce "tidy-up before the new tool", give owners instructions and deadlines, pilot with IT plus one department, name AI champions (GUIDE-008, GUIDE-009).

Concerns: "you take my access" (offer a request path), extra work (top sites only, 30 minutes each), surveillance fear (explain scope).

Humans in control: owners approve changes; IT security signs off before rollout.

Risks, regulation & compliance

• EU AI Act: the clean-up is minimal risk. Assistants may fall into other classes by use (HR use is Annex III high-risk, duties from 2 Dec 2027 under Regulation (EU) 2026/1744); Art. 50 transparency applies since 2 Aug 2026 (GUIDE-014).
• GDPR: oversharing personal data breaches need-to-know (Art. 5, 32 GDPR) and a leak may be reportable. Reviews touch employee access metadata; document purpose (GUIDE-005).
• Mitigations: change in waves with a rollback list; fix permissions rather than only hiding sites; guard against prompt injection and leakage (GUIDE-013); get licence scope in writing.

Getting started

1. Name an owner; define scope and "sensitive" with HR, finance and management.
2. Run the sharing reports and list the top 20 risks (half a day).
3. Hide the worst sites from AI search, then start owner reviews with a two-week deadline.
4. Remove open links and stale guests, label sensitive containers, document rules (PREP-024).
5. Re-run reports, pilot the assistant, repeat quarterly.

Sources

1. Microsoft Learn - Microsoft 365 Copilot oversharing blueprint - https://learn.microsoft.com/en-us/copilot/microsoft-365/microsoft-365-copilot-blueprint-oversharing
2. Microsoft Learn - SharePoint Advanced Management overview - https://learn.microsoft.com/sharepoint/advanced-management
3. Google Workspace Admin Help - Google Workspace security overview (admin sharing and audit controls) - https://workspace.google.com/security/
4. Microsoft Learn - Restricted Content Discovery - https://learn.microsoft.com/en-us/sharepoint/restricted-content-discovery
5. EPC Group - Deploy Microsoft Copilot safely in regulated environments (consultant source, secondary) - https://www.epcgroup.net/blog/deploy-microsoft-copilot-safely-regulated