Make your Microsoft data safe for AI.

The full deliverable runs to roughly 67 pages. This sample reproduces the executive and findings sections in full and abridges the design and build sections. Sections marked abridged are shortened or redacted here.

The organisation runs three Microsoft 365 tenants across its home region and two international regions. A data discovery exercise combining Microsoft administrative exports and data security posture management found that all three tenants carry material data governance risk, with several findings severe enough to block the planned rollout of AI over corporate data.

The risk is not any single AI product. It is enabling AI of any kind, Copilot, ChatGPT Enterprise, Claude or a connector that syncs SharePoint and OneDrive, over content never governed for AI-scale search. Copilot grounds on the tenant and inherits each user's permissions. Third-party AI reaches the same content through connectors. Both surface whatever a user can already reach. Across roughly 46,000 indexable locations, access has grown without governance, and one of seven gates is met.

Rating definitions

Every finding is rated two ways. Risk level is the severity against the organisation's data and operational risk. Deployment impact is the effect on enabling AI. A finding can be high risk without blocking AI, and can block AI without being the highest severity. The two are reported independently so remediation and release decisions stay separate.

Executive findings summary

The three stages of AI

AI adoption moves through three stages. Each adds capability and risk, and each depends on the stage below it.

Scope of this engagement

This assessment covers Stage 1 readiness: whether AI can be enabled safely over corporate data for productivity use across the estate. Stage 2 and Stage 3 controls depend on the specific AI tools and use cases the business intends to pursue and are scoped separately. The seven gates apply whether the AI is Microsoft 365 Copilot grounding on the tenant, or a third-party tool reaching the same content through a connector.

Access and method

Findings are based on data exported from SharePoint Online and OneDrive together with data security posture management findings. Controls that depend on Entra ID or Conditional Access could not be independently verified and are recorded on the basis of information provided by the client. Microsoft reports give a point-in-time view. Posture management adds correlations a point-in-time export cannot. At the time of writing, a first pass had scanned a fraction of the estate.

Design only

This deliverable defines the target and the sequence. Execution and remediation stay with the client, or with Cornerstone under a separate build engagement. There is no obligation to Cornerstone to act on it.

Tenant overview

Three discrete Microsoft 365 tenants with no cross-tenant federation. Combined AI indexing surface is roughly 46,000 locations.

Sharing exposure

Privileged application risk

A third-party SharePoint management application holds tenant-wide full control across all three tenants. A single compromise of that identity bypasses every site-level control in the design. The fix is not to remove the app but to harden the identity: workload Conditional Access, credential rotation on a set cadence, restricted network egress and quarterly attestation of continued need.

Systemic governance gaps

• No consistent site ownership or data owner assignment.
• No structured group model. Direct user permission assignments are common.
• No sensitivity label coverage. Most content is unlabelled.
• No lifecycle process for inactive sites or orphaned drives.
• No external sharing governance. Guest access is uncontrolled.

Findings from direct review of SharePoint admin centre configuration, independent of the data-level findings. Abridged here to the highest-severity items.

Critical and high configuration findings

Compliant settings confirmed

Not everything is a gap. The following were confirmed correct at org level and require no immediate change, subject to a site-level check during remediation:

• Default link type set to specific people at org level.
• Default link permission set to view.
• Guest access expiry configured at 30 days, with re-authentication.
• Guests cannot share items they do not own.
• Custom scripts blocked at site level.

Posture management scans data at rest across the estate and classifies it, finding correlations a configuration export cannot. Figures below are illustrative.

Finding distribution

Findings by category

Seven foundational gates must be satisfied before AI can be recommended over corporate data. The gates hold whether the AI is Copilot, ChatGPT Enterprise, Claude or a connector. Each is scored on evidence. One gate is met. Six remain open, three of them critical.

How each gate is scored

Ratings are evidence-based. A gate cannot rate above not started without an artefact from the environment. The scale runs not started, foundational, operational, optimised. The model is re-applied on a defined cadence after rollout, quarterly for the first year, then six-monthly.

Seven principles govern every design decision in the full document. They are derived from Cornerstone data management practice and adapted to the client estate.

Sections 8 to 16, design detail

The full document turns these principles into a target architecture: SharePoint information architecture, folder taxonomy and metadata, the group model, the access and permission model, data ownership, external sharing, governance, and the Purview design. The next pages show the shape of that design. The controlled vocabularies, naming conventions and configuration values are provided in the engagement, against your own estate.

SharePoint information architecture

A hub-and-site model organised by business unit and region. Hubs provide navigation, aggregated news and search scope. Content lives in associated member sites. Every site is classified into one of three tiers at creation.

Folder taxonomy and group model

A three-layer folder model (header, sub-function, working) applies to governed sites, with access assigned by group at each layer. Four group types carry access: site permission groups, role groups, team groups and data owner groups. Site and group names follow a fixed pattern by region, business unit and access level.

Access model

Three permission levels, granted only through groups. Groups nest so that access is managed in one place and follows the joiner, mover, leaver lifecycle.

Data ownership

Every site has a named data owner, accountable for content, access decisions and quarterly review. Ownership is tracked centrally through a dedicated group, not held in a person's head. Access reviews run quarterly for Tier 1, semi-annually for Tier 2.

External sharing, target state

Anyone links are removed and replaced with specific-person links. A domain allowlist limits external sharing to approved partners. Guest accounts carry expiry and periodic re-attestation. External sharing capability is restricted to an approved security group.

Governance model

A standing operating model keeps the estate safe after remediation: defined governance roles, a site lifecycle from request to archive, a disposition workflow for orphaned drives, and a governance calendar of recurring reviews. Policy is monitored continuously by posture management, with drift routed back to the owner.

Microsoft native tooling enforces. Posture management discovers and monitors at a breadth native tooling does not reach. The design uses both, each for what it does best.

Purview readiness design

The full document specifies the label taxonomy, auto-labelling logic, retention and a data loss prevention design that covers AI locations, sequenced so labels reach coverage before enforcement is switched on. Labels are what let any AI honour the client's protection model, so this design is the hinge of the whole programme.

Gate-clearing pathway

The gates clear in a set order. Identity and residency first, then oversharing and lifecycle, then classification and data loss prevention, then guardrails and operations. Enabling AI, of any kind, follows the clearance of the critical and high gates.

AI risk scenarios

The full document carries the settings reference and the step-by-step remediation the client's administrators execute. Sections 19 and 20 are the runnable core and are the most heavily abridged in this sample.

What these sections contain

• Organisation, site and document-library configuration reference, with the target value for each setting by tier.
• AI and Copilot exclusion settings for sites that must never be indexed.
• Bulk-change scripting to apply settings across thousands of sites.
• A seven-step external sharing remediation procedure, phased by risk.

A single illustrative line, to show the format:

The fix is sequenced into four phases across a year. Each phase maps to the gates it clears.

Responsibilities (RACI, summarised)

Selected risks and dependencies

Training plan

Adoption depends on people, not only controls. The full plan covers governance fundamentals for all staff, data owner responsibilities, administrator workshops, posture management operations, and safe AI use for the pilot cohort, each tied to the phase it supports.

Appendices

A: regional data summary. B: AI readiness checklist. C and D: naming references. E: the seven-gate model. F: glossary. G: the stage model. The seven-gate model in Appendix E is the assessment method itself, reproduced below.

This sample reproduces the findings and the shape of the design and roadmap. The detail that lets your team build is delivered in your own engagement, against your own data. Held back here:

Why the difference matters

A findings list tells you that you have a problem. A design tells you exactly how to fix it, in your environment, in the right order, with the controls named and the owners assigned. It applies whether the AI you enable is Microsoft 365 Copilot, ChatGPT Enterprise, Claude, Gemini or a connector that syncs your data. The sample proves the output is clear and actionable. The engagement gives you the build.

This document is an anonymised, illustrative sample. It does not describe any identifiable organisation. Figures have been altered and are not real measurements. Prepared by Cornerstone Cyber.