Permission Debt Comes Due When the Assistant Arrives: A UK Microsoft 365 Pre-AI Review
An AI assistant does not need new access to cause a new problem. In a Microsoft 365 tenant, it works with the permissions that are already there. That is the point of the design, and it is also the risk. When an assistant can search across SharePoint, Teams, OneDrive and email, it retrieves what the signed-in user is allowed to see. If the organisation has spent years granting access more generously than its documents deserved, the assistant will make that generosity visible — quickly, politely and at scale.
UK organisations preparing an AI rollout in 2026 should treat the assistant as an audit trigger, not a magic layer. The work before go-live is not primarily about prompts or licences. It is about permission debt: the sharing links, stale site memberships, guest accounts and open libraries that were tolerated because no one had a reason to look. The assistant gives everyone a reason.
Why the assistant changes the timing, not the duty
Microsoft 365 assistants are designed to respect the permissions model. They do not usually break inheritance or grant themselves access to a file. The difficulty is that many tenants contain permissions that were never deliberately designed. A site may have been shared with 'Everyone except external users' during a project. A Teams channel may still include contractors from a programme that closed two years ago. A SharePoint library may carry an anonymous link created for a one-off event and never removed.
None of that is necessarily a breach. It is a governance gap. But when an assistant begins answering questions across that estate, the gap becomes operational. A user who should not see a draft restructure document may receive a summary because a link gave the whole organisation access. A manager may be shown a file from a site they forgot they belonged to. The assistant is not being careless; it is being literal.
For UK organisations, the regulatory context matters. UK GDPR requires personal data to be processed securely and with appropriate organisational measures. The accountability principle means the organisation should be able to demonstrate how it meets those obligations. An AI rollout is a sensible moment to test whether the access model still matches the data protection by design and by default principle. If the answer is no, the rollout should wait until the access model is brought under control.
Week one: inventory what the tenant actually exposes
The first week is not a remediation week. It is a measurement week. Trying to fix permissions without a baseline produces activity but not evidence.
Start with the administrative surfaces that already hold the answer: the SharePoint admin centre, the Teams admin centre, Entra ID and Microsoft Purview. Build a register of sites, libraries, Teams and shared mailboxes that matter to the business. For each one, record the owner, the sensitivity of the content, the sharing settings, the membership model and any guest access. Then add the detail that tends to be missed: site-level sharing links, organisation-wide links, anonymous links and sites with no active owner.
The aim is not to produce a perfect inventory of every file. It is to identify the sites and libraries where broad access is both present and plausible to cause harm. A finance site with an 'Everyone except external users' link is a different proposition from a social committee site with the same setting. The register should capture that difference, not flatten it.
This is also the week to define the scope of the AI pilot. If the rollout is initially limited to a small group, say so in writing. The review can then be prioritised around the data those users can reach, rather than boiling the ocean across the whole tenant.
Week two: rank the sites by consequence, not by count
A common mistake is to order remediation by the number of files or the number of users. Volume is a poor proxy for risk. A single HR folder with salary data and an open link is more urgent than a large project site with controlled membership.
Use a simple scoring model that reflects the business. Consider the sensitivity of the content, the breadth of access, the presence of external users, whether the site is dormant, and whether the owner is still contactable. Sites that combine sensitive content with broad or anonymous access should rise to the top. So should sites that have no clear owner, because no one is currently accountable for reviewing them.
At this stage, it helps to involve the business rather than treating the work as an IT-only exercise. Site owners can confirm whether a link was intentional, whether a Team is still in use and whether a guest still needs access. Their answers should be recorded, not just acted upon. That record becomes part of the evidence trail.
Week three: remediate the highest-risk exposures
The third week is for changes, but not for indiscriminate clean-up. Focus on the highest-risk items identified in week two. Replace organisation-wide links with specific groups where the content warrants it. Remove anonymous links unless there is a documented business need. Expire stale sharing links. Review site memberships and remove accounts that no longer need access. Archive dormant sites or restrict them to a small owner group. Where guest access is still required, confirm the sponsor and the review date.
Technical controls can support this work, but they do not replace decisions. Sensitivity labels can help classify content and apply encryption or sharing restrictions. Conditional access and guest access policies can reduce the chance of recurrence. Site lifecycle policies can stop dormant sites from becoming permanent open archives. The important point is that each change should be traceable to a decision, and each exception should have an owner and an end date.
Communication matters here. Users who lose access to a site they had forgotten about will notice. A short explanation, a clear route to request access and a named owner for the site will reduce friction. Remediation that is seen as arbitrary will be worked around; remediation that is explained will be accepted.
Week four: record the work and set the gate
The final week is about evidence and control. Produce a before-and-after record for the sites in scope. Note what was found, what was changed, who approved the change and what remains as an accepted exception. This is not bureaucracy for its own sake. It is the practical form of accountability under UK GDPR: a record that shows the organisation identified a risk, took proportionate action and knows what it still needs to manage.
Use that record to set a clear gate for the AI rollout. The gate might be a named set of sites that must be remediated before the pilot starts, or a requirement that all pilot users only access sites with a current owner and a reviewed sharing model. Whatever the gate is, write it down and make it visible to the pilot sponsor.
The sweep should not end at go-live. Permission debt accumulates. A monthly review of high-risk sharing links, a quarterly membership attestation for sensitive sites and an annual site lifecycle review will keep the estate from drifting back to its previous state. The assistant will keep asking the same question of the tenant: who is allowed to see this? The organisation should be able to answer without guesswork.