Skip to main content
§ Video

What Microsoft 365 Copilot can actually see inside your tenant

Watch on YouTube ↗

Microsoft 365 Copilot can summarise a meeting you missed, find a proposal buried in SharePoint two years ago, and draft a reply grounded in an email you barely remember receiving. It can feel as though one enormous AI has quietly read the entire company.

It has not. But understanding what it has done matters, because the gap between the fear and the reality is where the actual risk lives — and the actual risk is more mundane and more likely than the one people worry about.

A scoping note first, since Microsoft's naming is its own small escape room: this describes Microsoft 365 Copilot with a licence that grounds answers automatically in your work data. The basic Copilot Chat tiers are primarily web-grounded and use work content only when you explicitly hand it over.


The journey of one request

Picture your prompt arriving at an office in a single folder. The office checks who you are, searches the company's systems, gathers evidence, and passes it to a writing room. Each of those steps is a real component: Microsoft Entra for identity, Microsoft Graph for access, Work IQ for context, Microsoft Search and the Semantic Index for retrieval, an orchestration layer, and a routed foundation model at the end.

Say you ask: "Prepare me for tomorrow's Northwind renewal meeting. Summarise the latest discussion, open risks and agreed pricing."

Reception is identity and policy. You signed in with an Entra work account. Copilot operates inside the Microsoft 365 service boundary using your organisational identity. Conditional access rules and MFA still apply. Copilot does not receive a master badge because its name appears on every second Microsoft slide. Your account, tenant, licences and permissions shape what the request can reach — and if you cannot open a SharePoint file through normal access, Copilot is not supposed to fetch it for you either.

Graph is the access layer, connecting Microsoft 365 data and the relationships between it. It is not one enormous Copilot database; it is a common interface over the services you already have.

Work IQ supplies context — the coordinator who knows "Northwind" may refer to an account, a renewal and a particular set of people. It is not the model writing the answer, and it can be switched off. With it off, Copilot still uses public web information but stops automatically referencing your work data.

Search and the Semantic Index find evidence. Exact keyword matching works if the pricing file literally says "Northwind renewal". It does not help if the document says "customer extension" and you asked about "renewal risk". The Semantic Index combines lexical matching with vector-based semantic retrieval, so it can search for related meaning as well as identical wording. It is not a readable clone of every byte in your tenant, and it grants no new access. Unsupported, archived or not-yet-indexed material may simply not appear.


Permission trimming — and the thing it cannot fix

Before results come back, every candidate passes a badge check. Retrieval is filtered to content the signed-in user may access; Copilot and the Semantic Index honour existing Microsoft 365 permissions, tenant boundaries and user identity during grounding. That is what stops Copilot being an automatic privilege-escalation machine.

It is also where the real risk hides.

Imagine a confidential restructuring plan was accidentally shared with "everyone except external users". The badge reader checks your account. Your account is allowed. The document returns.

"Copilot respects permissions" does not mean "Copilot can never surface sensitive information." It applies the permissions your organisation actually configured — including the ones configured by accident, years ago, by someone who has since left. Copilot changes the speed of discovery. It does not transform messy permissions into good judgement.

This is the oversharing problem, and it is the single most important thing to fix before a rollout rather than after. Information users could already reach becomes dramatically easier to find, combine and summarise at scale.


Grounding is not training

The authorised evidence goes back into the folder. The orchestrator does not send your entire tenant and fourteen years of email into one prompt — it selects a limited evidence packet and adds it to the request. This pattern is retrieval-augmented generation; Microsoft generally calls it grounding.

Grounding is not retraining the model on your SharePoint. Microsoft states that Graph data is not used to train Copilot's models.

But "not used for training" does not mean "nothing is stored." Microsoft documents interaction history containing the prompt, the response and the citations. That data is encrypted and governed through Purview, and its retention depends on policy, processing and legal holds — there is no honest universal claim that everything disappears after a fixed number of days. The accurate summary: your work data supplies request-time grounding, and the interaction itself remains enterprise data with real storage, residency and processor considerations.

Citations deserve the same scepticism. Copilot can return links to the sources used for grounding, and those references let you inspect the evidence — but a citation proves the source participated, not that every nearby sentence follows from it. Treat citations as an audit trail into the evidence, not a stamp reading "truth approved". For anything that matters, open the source.


Where the failure modes are

Grounding makes Copilot more useful. It does not make it infallible, and the failure modes are worth naming because each one has a different fix.

  • Retrieval miss — the prompt was vague and the right evidence never came back.
  • Oversharing hit — the retrieval worked perfectly and that is the problem; access was too broad.
  • Stale source — the file is real, indexed, permitted and eighteen months out of date.
  • Bad synthesis — accurate evidence, summarised badly, or a hallucinated bridge invented between two separate claims.
  • Connector mapping errors — a synced connector copies external content and its access-control information into the Graph index. Security then depends on correct identity mapping. A courier can still deliver a folder labelled incorrectly.
  • Hostile content — retrieved web pages can themselves contain instructions aimed at the model. Microsoft documents content filters and jailbreak defences; controls reduce risk, they do not produce a mathematical guarantee.

Agents raise the stakes again. Once an agent can call an API, the office can perform an action rather than just return information — using either delegated user identity or permissions granted to the agent itself. Least privilege stops being advisory. A fluent model choosing an API parameter is still software making a choice: give it the keys to the stationery cupboard before the treasury.


What to do about it

Two separate retrieval paths also deserve a governance decision: work grounding stays inside Microsoft 365, while web grounding sends a shorter query to Bing. Microsoft states the generated query normally excludes the full prompt and tenant identifiers, but regional commitments do not apply to Bing the same way. Admins can control this.

Before rollout, fix broad access and classify sensitive data. During use, verify claims that matter by narrowing the prompt and checking numbers in the original system. And assess the whole thing through governance and, where personal data is involved, a DPIA — the outcome depends on your tenant configuration, contracts, data locations and use case, not on a vendor's marketing page.

There is a jurisdictional layer to this too, which we cover separately in who can access your AI data: where data sits is a narrower question than which laws govern it and who can technically reach it.

When someone tells you Copilot knows everything, remember the folder. It receives selected evidence for one request, through disclosed controls and undisclosed implementation details. Powerful, yes. Psychic, no.


Independent educational analysis, not legal advice. Privacy and compliance outcomes depend on your tenant configuration, contracts, data locations, security controls and specific use case.