0.3Data security & privacy
The essentials in 30 seconds
- “Enterprise data” commitment: in Microsoft 365 Copilot, your prompts and answers are not used to train the models and stay processed within your tenant's perimeter.
- Copilot sees what YOU see: it opens no new door, but it shines a light on poorly closed doors — the No. 1 real risk is oversharing (documents shared “with everyone”).
- Purview labels respected and propagated: a summary of a Confidential document inherits the same level — and nothing confidential ever goes into a public chatbot.
Threaded mini-case — Hicham runs his self-check in Algiers
Hicham, administrative assistant in Algiers, is about to use Copilot on real files: first, he runs his security self-check. He identifies his three most sensitive documents (payroll, contracts, customer data) and checks who they are actually shared with in OneDrive and SharePoint — he tightens an inherited “whole organisation” share. He rereads the internal AI charter, notes that exchanges may be logged, and sets his rule: company data only in Microsoft 365 Copilot, never elsewhere. His workstation is healthy before the first internal prompt.
Standalone case for this section: it returns in the lesson, the diagram, the script and the 48h challenge of section 0.3.
1. The “enterprise data” commitment
Microsoft 365 Copilot rests on a clear contractual commitment: enterprise data protection. Your prompts and answers are not used to train the models, and data is processed within the contractual framework of your tenant — the same guarantees as the rest of Microsoft 365. That is the radical difference from a public chatbot: pasting confidential company data into one is, in most charters, professional misconduct. Microsoft 365 Copilot stays within the organisation's secure perimeter. Hicham remembers the boundary: inside the tenant, protected; outside, forbidden for confidential material.
2. Copilot sees what YOU see
Respect for permissions is absolute: Copilot only accesses content your account already has the right to. It opens no new door — but it shines a light on poorly closed doors. Two people get different answers from the same Copilot, depending on their rights. Hence the real issue revealed by rollouts: oversharing. Documents shared too broadly (“Everyone”) suddenly become findable via AI. The No. 1 risk is not Copilot: it is the quality of your internal sharing. Hicham understood: before AI, tidy up access.
3. Labels and legal framework
Microsoft Purview sensitivity labels (Public, Internal, Confidential…) are respected by Copilot — and propagated to the content it generates: a summary of a Confidential document inherits the same level. On the legal side: Microsoft does not use your prompts for training, but your employer may log them for compliance purposes (like emails). Always refer to the internal charter — and never forget the golden rule: nothing confidential in a public Copilot. Hicham applies the label before sharing any generated content.
🔎 Concrete example — 🔎 Concrete example — The over-shared document
A salary file from Hicham's department was mistakenly shared with “Everyone” two years ago. Nobody had noticed.
Before Copilot: nobody stumbled on it — you had to know its exact location.
With Copilot: “What are the team salaries?” may surface it, because it is genuinely accessible to all.
The lesson: Copilot hacked nothing — it revealed a poorly closed share.Reading: the AI opened no door: it shone a light on a door left wide open for two years.
Lesson: hence access audits BEFORE any Copilot rollout — the project lies in sharing.
Section vocabulary
- Enterprise data protection
- The Microsoft 365 Copilot commitment: prompts and answers not used to train models, data processed within the tenant's contractual framework.
Simple example: Hicham's prompts on internal files stay within his organisation's perimeter.
What it is not: a promise that would also cover public chatbots.
Reflex: company data = only in Microsoft 365 Copilot.
- Respect for permissions
- Copilot only accesses content YOUR account already has the right to: it opens no new door — but it shines a light on poorly closed doors.
Simple example: Hicham and his manager get different answers from the same Copilot, depending on their rights.
What it is not: universal administrator access to all documents.
Reflex: if the AI shows too much, look for the poorly closed share, not the bug.
- Oversharing
- Documents shared too broadly (“Everyone”): the No. 1 risk revealed by Copilot rollouts.
Simple example: The Algiers salary file, open to the whole organisation for two years.
What it is not: a cyberattack against Microsoft 365.
Reflex: audit broad shares before any intensive AI use.
- Sensitivity labels
- Microsoft Purview labels (Public, Internal, Confidential…): Copilot respects them and propagates them to generated content.
Simple example: The summary of Hicham's “Confidential” file inherits the same level.
What it is not: simple folder names with no effect.
Reflex: label before generating, check before sharing.
- Free / public Copilot
- Outside the business framework: NEVER paste confidential company data into it.
Simple example: Hicham only types neutral examples there, never a payroll file.
What it is not: a space as protected as the organisation's tenant.
Reflex: faced with a public chatbot, ask “is this shareable?” before pasting.
In practice — In practice — Run your security self-check
- Identify your 3 most sensitive documents (payroll, contracts, customer data): who are they actually shared with? Do as Hicham in Algiers.
- In OneDrive or SharePoint, check “Everyone” or “Whole organisation” shares on your folders: tighten if needed.
- Reread the internal charter on AI use: what does it say about public Copilot and logging?
- Set your personal rule: no company data in a chatbot outside the tenant, ever.
Ready-to-say script — reassure on the framework, warn on sharing (30 s)
“Good news: in Microsoft 365 Copilot, our prompts do not train the models and stay in our tenant — and the AI only sees what each person has the right to see. But it reveals poorly closed shares: a document open to everyone becomes findable. Like Hicham, I checked my three sensitive folders and tightened one overly broad share. Before rolling out, we audit access with the IT department.”
Say while showing the three-guarantees diagram; end on the action: the sharing audit.
Copy-paste template — test your sharing (if 365)
Prompt 1 (over-sharing test — work account, in Microsoft 365 Copilot):
Which recent documents mention the bonus file? List them with their locations.
→ Any over-accessible result = share to tighten.
Prompt 2 (personal rule to copy out):
My rule: company data only in Microsoft 365 Copilot (work account). Never put confidential material in a public chatbot.Without a 365 licence
Hicham on the free version: no internal test possible — he applies the manual self-check (checking OneDrive/SharePoint shares) and the “never confidential outside” rule.
Before rollout
Algiers team as 365 candidates: run prompt 1 on two sensitive pilot files, list the broad shares found, and have the tightening validated by the IT department before generalising.
Remote work
At home on a personal device: doubled vigilance — work account only in the organisation's applications, never copy internal folders to personal tools.
New joiner
First month in the role: ask for the AI charter, find out who administers sharing, and copy out the personal rule before the first prompt on a real file.
✓ Best practices
- Audit “Everyone” shares before any rollout: that is where the real leaks hide.
- Apply Purview sensitivity labels: Copilot respects them and propagates them automatically.
- Check the confidentiality level of generated content before sharing it.
✗ Common pitfalls
- Pasting company data into a public Copilot: professional misconduct in most charters.
- Believing Copilot “hacks” documents: it only shows what you already have the right to.
- Rolling out Copilot without auditing shares first: the AI then reveals poorly shared documents.
Key takeaways
- Microsoft 365 Copilot: no model training on your data, processing within the tenant perimeter.
- Copilot sees what YOU see: the real issue is the quality of your internal sharing.
- Company data in a public chatbot = professional misconduct in most charters.
- Sensitivity labels (Purview) also govern AI-generated content.
48h challenge — apply tomorrow
Run your self-check like Hicham: 3 sensitive documents, shares checked in OneDrive/SharePoint, charter reread — and copy your personal rule above your screen.
- I can explain the “enterprise data” commitment in one sentence.
- I have checked the sharing of my 3 most sensitive documents.
- My personal “never confidential outside the tenant” rule is written down and applied.
Frequently asked questions
Are my prompts read by Microsoft or my employer?
Microsoft does not use them to train models. Employer side: Copilot interactions may be logged for compliance purposes (like emails) — refer to the internal charter.
Can Copilot disclose a confidential document to a colleague?
Not by itself: a colleague will never obtain via Copilot content their rights do not give access to. The danger comes from documents shared POORLY upstream — hence access audits before rollout.
Where can confidential company data go?