Module 00 · Lesson 00.6
Adopt the safety and validation check
Use a short preflight before prompting and a risk-scaled review before copying, sharing, deciding, publishing, or automating.
Protection is specific to the surface and source
Microsoft applies enterprise data protection to signed-in organizational Microsoft 365 Copilot and Copilot Chat experiences. Prompts, responses, and Microsoft Graph data are not used to train foundation models. That does not turn every Copilot-branded surface, web query, installed agent, or pre-existing permission into the same risk.
Permissions
Copilot honors access; it does not repair oversharing
If a file is already available too broadly, Copilot can make that existing access easier to use. Permission hygiene, sensitivity, DLP, and source review still matter.
Web queries
Public-web search has a different handling path
Microsoft removes user and tenant identifiers from generated Bing queries, but Bing operates separately from Microsoft 365 customer-data processing. Do not send confidential details into a web-search query.
Agents
Review the publisher, data, tools, and terms
An agent can have its own knowledge, actions, connectors, privacy statement, terms, permissions, approval state, and consumption model. Visibility in Copilot is not a universal safety certificate.
Scale the review to the consequence
A brainstorming list and a customer-facing legal notice should not pass through the same review.
| Risk tier | Examples | Minimum review |
|---|---|---|
| Low | Fictional practice, private brainstorming, a disposable outline | Check relevance, obvious errors, and source boundary before reuse. |
| Moderate | Internal recap, project plan, analysis draft, routine email | Trace facts, verify names/numbers/dates, check permissions and missing context, and obtain the normal owner review. |
| High | Legal, financial, medical, HR, security, regulated, public, contractual, or automated action | Use only approved systems and data; require qualified human review, documented evidence, checkpoints, policy controls, and explicit approval. Copilot must not be the final authority. |
The preflight and validation check
Use the left column before prompting. Use the right column before the output leaves your private working context.
Before the prompt
- Surface: I know the active account, product label, and organization policy.
- Data: I am allowed to use this material in this surface.
- Permission: The source is not overshared, ownerless, or outside my legitimate need.
- Source: I named what Copilot may use and what it must ignore.
- Goal: The job, audience, output, exclusions, and definition of done are clear.
- Consequence: I chose a review and approval path that matches the risk.
Before action or sharing
- Source: Important statements trace to the original evidence.
- Verified: Names, numbers, dates, citations, calculations, and quotations were checked.
- Context: Caveats, dependencies, contradictions, exceptions, and affected people are visible.
- Resilient: The result exposes assumptions and still makes sense if inputs change.
- Permission: The output’s destination and audience are allowed to receive its contents.
- Approval: The correct human owner reviewed the final artifact or action.
Add checkpoints to consequential work
For longer or higher-risk tasks, Microsoft’s current guidance recommends explicit boundaries for goal, freshness, context, audience, and checkpoints.
Goal:
Approved sources:
Required freshness date:
Context and known constraints:
Audience and allowed destination:
Facts that require independent verification:
Actions Copilot must not take:
Checkpoint 1 — outline and sources:
Checkpoint 2 — draft and unsupported claims:
Checkpoint 3 — qualified human approval:
Stop condition:A checkpoint is meaningful only when the work can stop. Do not ask Copilot to complete a sensitive end-to-end process and then treat a final glance as human control.
Exercise: classify three tasks
For each scenario, decide whether to proceed, modify the setup, escalate, or stop. Write the reason and the required reviewer.
Scenario A
Fictional product-name ideas
You are using invented product details in an approved work Copilot Chat surface. The list will remain private and is not a trademark clearance.
Likely treatment: low risk. Proceed, label the result as brainstorming, and keep legal/name clearance separate.
Scenario B
Internal project recap
The meeting transcript is organization-owned, everyone had notice, access is appropriate, and the recap will guide real assignments.
Likely treatment: moderate risk. Verify decisions, owners, dates, and permissions; obtain the project owner’s review.
Scenario C
Employee-performance recommendation
A manager wants an external agent to combine HR records and recommend a performance action.
Likely treatment: stop and escalate. The data, external agent, employment consequence, policy, fairness, and qualified approval requirements exceed a routine Copilot exercise.
Add the checklist to your course workspace
Tailor it to your organization. Replace generic owner labels with real roles, policy links, and escalation paths that you are allowed to record.
| Artifact field | Good evidence | Weak evidence |
|---|---|---|
| Approved surface | Named product, identity, current label, policy owner, and checked date | “Microsoft Copilot is secure” |
| Allowed data | Specific classifications and approved training sources | “Nothing sensitive” |
| Verification | Named source, facts checked, assumptions recorded, correction history | “Read it over” |
| Approval | Named role and point in the workflow where work can stop | “Human in the loop” |
| Agent review | Publisher, terms, data, tools, actions, admin state, and cost checked | “It appears in the Agent Store” |
Knowledge check
Does enterprise data protection mean a broadly shared file is safe?
No. Copilot honors existing permissions. If the file is already overshared, Copilot can make the existing access easier to use. Remediate the permission problem.
Why are web queries a separate review concern?
Bing web-query handling is separate from Microsoft 365 customer-data processing. Even when identifiers are removed, confidential details should not be sent as search terms.
What makes a human checkpoint real?
The reviewer has the evidence, authority, time, and ability to stop or change the work before the consequential action occurs.
Official Microsoft sources
- Enterprise data protection in Microsoft 365 Copilot and Copilot Chat
- Security for Microsoft 365 Copilot
- Validate Copilot output before you act on it
- Design Copilot boundaries and checkpoints
- Get started with agents in Microsoft 365 Copilot
Verified July 30, 2026. Educational guidance only; your organization’s policies and qualified reviewers control real use.