How to prepare your organisation before turning on Microsoft 365 Copilot

Microsoft 365 Copilot does not need anyone to “connect” it to your documents: it can already read everything the user can see in SharePoint, OneDrive, Teams and email. That is its great advantage and also its main risk. If a folder containing bids, payroll or confidential reports is shared with too many people, Copilot will use it to answer all of them, even if they do not remember they have access.

That is why a Copilot rollout starts long before licences are assigned. These are the steps we recommend.

1. Clean up permissions

This is the step that prevents the most surprises:

• Review SharePoint sites, Teams teams and folders shared with “everyone in the organisation” or with very broad groups.
• Find open sharing links and set them to expire.
• Restrict the content Copilot can draw on in the most sensitive sites (for example, by excluding the site from search results or with Restricted Content Discovery).
• Apply the principle of least privilege: each person sees what they need for their work, not what is convenient.

2. Classify information

Microsoft Purview sensitivity labels make Copilot respect each document’s encryption and restrictions. A simple four-level scheme is enough to start with:

Level	Examples	Handling
Public	Information already published, regulations, press releases	No restrictions
Internal	Procedures, technical documentation without personal data	Organisation staff only
Confidential	Contracts, case files, bids in progress	Access by role and need to know
Restricted	Health data, data on minors, security configurations	Minimum access, encryption and auditing

Auto-labelling helps, but it does not replace the classification approved by the person responsible for each area.

3. Prevent leaks and log usage

• Data loss prevention (DLP): policies that warn users or block them from pasting or uploading personal or confidential information into external AI services.
• Control of unauthorised tools: block or monitor access to other AI assistants from corporate devices and the corporate network, so that Copilot is not the front door while the others are an open window.
• Audit logs: enable and retain Copilot usage logs to investigate incidents and handle complaints.

4. Set usage rules

Few, clear and mandatory:

1. AI supports; it does not replace professional judgement or a signature.
2. Authorised tools only, with a corporate account and on organisation devices.
3. Never enter credentials or passwords.
4. Do not send or sign anything generated with AI without reading it in full and checking data, figures and references against the source.
5. Basic training before use (also required by Art. 4 of the EU AI Act).
6. Report any incident or possible data leak immediately.

It is worth complementing them with a table of what information can be entered into each AI tool.

5. Inform staff and their representatives

Copilot processes employee data: emails, meetings, chats. Before enabling features such as meeting transcription and summaries across the board, you need to inform staff and their representatives, decide which meetings are transcribed, who can access the summaries and how long they are kept, and expressly rule out their use to evaluate people.

6. Train with real cases

A four-hour hands-on training session, in two blocks, works well:

Block 1. Email, documents and presentations (2 h)

• What Copilot is and is not: real capabilities and limits, and how it works with existing permissions.
• Outlook: summarise long threads, draft replies and adjust the tone.
• Word: drafts based on your own documents, summaries and consistency review.
• PowerPoint: presentations from a document, with speaker notes.
• How to ask well: context, format, tone and constraints.

Block 2. Teamwork and responsible use (2 h)

• Teams: meeting recaps, decisions and tasks; catching up on a channel.
• Microsoft 365 Copilot Chat to query the organisation’s documentation.
• What Copilot can and cannot see: permissions rule.
• Personal data, confidential information and copyright.
• Where to place human review before publishing or sending.

Each block alternates demonstrations and exercises in a sandbox with fictitious documentation, in groups of no more than 16 people.

7. Measure and adjust

During the first few weeks, monitor actual usage, incidents and the most frequent questions. If Copilot returns information that someone should not see, the problem is almost never Copilot: it is a misconfigured permission that already existed.

Real risks, with examples

• Inherited oversharing. A Teams team created years ago for a project includes half the organisation and contains a draft budget with salaries. Nobody remembered it; Copilot does find it when asked about “staff costs”.
• “Anyone with the link” links. A document shared this way is within reach of anyone who has ever opened it, and it ends up in their answers.
• Personal data in summaries. A meeting summary captures comments about a person’s health or performance and is distributed to the whole channel.
• Prompt injection. An external email or document with hidden text tries to make the assistant do something unintended, such as summarising confidential information in its answer.
• Overconfidence. A well-written draft with a made-up figure or regulatory article is sent without review.

None of these risks is new: Copilot makes them visible and faster. That is why preparation is, above all, permission and information hygiene.

A phased rollout plan

Phase	What is done	Indicative duration
0. Diagnosis	Report on overexposed sites and teams, open links and unlabelled sensitive documents	2–4 weeks
1. Hygiene	Permission clean-up, sensitivity labels, DLP and auditing enabled	4–8 weeks, in parallel with phase 2
2. Pilot	20–50 users from different areas, with training and defined use cases	6–8 weeks
3. Evaluation	Measurement of usage, perceived savings, incidents and adjustments to rules	2 weeks
4. Expansion	Rollout in waves, with training before each one	Depending on size

The pilot should include people from areas that handle sensitive information (HR, finance, legal): that is where the permission problems that must be solved before expanding show up.

Checklist before assigning licences

• [ ] Report on SharePoint sites and Teams teams with broad access reviewed and corrected.
• [ ] “Anyone” and “People in your organization” sharing links expired or restricted.
• [ ] Sensitivity labels published and applied at least to restricted information.
• [ ] Data loss prevention policies active for personal and confidential data.
• [ ] Most sensitive sites excluded from or restricted for Copilot search.
• [ ] Copilot audit logs enabled, with a defined retention period.
• [ ] Usage rules approved and communicated.
• [ ] Staff and their representatives informed about the processing of their data.
• [ ] Training delivered to pilot participants.
• [ ] Incident channel and designated owner.

How to measure whether it works

• Adoption: weekly active users and most-used features (email summaries, meetings, drafts in Word).
• Perceived value: a short survey at the end of the pilot on time saved and quality of output.
• Quality: number of incidents caused by wrong answers or information that should not have appeared.
• Security: DLP alerts and access to labelled content.

If adoption drops after the first few weeks, it is almost always due to a lack of concrete use cases, not a lack of interest. That is why training should work with real tasks from each area.

Copilot within AI governance

Copilot is not an isolated case: it is one of the organisation’s authorised tools. It should appear on the list of tools with its permitted information level (usually internal information), new features that Microsoft turns on in updates should be reviewed before enabling them, and uses involving personal data or supporting decisions about people should follow the same assessment process as any other AI system. We explain this in the lifecycle of an AI project.

At Bidatia we support Copilot rollouts from start to finish: permission review, Purview, usage rules and training. If you also want to fit it into a broader framework, see which regulations apply to AI and data.