What Copilot does and which product you are buying
Microsoft 365 Copilot is a generative AI assistant that works inside Word, Excel, PowerPoint, Outlook, Teams and SharePoint. With a Microsoft 365 Copilot licence it can ground its answers in your organisation's data through Microsoft Graph: emails, files, chats and meetings that the user can already open. It drafts documents, summarises meetings and threads, and finds content without a manual search.
Microsoft has renamed its Copilot products several times, so check what you are buying. Microsoft 365 Copilot Chat is available to users with an eligible work account and is grounded in the web, plus any files a user uploads; agents can extend it. The paid Microsoft 365 Copilot licence adds grounding in your tenant data and Copilot inside the Office apps. Most of the risk in this guide comes from that tenant grounding.
Copilot works within the user's existing permissions. It does not read everything on every request: it retrieves content that is relevant to the prompt and that Microsoft's Semantic Index and search can reach. But it makes content easier to find, so a file that was technically accessible and practically hidden becomes a short question away.
Copilot does not create new access rights. It makes existing ones easier to use, both by staff who stumble on content and by an attacker who has taken over an account.
Is Copilot worth it? Measure net value
It depends on your use cases, and a pilot is the way to find out. Published figures vary. In a Microsoft survey of about 1,300 Copilot users, respondents reported saving 14 minutes a day on average and the most efficient users about 30; Microsoft found that around 11 minutes a day was the point where users said they saw value. The UK cross-government Copilot experiment reported an average of 26 minutes a day. Both figures are self-reported estimates, and your results will depend on the work your people do.
Time saved is a gross figure. What you need is net value: the benefit minus licences, support, training and the time people spend checking Copilot's output, which can be wrong or incomplete. Define a baseline before the pilot and measure against it.
| Measure | What to compare |
|---|---|
| Time to an approved output | Time to produce an approved report or reply, including review, with and without Copilot |
| Corrections | How often output needs factual fixes or rework |
| Adoption | Active use per licensed user over several weeks, not first-week curiosity |
| Quality | Reviewer ratings or error rates for a sample of outputs |
| Net value | Measured benefit minus licences, support, training and verification time |
Deciding not to deploy broadly, or deploying only to the roles where the numbers work, is a legitimate result of a pilot. Copilot is a paid add-on licensed per user, so buy against defined use cases rather than buying broadly and measuring later.
You need a business sponsor who owns the use cases and the value case, and an IT lead who owns permissions, security and prerequisites. Without the first you get a technically sound deployment nobody uses. Without the second you push adoption into an unprepared data environment.
Permissions: the risk you inherit
Most Microsoft 365 tenants carry years of permission drift: sites shared with "Everyone except external users", libraries opened for a project and never closed, anyone-with-the-link sharing that is still active, and Teams whose membership has grown beyond the original group. Copilot does not cause these problems. It makes them visible to anyone who asks the right question, such as a salary question that finds a payroll file nobody locked down.
Review and fix permissions before the pilot starts, beginning with the sites that hold personal, financial, HR and legal data. Our article on Copilot governance and confidentiality covers the oversharing scenarios and UK GDPR implications in more depth. SharePoint Advanced Management, which Microsoft includes for tenants with Copilot licences, gives you tools to find and contain the problem:
| Tool | What it does |
|---|---|
| Data access governance reports | Identify sites with broad access, sensitive content or heavy sharing |
| Site access reviews | Ask site owners to review and correct who has access |
| Site attestation policies | Ask owners to confirm periodically that the site is still needed and that ownership, membership and sharing are correct; non-response can trigger notifications, restrictions or archiving |
| Restricted Content Discovery | Temporarily keep selected sites out of Copilot and organisation-wide search while you remediate |
| Restricted Access Control | Limit a site to members of a specified group |
None of these replace remediation. Restricted Content Discovery hides a site from discovery but leaves its permissions as they were. An owner's attestation confirms the site and its membership; it does not prove that every document inside is correctly classified.
Labels, DLP and retention
Sensitivity labels from Microsoft Purview classify content. Depending on how you configure them, they can also add visual markings, apply encryption that restricts who can open a file, and drive DLP policies. A label that only classifies does not restrict access by itself.
Copilot respects labels in specific ways, and the behaviour differs by app. When Copilot in Word, PowerPoint or Outlook creates new content from a labelled item, the new content inherits the label; when it draws on several labelled sources, it takes the highest-priority one. Copilot Chat shows the highest-priority label of the sources behind a response, but labels on Teams meetings and chats, and container labels on teams and SharePoint sites, are not passed on to the content Copilot returns. Users can replace an inherited label, and remove it where labelling is not mandatory, and inheritance does not work for content encrypted with user-defined permissions. Encrypted content is only returned to users whose usage rights include copying content (the EXTRACT right). Test your configuration with an authorised user and an unauthorised one before the pilot, rather than assuming the behaviour.
Keep the label taxonomy small enough that people can choose correctly without thinking hard. A long list of labels and sub-labels produces random choices.
Purview DLP has a policy location for Microsoft 365 Copilot, separate from your endpoint and email DLP. It can stop Copilot from processing files and emails with specified labels, and Microsoft has added options to restrict prompts that contain sensitive information types and to control web searches; some of these options are still in preview. Check the current limits. For example, Microsoft documents that a blocked file can still appear as a citation, that label-based blocking covers emails sent on or after 1 January 2025 but not calendar invites, that in Word, Excel and PowerPoint the policy is evaluated when a file is opened, and that DLP does not scan files uploaded directly into a prompt. Licensing requirements apply.
Old content is a quality problem as well as a risk: Copilot will surface an outdated policy alongside the current one if both are accessible. Replace any blanket "archive everything older than X" rule with retention policies tied to purpose, owner and legal holds. UK GDPR requires that you keep personal data no longer than necessary for the purpose, so your retention rules and your Copilot preparation should line up.
Technical prerequisites checklist
Verify these before the pilot. Microsoft's Copilot requirements page is the reference and changes over time.
- LicencesAn eligible Microsoft 365 or Office 365 base licence, plus a Microsoft 365 Copilot licence assigned to each pilot user.
- IdentityUsers in Microsoft Entra ID with MFA. Review Conditional Access policies and their exclusions so Copilot works for the right users and a stolen account is harder to use.
- NetworkAllow the Microsoft 365 endpoints Copilot needs. Check proxies, SSL inspection and supported browsers, because inspection and blocked endpoints can break Copilot features in ways users find hard to report.
- Microsoft 365 AppsCopilot in the desktop apps needs a supported build on Current Channel or Monthly Enterprise Channel; it is not available on Semi-Annual Enterprise Channel. If you run Semi-Annual, assess which line-of-business add-ins and macros depend on it before you move. That testing is often the longest lead time in the project.
- Audit and information protectionAudit logging on, so you can see Copilot interactions. Sensitivity labels and DLP configured as above.
- Agents and connectorsDecide who may build, share and approve agents and Graph connectors before anyone starts, and which permissions they may use.
Prompt injection and EchoLeak
Prompt injection is the attack class to understand. Instructions hidden in an email, document or web page that Copilot reads can try to change what it does: give a misleading summary, leave out a warning, or push the user towards a malicious link. Hidden text, such as white-on-white text or metadata, makes this harder for a human to spot. Whether an injection succeeds, and what damage it can do, depends on Microsoft's product protections, what the user can access, and above all which tools and actions are available. A Copilot that can only read and summarise has a smaller blast radius than an agent that can send email or update records. The OWASP GenAI Security Project lists prompt injection as the top risk for LLM applications.
EchoLeak (CVE-2025-32711) shows what a chained attack can look like. Aim Security researchers disclosed it in June 2025: a crafted email could cause Microsoft 365 Copilot to include data from the user's context in a request to an external server, with no click from the victim. The chain combined prompt injection with bypasses of Microsoft's injection classifiers and link redaction. Microsoft fixed it on the service side before disclosure and reported no evidence of exploitation in customer environments. It does not mean every Copilot summary can leak data. It does show that an AI assistant inside your data environment needs its own threat model.
Practical controls: fix permissions, label and protect sensitive content, keep Copilot and agents away from actions they do not need, log interactions, and give users a simple way to report odd behaviour.
Agents need their own assessment
Agents built in Copilot Studio or added from Microsoft and third parties connect Copilot to other systems: a ticketing tool, the CRM, HR or finance. Each one needs its own assessment, separate from the Copilot rollout.
Start by working out which identity runs each action. An agent can act with the user's delegated permissions, through a connector's stored credentials, or as an app or workload identity. The answer decides what an attacker or a confused user can make the agent do.
- Query-only agents should have read-only access to the sources they need
- Agents that write or trigger workflows need narrowly scoped permissions, rate and value limits, and authorisation checks enforced outside the model
- High-impact actions, such as payments, deletions or sending data externally, need human approval
- Credentials belong in a secrets store and should expire; never put them in an agent definition
- Share agents with the people who need them, not the whole tenant
Keep a register with one entry per agent: owner, purpose, data sources, identities, connectors, permitted actions, limits, logs and how to suspend it. Our article on agentic AI security risks covers the threat model in more detail.
Pilot, adoption and timeline
Pick a pilot group that represents the roles you would license later, not only enthusiasts. Give each team specific use cases, such as meeting summaries, first drafts of standard reports or email triage, and include at least one security-sensitive scenario, for example a user asking for content they should not see. Write down the success criteria before the pilot starts.
Output quality depends on how people ask. Short prompt training during the pilot, and a shared library of prompts that work for your tasks, help more than an announcement email. People in each team who share what works are more persuasive than IT.
Measure with care. The Copilot Dashboard in Viva Insights shows adoption and estimated impact based on Microsoft's assumptions; it does not measure time saved for each person. Compare against your baseline, or against a group without Copilot, and do not use the data to rank individuals' productivity.
How long all of this takes depends on your permission debt, the state of your labels, your update channel and how many agents you plan. Plan from your own scope. As an illustration only, a mid-sized organisation with moderate permission problems might spend a few weeks on remediation, run a pilot for six to eight weeks, and then expand in stages.
Go or no-go before broad rollout
Use the pilot to answer each question below. Any "no" is a reason to fix first or narrow the rollout.
| Area | Ready when |
|---|---|
| Permissions | High-risk sites are reviewed and remediated, and broad sharing is under control |
| Data protection | Labels and Copilot DLP behave as expected in tests with authorised and unauthorised users |
| Identity | MFA and Conditional Access cover all Copilot users |
| Product readiness | Apps are on a supported channel and network checks pass |
| Legal compliance | Your DPIA or legitimate interests assessment covers Copilot processing, and your records and privacy notices are updated where needed |
| User readiness | Training, use cases and a route for reporting problems are in place |
| Security validation | Pilot logs and the security scenario showed no unexpected access |
| Business value | Net value against the baseline justifies the licences for the roles you plan to add |
| Agent governance | Every agent has an owner, a register entry and a completed assessment |
After rollout, review permissions and the agent register on a schedule, track net value rather than licence counts, and keep the reporting route open.
Whether Copilot is worth it depends on your use cases, and a measured pilot will tell you. Whatever the answer, fix permissions, protect sensitive content and settle agent governance first, because Copilot will find whatever your users can already reach.