How Does Microsoft 365 Copilot Work? A Plain Guide

TL;DR:

• Microsoft 365 Copilot sits between your staff, your Microsoft 365 data and an AI model. It takes a prompt, finds relevant work content, sends both to the model, then processes the answer and returns it inside Word, Outlook, Teams or the Copilot app.
• Free Copilot Chat, included with eligible Microsoft 365 work accounts, uses the web and whatever a person pastes, uploads or has open. The paid Microsoft 365 Copilot licence adds automatic access to that person's emails, files, chats, meetings and calendar through Microsoft Graph.
• Copilot can only show a person what they can already open. If files are overshared, Copilot finds them fast. Fixing permissions is the real rollout project.
• Microsoft says prompts, responses and Graph data are not used to train its foundation models. For customers outside the EU, model processing can happen in the US, the EU or other regions, and Kenya is not on Microsoft's data-at-rest list for Copilot (as of November 2026).
• Admins control web search, agents, third-party models such as Anthropic's Claude, labels, data loss prevention, retention and audit. Check each setting before the paid licence goes live.

Who this guide is for

This is for managers, IT leads and operations heads in Kenyan organisations that use Microsoft 365 and are deciding how to switch on Copilot.

Our companion post covers training, pricing and a 90-day rollout plan: How Kenyan teams should approach Microsoft Copilot training. This one goes under the bonnet, so your decisions rest on how the system really behaves.

First, the names

Microsoft renames Copilot products often. As of November 2026, Microsoft Learn describes three licence levels for work accounts, plus extra pay-per-use features on top of the paid licence:

The Microsoft 365 Copilot app is now simply called Microsoft Copilot. In this guide, "the paid licence" means the Microsoft 365 Copilot add-on licence.

How Copilot answers a prompt, step by step

Copilot looks up relevant content the person is allowed to see, adds it to the prompt, sends that to an AI model, then processes and returns the answer. Microsoft's documentation on the semantic index describes five steps:

1. You ask. A staff member types a prompt in Word, Outlook, Teams or the Copilot app.
2. Copilot gathers context. It searches Microsoft Graph and the semantic index for relevant emails, files, chats and meetings. This is called grounding. It only searches content this one person can already open.
3. Copilot sends a bigger prompt to the model. Your words plus the snippets it found go to a large language model.
4. The model drafts an answer.
5. Copilot finishes and returns it. It goes back to Graph and the semantic index for post-processing, then sends the answer to your app. Stored answers include citations to the sources used. Microsoft says requests are encrypted in transit and customer data stays encrypted at rest.

Two things follow from this design:

• The model sees a small bundle each time. For each prompt it gets a targeted set of extracts, never the whole company.
• Answer quality depends on file quality. If SharePoint holds five versions of the same policy, Copilot may ground its answer in the wrong one.

A worked example

Imagine a procurement officer with the paid licence asks Copilot in Teams: "What did we agree with Supplier A about delivery dates?"

Copilot searches her mailbox, her Teams chats, meetings she attended and SharePoint sites she can open. It finds an email thread, a meeting recap and the signed contract. It sends her question plus short extracts from those three to the model, and returns a summary with links to each source.

A colleague in finance asks the same question. He was not on the thread, missed the meeting and has no access to the procurement site. Copilot finds nothing from those sources for him. Same question, different answer, because the permissions differ.

Now suppose the procurement site was shared with the whole organisation by mistake three years ago. The finance colleague can technically open it, so Copilot can find the contract for him. That is the oversharing problem in one picture.

What is Microsoft Graph, in plain words?

Microsoft Graph is the connected map of your Microsoft 365 tenant: people, emails, files, chats, meetings, calendars and how they relate. Microsoft says the paid Copilot uses Graph to bring "personalized organizational data context into the prompt".

Think of Graph as the company filing room and permissions as the keys each person carries. Copilot can only enter rooms that person's keys open, but it reads every shelf in those rooms very quickly.

The semantic index

The semantic index is a search index that understands meaning as well as exact words. Microsoft's example: ask for an email where a colleague "praised" a vendor, and it also looks for words like "elated" and "amazed". As of November 2026, Microsoft Learn says:

• It is built automatically for organisations with paid Copilot licences, needs no setup, and cannot be switched off as a whole.
• It only shows results to people who already have access.
• Index data stays inside your tenant.
• Organisations without Purview data loss prevention can exclude a sensitive SharePoint site (Microsoft gives payroll, HR and financial information as examples) by setting it not to appear in search results. That removes it from normal search and the semantic index together.

Free Copilot Chat versus the paid licence

The key difference is whether Copilot goes looking for your work content by itself.

Source: Microsoft Learn, "What is Microsoft Copilot?", updated 10 November 2026.

For prices and how to pick the first licence holders, see our Copilot training guide for teams.

Permissions and oversharing: the real risk

Copilot works with the access people already have, so old sharing mistakes become easy to find. Microsoft states it plainly: "If content is already overshared, Copilot can surface it to anyone who has access."

Oversharing builds up quietly: a project site shared with everyone, a payslip folder with a company-wide link. With the paid licence, a staff member asking "what are the salary bands for managers?" may get an answer from a file they never knew they could open.

How Copilot treats protected files

Microsoft Purview sensitivity labels and encryption change what Copilot can do. As of November 2026, Microsoft Learn says:

• Encrypted files need two rights. The person needs both VIEW and EXTRACT rights for Copilot to summarise an encrypted file. With VIEW only, Copilot does not summarise it but can return a link.
• Labels carry over. New content Copilot creates from labelled files can inherit the highest-priority label.
• Data loss prevention can stop summaries. A Purview DLP policy for Copilot can stop it summarising labelled files and emails.
• Teams is a gap. Copilot does not currently recognise sensitivity labels that protect Teams meetings and chats.

Microsoft's deployment blueprint starts with fixing oversharing, and it says SharePoint Advanced Management is included with the Copilot licence. Microsoft Learn names tools such as Restricted Content Discovery (hiding chosen sites from Copilot and organisation-wide search while you review them) and site access reviews. Our companion post has a full permissions checklist, so we will not repeat it here.

Where your data goes

What is stored

• Prompts and responses, with citations, are stored encrypted under the same terms as your email and files. They form each person's Copilot activity history.
• A hidden folder in each user's mailbox holds them. Admins can search them with eDiscovery and manage them with retention policies.
• Audit logs record who used Copilot, when, and which files or sites it used. They do not contain prompt text; admins need eDiscovery or Purview's Data Security Posture Management for AI to see that.
• Users can delete their own Copilot history in the Microsoft My Account portal.

Microsoft states: "Prompts, responses, and data accessed through Microsoft Graph aren't used to train foundation LLMs." It also says Copilot has opted out of the human abuse-monitoring review available in Azure OpenAI.

Where processing happens

• Microsoft routes model calls "to the closest data centers in the region" but can use other regions when demand is high.
• EU traffic stays inside the EU Data Boundary. Microsoft says "Customers outside the EU may have their queries processed in the US, EU, or other regions."
• Microsoft's data-at-rest commitment for Copilot lists sign-up countries such as South Africa, the UAE, India, the UK and the US. Kenya is not on that list (Microsoft Learn page dated 8 October 2026). Your admin can check where your tenant's data actually sits in the Microsoft 365 admin center.

Web search goes to Bing, under different terms

With web search on, Copilot writes a short search query of a few words from your prompt, removes your user name, domain and tenant ID, and sends that to Bing. It also sends your device's IP address so Bing can localise results. Microsoft's Product Terms say these queries are not used to build advertising profiles, not shared with advertisers and not used to train generative AI foundation models.

However, Microsoft's Data Protection Addendum does not cover these web queries. For them, Microsoft acts as a data controller, under the Microsoft Services Agreement plus extra commitments in its Product Terms. Admins can turn web search off, or off for work mode only, with the "Allow web search in Copilot" policy in the Cloud Policy service.

Third-party models, including Claude

As of November 2026, Microsoft says Claude models from Anthropic are "on by default for most customers in commercial cloud (excluding EU/EFTA and UK)". Microsoft adds that they appear only in some licensed experiences. Unless an admin changed the setting, a Kenyan tenant is likely to have them on where they apply. They run under Microsoft's Product Terms and Data Protection Addendum, but are excluded from the EU Data Boundary and in-country processing commitments. Some advanced "Anthropic models with Data Retention" stay off until an admin opts in under Anthropic's own terms.

An AI Administrator or Global Administrator can switch Anthropic off, or limit it to certain groups, under Copilot settings, "AI providers operating as Microsoft subprocessors". Either choice is reasonable. Make it on purpose.

What this means under Kenya's Data Protection Act

Copilot does not change your duties under the Data Protection Act, 2019. It adds new places where personal data is processed. This is not legal advice; talk to your data protection officer or lawyer.

• Transfers out of Kenya. Section 48 allows personal data to go to another country only on set conditions, such as proof to the Data Commissioner of appropriate safeguards. The law firm ENS reports that the ODPC's "Guidance Notes for Cross-border Data Transfers, 2026", published on 8 September 2026, treat storing personal data on cloud servers outside Kenya as a cross-border transfer. ENS describes the notes as explaining how the ODPC expects the Act to be applied; they are guidance, not a new law.
• Sensitive personal data. Section 49(1) says sensitive personal data may be processed out of Kenya only with the person's consent and confirmation of appropriate safeguards. Take extra care with HR, medical and welfare files.
• Impact assessments. Section 31 requires a Data Protection Impact Assessment before processing likely to result in high risk to people's rights and freedoms. Copilot over HR or customer records is a good reason to check. Microsoft publishes a "Microsoft Copilot & Copilot Chat Risk Assessment Quickstart" in its Service Trust Portal that can feed into one.

For more on the Act and everyday AI tools, see Can staff use ChatGPT under Kenya's Data Protection Act?.

What admins control

Your IT lead should know the current setting for each of these (Microsoft Learn, November 2026):

Built-in safeguards, and their limits

Microsoft lists protections that run alongside the model. It says some of them are not available in every Copilot scenario.

• Prompt injection defences that try to block jailbreak prompts and hidden instructions inside documents or emails.
• Harmful content filters for hate, sexual content, violence and self-harm.
• Workplace harm filters. In certain Copilot scenarios, Microsoft says it restricts Copilot from judging an employee's performance, attitude or emotional state from their workplace communication. Managers should not ask Copilot "who is disengaged?"

None of this makes the output correct. Microsoft says generative AI responses "aren't guaranteed to be 100% factual." A person checks every figure, clause and name before anything goes out.

Five questions to settle before the paid licence

1. Who owns the permissions clean-up? Name one IT person and one business owner for each major SharePoint site.
2. Which sites stay hidden at first? HR, payroll, board papers, legal and disciplinary files are common candidates.
3. What is our position on web search and third-party models? Decide, write it down and tell staff.
4. How long do we keep prompts and responses? Set a Purview retention policy that matches how you treat email.
5. What is the staff rule? Work accounts only, check everything, and no decisions about people made by Copilot alone. Our AI acceptable use policy template for Kenya is a starting point.

For the leadership view, see What should a CEO ask before an AI rollout in Kenya?.

Prompts that use Copilot's design well

Knowing how grounding works makes prompts better. These need the paid licence. Replace the brackets with your own details.

Point at the source. Microsoft notes that typing "/" lets you reference a specific file directly.

Using /[Q3 Operations Report], list the three biggest cost overruns, with the section each figure comes from. If a figure is not in the report, say "not in the source".

Scope to one library. Microsoft says attaching a SharePoint library or folder narrows the search and, in the web experience, uses its column metadata.

Using only the [Contracts 2026] library, list every contract that expires before [date], with the supplier name and file link. Leave out drafts.

Ask for sources, then check them.

Summarise what was agreed in my meetings with [team] this month. Cite the email, chat or meeting for each point, and mark anything uncertain with [CHECK].

These follow the four Cs we teach: Clarity, Context, Constraints and Customisation.

What goes wrong, and how to avoid it

Questions people ask

Can Microsoft 365 Copilot read everyone's email?

No. Copilot only uses content the signed-in person can already open. With the paid licence it searches that person's own mailbox and the files, chats and meetings they can reach. Free Copilot Chat does not search mailboxes or files on its own.

Is free Copilot Chat safe for confidential work?

With a work account, Copilot Chat is covered by Microsoft's enterprise data protection, the same contract terms as your email, and Microsoft says prompts are not used to train its foundation models. The bigger risk is staff using personal accounts or unapproved tools.

Is our Copilot data stored in Kenya?

As of November 2026, Kenya is not among the countries listed in Microsoft's data-at-rest commitment for Copilot, and model processing outside the EU can happen in the US, the EU or elsewhere. Your admin can check your tenant's data location in the Microsoft 365 admin center.

Can admins see what staff type into Copilot?

Yes, with the right Purview tools and roles, because prompts and responses are stored in a hidden mailbox folder that eDiscovery can search. Say so openly in your AI policy.

Can Copilot make HR decisions?

It should not. Microsoft restricts Copilot from judging employees from their communications, and section 35 of the Data Protection Act gives people the right not to be subject to a decision based solely on automated processing, including profiling, that produces legal effects for them or significantly affects them.

Sources

All sources last checked 11 November 2026.

• Microsoft Learn, "What is Microsoft Copilot?" (updated 10 November 2026)
• Microsoft Learn, "Data, Privacy, and Security for Microsoft Copilot" (updated 4 November 2026)
• Microsoft Learn, "Data protection and auditing for Microsoft Copilot" (updated 9 November 2026)
• Microsoft Learn, "Semantic indexing for Microsoft Copilot"
• Microsoft Learn, "Enterprise data protection in Microsoft Copilot and Microsoft Copilot Chat"
• Microsoft Learn, "Data, privacy, and security for web search in Microsoft Copilot and Microsoft Copilot Chat"
• Microsoft Learn, "Claude models from Anthropic in Microsoft Online Services"
• Microsoft Learn, "Secure & Governed Data Foundation for Microsoft Copilot: deployment blueprint"
• Microsoft Learn, "Data Residency for Microsoft 365 Copilot Offerings" (8 October 2026)
• Microsoft Learn, "Microsoft 365 services data locations"
• Office of the Data Protection Commissioner, The Data Protection Act, No. 24 of 2019 (sections 31, 35, 48 and 49)
• ENS via Afriwise, "Data Without Borders? What Kenya's New Cross-Border Data Transfer Guidance Means for Businesses" (28 October 2026)

Where FuKazee fits in

FuKazee trains Kenyan teams on Microsoft 365 and Copilot for the work in front of them. We assume nobody in the room knows AI yet, we avoid jargon, and we start with free Copilot Chat so everyone can practise before anyone buys licences. We teach the four Cs, how grounding and permissions work, and how to check every output. Permissions settings and legal sign-off stay with your IT team and counsel; we help staff understand why those decisions matter. Since 2022 we have trained more than 3,000 people.

See how our sessions work on the Microsoft Copilot training in Kenya page.

To plan Copilot training for your team, email us at lead@fukazee.com