Chapter 15: Microsoft Copilot From Beginner to Advanced User
A quick note before you begin
Snapshot taken in the third quarter of 2026. Copilot’s names, licenses, and boundaries have changed more than once in the time it took to write this book, and at least one significant change removed a capability people were already using. Verify at Microsoft’s official pricing pages and at the Microsoft 365 roadmap, which is unusually good and publishes what is coming.
Four different products called Copilot
Microsoft has attached this one word to several genuinely different things, at several different prices. When a colleague says the company has Copilot, the useful reply is: which one?
The free chat assistant. A general web assistant at copilot.microsoft.com and in Windows. It answers questions, drafts text, and searches the web. It cannot see your organization’s files or mail, and it is not what the advertising is about.
Copilot in consumer Microsoft 365. AI inside your own Word, Excel, PowerPoint, Outlook, and OneNote, included with personal and family subscriptions, subject to a monthly allowance of credits. A higher-priced consumer plan removes those limits. This is for your own documents; it cannot reach a work account’s data.
Microsoft 365 Copilot. The work version, sold as a per-seat addition to a qualifying business or enterprise plan. This is the expensive one, and the only one that is grounded in your organization’s data.
GitHub Copilot. A coding assistant, separately licensed and separately billed. Unrelated to your spreadsheets.
Alongside these sits Copilot Studio, for building custom agents, billed by consumption rather than by seat.
The distinctions are not pedantic. In April 2026 Microsoft moved the in-application features inside Word, Excel, PowerPoint, and OneNote behind a paid license, and a large number of people opened Word one morning to find something they had been using every day was simply gone. Knowing which product you hold is the difference between a workflow and a disappointment.
Grounding: the whole argument, and the whole risk
The paid work version’s case rests on one idea. It is connected, through Microsoft Graph, to your organization’s mail, documents, chats, meetings, and calendars, and it answers from them.
That changes the character of the questions you can ask. Not write me a project update, but write me a project update from the last three weeks of this Teams channel, the two documents in this folder, and Thursday’s meeting. The alternative to that is a person spending forty minutes assembling context by hand.
The risk arrives in the same sentence. Copilot inherits your existing permissions: it can see what you could already have opened. In most organizations, more is shared more broadly than anyone believes, and it has been invisible only because nobody was searching for it. Copilot searches for it. Deployments frequently uncover a permissions problem that predated the AI by a decade, which is an argument for tidying up before rolling out, not for avoiding the product.
Copilot in Word
Drafting. Generate a first draft from a prompt, and — more usefully — from documents you point it at. Draft a proposal based on this specification and last year’s version is the shape of request that earns its keep.
Summarizing. Long documents, contracts, and PDFs reduced to their substance, with the ability to ask follow-up questions about specific sections. This is the most-used feature in the product for a reason.
Rewriting. Select a passage and ask for it shorter, plainer, more formal, or restructured. Working on a selection rather than the whole document is the difference between editing and regenerating.
Review workflows. Copilot can now edit with Track Changes enabled, so its revisions arrive as proposals you accept or reject rather than as silent alterations. It can also read, add, and reply to comments, and apply an edit that a comment asks for. If you circulate documents for review, turn this on and leave it on: visible changes are the only kind you can supervise.
Document furniture. Tables of contents that update as the document evolves, headers, footers, page numbers, footnotes. Unglamorous, and a real recovery of time for anyone producing long formatted documents.
Agent mode. Multi-step work inside the document, planned and then executed, rather than one instruction at a time.
Copilot is genuinely useful for first drafts of structured business documents assembled from material you already have, long summaries, and mechanical formatting. It remains weak where your individual voice, judgment, or an exact number or name matters. Familiar AI failure modes still apply inside Microsoft applications.
Copilot in Excel
Excel deserves its own treatment, because the gap between what Copilot does well here and what people expect it to do is wider than anywhere else in the suite.
Formula generation and explanation. The highest-value everyday use, and the one to start with. Describe the calculation in plain language and get a formula, or paste in something inherited and incomprehensible and ask what it does. Ask for the explanation alongside the formula, then put it in yourself. You will learn Excel faster this way than from any course, and you retain the ability to see what the sheet is doing.
Analysis and presentation. Highlight trends, suggest charts and PivotTables, add a calculated column, apply conditional formatting, sort and filter, and answer questions about a range.
Advanced analysis with code. For heavier statistical work, Copilot can generate and run Python within the workbook rather than producing numbers from its own impression of what they should be. The distinction between a computed figure and an asserted one is central whenever AI hands you a number. Ask which process produced it.
Agent mode. The most consequential and the most dangerous feature in the application. Given a broad instruction, it will build: create sheets, write linked formulas, construct a dashboard, produce a model. It is remarkable when you are starting from nothing.
Four cautions, learned the hard way by a great many people:
• It edits your live workbook, and there is generally no preview. Work on a copy. Always.
• It builds better than it maintains. On a fragile, mission-critical workbook that has evolved over years, incremental help is safer than delegation. Ask it to explain and to suggest; do the surgery yourself.
• It has structural requirements. Data laid out as a proper table, files stored where the service can reach them, and a saved rather than an unsaved workbook. Most reports of Copilot “not working” in Excel are one of these three.
• Verify the arithmetic. A wrong number in a spreadsheet looks exactly like a right one, and it propagates into every cell that depends on it. Treat spreadsheet calculations as high-consequence work whenever other decisions depend on them.
Copilot in Outlook, Teams, and PowerPoint
Outlook. Summarize a long thread, draft a reply from a rough instruction, rewrite a selected portion rather than regenerating the whole message, and get a briefing before a meeting. Thread summarization is what people keep.
Teams. This is where the paid version most often justifies itself. Real-time transcription, a recap of a meeting you missed, decisions and action items with owners attached, and the ability to ask what was said about a topic across an hour of conversation. For organizations that live in meetings, this single capability decides the purchase. Note the obvious consequence: meetings become searchable records, which changes what it is wise to say in them.
PowerPoint. Build a deck from a Word document, apply the organization’s approved template, generate images and speaker notes, and restructure an existing presentation. Treat the output as a first draft with the right slides in roughly the right order, not as a finished deck.
Level three: Delegation
Agents. Configured assistants for a defined job — a planner, a researcher, a departmental specialist — built in Copilot Studio and run against your organization’s systems. This is where Microsoft is placing its strategic weight, and where costs move from predictable per-seat licenses to consumption.
Copilot Cowork. Handing over whole multi-step tasks rather than conducting them turn by turn.
Notebooks. A place to gather documents, mail, meetings, and pages as a defined source set, so answers come from that material rather than from everything at once. The same grounding principle described in Chapter 14 under NotebookLM, and equally effective for the same reason.
Connectors. Links to systems outside Microsoft’s world, so Copilot can answer from your ticketing system or your knowledge base.
Administrative controls. Data-loss prevention applied to prompts, sensitivity labels, audit trails, and reporting on what is being used. For regulated industries this tooling is a genuine differentiator and is frequently the reason Copilot wins the evaluation.
Where it shines
Held loosely:
• Meetings: transcription, recaps, action items. The strongest single feature in the product
• Any organization whose work already happens in Outlook, Teams, and SharePoint
• Governance, compliance, and administrative control, at a level competitors do not match
• Formula help in Excel, and summarization of long documents in Word
• Being present where the work is, which removes the copy-and-paste step and the exposure that comes with it
Current limitations
Cost stacking. The advertised per-seat price is an addition to a qualifying subscription you must already hold. The real figure is the two together, and it is often two to three times the number people quote. Price it properly before proposing it.
Uneven quality across applications. The chat and meeting capabilities have been consistently stronger than the in-document ones. Expectations set by the marketing have collided with this repeatedly, and disappointed pilots are usually disappointed about Word and Excel specifically.
Grounding is only as good as your housekeeping. If your files are scattered, duplicated, and misfiled, Copilot will confidently summarize the wrong version. It has no way to know which document is the current one unless your organization does.
Naming and entitlement confusion. Covered above, and worth repeating because it wastes more time than any technical limitation.
A structural point is worth understanding. Copilot runs on models built by other companies, including OpenAI and, more recently, Anthropic. The model varies by feature and may sometimes be selectable. The assistant you use and the intelligence underneath it are increasingly separate products, and the company selling the assistant may not have built the model.
The subscription decision
For individuals. Check what you already hold. Consumer Microsoft 365 subscriptions now include AI in the applications with an allowance attached, and the higher consumer plan removes the limits. If you were about to buy a general assistant subscription and you already pay for Office, work out what you have before adding another.
For organizations. Do not buy it for everyone at once. License a pilot group whose work is genuinely meeting-heavy or document-heavy, define in advance what improvement would look like, and measure it after a quarter. Then license by role. Organizations that license it broadly often discover that a substantial share of those seats receive little use, and the arithmetic on a per-seat product is unforgiving.
And before the pilot: run the permissions review. Discovering your access problems through a pilot is expensive; discovering them through Copilot answering the wrong person’s question is worse.
Privacy and governance settings to check
Which account, again. Consumer Copilot and work Copilot are different products with different data terms. Work material belongs in the approved work account.
Enterprise data protection. On business and enterprise licenses, your organization’s prompts and content stay within your tenant and are not used to train the underlying models. Consumer terms are different. Know which set applies to you and do not assume the stricter one.
Retention and discoverability. Copilot interactions in a work tenant are records: retained, subject to your organization’s policies, and discoverable. Meeting transcripts especially.
Sensitivity labels and data-loss prevention. If your organization has them, they apply to Copilot too. If it does not, that is worth raising before deployment rather than after.
Chapter 25 assembles this into a policy you can actually follow.
Exercise 15.1: The formula loop
Open a spreadsheet you use. Find a calculation you have been doing manually, or a formula you inherited and never understood.
Ask for the formula and an explanation of each part, then type it in yourself and check it against a case where you know the answer. Repeat weekly for a month. This is the fastest route to competent Excel that currently exists.
Exercise 15.2: Summarize a meeting you attended
Take a recorded meeting you were in and ask for a recap with decisions and action items.
Compare it against your own memory and notes. What did it catch that you missed? What did it attribute to the wrong person, or record as a decision when it was only a suggestion? That second list is your calibration, and it is the more valuable one.
Exercise 15.3: Establish which product you have
Find out, definitively: which Copilot license is attached to your account, whether it can see your organization’s files, and what your administrator has enabled or blocked.
It takes one email and it prevents months of confusion about why the instructions in a guide do not match what is on your screen.
Learn More About Artificial Intelligence
Learn more about the CLEAR, TRUST and LEARN methods for Artificial Intelligence