Microsoft Copilot: The Complete Training Guide for 2026

Two years after Microsoft 365 Copilot landed on enterprise price lists, the question is no longer whether to license it. It is who gets a seat, what they will do with it, and how to measure whether the licence pays for itself. The companies that figured out the second and third questions early — Klarna, KPMG, Vodafone, and a long tail of mid-market firms — have stopped publishing breathless case studies and started quietly raising productivity targets. The companies that haven't have a few thousand seats and a vague feeling that nothing has changed.

This guide treats Copilot as one product family across the Microsoft stack — the same brand, very different licences, very different capabilities. By the end you'll know what each variant actually does, what it costs in 2026, where it earns its money in Word, Excel, Outlook, Teams and PowerPoint, and where ChatGPT is still the better tool. If you came for the AI-900 cert, that's covered too.

Table of contents

• What Copilot is across the Microsoft stack
• Pricing: Personal vs Pro vs Microsoft 365 Copilot
• Copilot in Word — the moves that save the most time
• Copilot in Excel — formulas, analysis, charts
• Copilot in Outlook — triage, drafting, summarising
• Copilot in Teams — meeting recaps and action tracking
• Copilot in PowerPoint — one-shot decks from a brief
• Copilot Studio — when you need a custom agent
• Security, governance and the rollout you didn't plan for
• Certifications and AI-900 prep
• Copilot vs ChatGPT — where to use which
• The five rollout mistakes that kill Copilot projects
• Frequently asked questions
• The bottom line

What Copilot is across the Microsoft stack

The word "Copilot" gets attached to at least seven distinct Microsoft products. They share a brand, a chat-style interface, and a base set of OpenAI models behind the scenes — and almost nothing else. If you don't separate them mentally, every conversation about pricing, security, or capability turns into a fog.

The cleanest mental model is to split them by who pays and what data the model sees.

The pivot in 2025 was to separate the agentic offering. Copilot Studio is now the front door for building autonomous agents inside Microsoft's ecosystem — the equivalent of OpenAI's Custom GPTs but with proper tenant security, connectors to SharePoint and Dynamics, and a price point aimed at IT departments rather than individual users.

If a colleague says "we have Copilot," ask which one. Free Copilot helps draft a poem; Microsoft 365 Copilot drafts a board paper from your last quarter's Teams meetings, your finance spreadsheets, and an email thread with the CFO. The licence cost difference is roughly 30x for a reason.

The product line in 2026 is also beginning to consolidate. Microsoft renamed the consumer chat experience to simply "Copilot" in late 2024, dropping "Bing" from the brand entirely. The enterprise product retained "Microsoft 365 Copilot." Copilot Studio remains the platform for building custom agents, while the role-based variants (Sales, Service, Finance) are increasingly positioned as feature add-ons rather than separate products. Watch for further simplification through 2026 — Microsoft's own internal product reviews suggest the "Copilot for X" naming is on the chopping block in favour of skill-based packaging.

Pricing: Personal vs Pro vs Microsoft 365 Copilot

Copilot pricing as of early 2026 has settled into three brackets. The free tier exists, the personal tier is genuinely useful for some people, and the enterprise tier is where most of the productivity claims come from.

Two prerequisites bite people. Microsoft 365 Copilot requires an underlying productivity licence — Business Standard, Business Premium, E3, or E5 at minimum. If your shop runs Business Basic (web-only Office), Copilot won't sit inside Word. The total cost is the base licence plus $30, not $30 flat.

The third prerequisite that catches mid-market companies: regional pricing. The $30 figure is US-list. UK pricing is typically £24.70, EU pricing varies by country (around €28-32), and pricing in Australia, Canada and emerging markets shifts with FX. Volume discounts at Enterprise Agreement scale typically deliver 10-15% off list, but published price is what most sub-500-seat deployments actually pay. Budget the local-currency figure, not the US headline.

The annual commitment is the second one. Microsoft pulled month-to-month M365 Copilot in 2024. You can buy a one-month trial through select partners, but full deployment requires a 12-month commitment per seat. That makes pilot design more important — you can't easily ramp 50 seats up and down.

For the ROI math, the rough rule that has held up since launch: if a knowledge worker saves 30 minutes a day using Copilot, the licence pays for itself at any salary above roughly $40,000. Most internal studies show 60 to 90 minutes a day for the active users, but only 30% to 50% of seats are actively used. The hard part of ROI is not the per-user maths; it is making sure the seats go to people who will actually use them.

Copilot in Word — the moves that save the most time

Copilot in Word is the variant most people meet first. It sits as a button in the ribbon and a side panel, drafts text from a prompt, rewrites selections, and summarises long documents. After two years of internal data, three patterns dominate the savings.

The first is drafting from a brief plus a reference document. Hand Copilot a one-paragraph brief and a reference attached from your OneDrive ("base it on Q3-board-paper.docx"), and it produces a 1,500-word first draft in your house style faster than you can outline. The output is mediocre for anything genuinely original; it is excellent for the 80% of corporate writing that is structured and predictable — status updates, project briefs, internal announcements.

The second is Rewrite Anywhere, an underused feature that landed in late 2024. Highlight any paragraph and Copilot offers three rewrites without leaving the document — shorter, longer, more formal, in a specific tone. It replaces the back-and-forth of pasting into ChatGPT and pasting back. The friction reduction is the productivity gain.

The third is summarising. Drop a 40-page contract into Word, hit Summarise, and Copilot returns a structured outline: key obligations, dates, exceptions, weird clauses. The legal team at a UK insurer reported in early 2025 that initial-pass contract review dropped from 90 minutes to 25 minutes per contract — not because Copilot was right about everything, but because it routed attention to the parts that mattered.

Where it still falls short: anything that requires actual research is unreliable. Copilot in Word will invent statistics if you ask for "supporting data," because the model is producing text, not searching. For research-driven writing, drive the research separately (in Copilot chat with web access, or ChatGPT) and only use Word's Copilot for drafting. The deeper playbook is in our Word-specific guide.

The killer technique most users haven't found yet is "Draft with these as references." When drafting a new document, click the paperclip icon in the Copilot panel and attach up to three OneDrive documents — past briefings, your style guide, source data. Copilot grounds in those documents and the output picks up your house structure, paragraph length, and default disclaimers. The shift from generic LLM prose to recognisably-yours writing is dramatic. Most internal teams that struggle with Copilot in Word have not tried this and dismissed the tool based on first impressions of unattributed output.

Copilot in Excel — formulas, analysis, charts

Excel was the late arrival. Copilot in Excel went generally available in mid-2024 and only became reliable on tables that aren't perfectly formatted in early 2025. By 2026 it has stopped being a demo and started being a tool — but only on data that lives in proper tables.

What it does well: generating formulas from natural language ("for each row, calculate the percentage difference between this month and last month"), summarising a dataset ("what are the three most common reasons for cancellation in this column?"), creating PivotTables from a question, and drafting a chart.

What it still does badly: anything that requires multi-step manipulation across sheets, anything involving macros, and anything where the data isn't in a Format-as-Table-applied range. If your spreadsheet looks like a printable invoice with merged cells and a logo, Copilot is useless. If it looks like a database — header row, one row per record, no merges — Copilot is genuinely fast.

The single highest-leverage move is asking for formulas in plain English. The 2024 Microsoft research found that 73% of office workers use only basic Excel functions; the long tail of Excel productivity is locked behind formula syntax most people never learn. Copilot collapses that gap. "Calculate the rolling 30-day average of column D, ignoring blanks" produces the right LET/AVERAGEIFS construction in two seconds. Our 12 techniques for Excel walks through the moves with the highest savings.

The mental shift Excel users have to make is to stop asking "what formula do I need" and start asking "what answer do I need." The former requires formula-language fluency; the latter requires only knowing the question. This is the productivity unlock for the 70% of office workers who never learned VLOOKUP, INDEX/MATCH, or LET. Their Excel competence has just been quietly upgraded by 30%.

The 2025 Excel update added a related capability worth knowing: Python in Excel can be invoked via Copilot. Ask "use Python to forecast the next 6 months of revenue based on the last 24 months" and Copilot writes a Python forecasting block (using statsmodels or sklearn) that runs in the cloud and returns results into a cell range. This is genuine data science capability accessible via natural language. For most analysts the learning curve is gone.

Copilot in Outlook — triage, drafting, summarising

Email is where Copilot earns its licence for most people, in part because email is where most people lose the most time. Three workflows dominate.

The first is morning triage. Open Outlook on Monday, click the Copilot Summary button at the top of the inbox, and within five seconds you get a one-screen brief: what's urgent, what's in a long thread, who is waiting on a reply, and what scheduled events have shifted. For a typical office worker with 60 to 100 unread messages from the weekend, this saves the 20 minutes of triage they would otherwise spend reading subject lines.

The second is drafting in your voice. Copilot has read the emails you've sent over the last 12 months. When you ask it to draft a reply, the default tone matches your writing — short if you write short, formal if you write formal. It is not perfect, but it skips the cold-start problem of writing the first draft. One bank reported in late 2024 that average reply length dropped 18% after Copilot rollout, and reply time dropped 41%, because people were editing instead of composing.

The third is thread summarisation. The 47-message thread that someone CC'd you on yesterday is now a button-press summary plus a "what does this person want from me?" reading. The privacy boundary holds — Copilot only sees what your account can see — and the time saving on cross-functional projects is substantial. Our Outlook-specific deep dive has the full set of triage and drafting prompts.

The fourth, and the most underrated, is Coaching mode. Write a difficult email — a pushback, a polite no, a concern raised to a manager — and click Coaching before sending. Copilot reads the draft and returns feedback: tone, possible misreadings, missing context, things that might land worse than intended. It is not infallible, but the cost is 30 seconds and the upside is catching a sentence that would have been a problem. For high-stakes emails, this is a free second opinion that costs nothing.

What unifies all four Outlook workflows is that the underlying time saving is "context already loaded." Copilot can read the thread, the calendar, the linked SharePoint document, and the historical sent-mail style without you assembling any of it. That is the privacy-preserving substrate that consumer ChatGPT cannot replicate, and it is why most knowledge workers find Outlook the single Copilot variant they would refuse to give up.

Copilot in Teams — meeting recaps and action tracking

Copilot in Teams is the feature most likely to change how a meeting culture works. The recap and action-item extraction is good enough that note-taking by hand is no longer the default — and that single change ripples through how meetings are scheduled, run, and followed up.

During a meeting, Copilot can answer questions in real time: "What did Maria mean by 'option B'?" It pulls from the live transcript. After the meeting, it produces an automated recap with topics, decisions, action items, and who said what. The action items, importantly, name owners and due dates if anyone said them out loud — which has the side effect of forcing meetings to actually conclude.

The one rough edge is long meetings. Copilot in Teams handles up to 4 hours of transcript reliably; beyond that, summary quality degrades. For all-day workshops, recap by section rather than as a whole.

The other practical use is catching up. If you joined a meeting 20 minutes late or missed it entirely, Copilot will brief you on what happened up to your moment of arrival. The 2025 Microsoft work trend index put the time saving at 35 minutes per meeting missed, on average — substantial in any meeting-heavy culture. The full setup walkthrough is in our Teams meeting recap guide.

The behaviour change Copilot in Teams provokes is the most interesting second-order effect of the entire product family. Meetings now end with explicit commitments because everyone knows the recap will surface them. Vague meetings still happen, but the cost of being vague — being publicly listed as having an unclear action item — has gone up. Several large companies in 2025 reported a measurable drop in unclear handoffs purely because the recap exposed them. This is the rare case where a tool genuinely changes culture rather than just augmenting it.

Copilot in PowerPoint — one-shot decks from a brief

PowerPoint is the Copilot variant that produces the most variable results. The "create a deck from a Word document" feature works — but the deck you get is generic, branded with the default template, and obviously machine-generated. For internal use it is fine; for an external pitch it is a starting point, not a finished product.

Where it earns its keep is restructuring. Hand Copilot an existing deck and ask it to "shorten this from 40 slides to 15, keep the key data, drop the recap sections" and it does so accurately. Likewise the redesign function: "apply the corporate template" produces an instant brand-compliant version of a draft.

The honest assessment: PowerPoint Copilot is a 2x productivity gain on internal slides and a 0.5x productivity gain on external slides, because the polishing time you save up front, you spend later trying to fix Copilot's slightly-off layout choices.

The feature that might rebalance this assessment in 2026 is "Designer-quality" generation, which Microsoft started rolling out in late 2025. The new layouts are visibly less template-y; the colour and font selection is contextual rather than generic. For internal-quality slides this is now production-ready. External-quality decks — investor pitches, customer proposals — still need a designer, but the gap is narrowing. By late 2026 the calculus will likely shift further toward Copilot for both audiences.

Copilot Studio — when you need a custom agent

Copilot Studio is the place where Microsoft 365 Copilot stops being a chat assistant and becomes a platform. It is the answer to two questions: how do you build an agent that knows about your specific company knowledge base, and how do you give it actions inside your business systems?

Practically, Copilot Studio lets you build a chat agent with three things plugged in: a knowledge source (a SharePoint site, a Dataverse table, a public website), a set of topics (conversation flows you've authored), and connectors to take action (raise a ticket, look up an order, send an approval). The output is an agent that lives inside Microsoft 365 Copilot's chat interface, alongside the built-in skills, with branding and a name you choose.

The early use cases that survived 2025 are narrow: HR policy bots ("what is the parental leave policy?"), IT helpdesk tier-zero ("how do I reset my VPN?"), and field-service agents that pull from a service manual. The use cases that didn't survive: anything requiring real reasoning across multiple sources. Copilot Studio is excellent at retrieval and routing; it is not a substitute for an LLM with full agent autonomy. For the broader landscape, see our AI agents pillar.

Pricing is not trivial. The base $200/tenant/month includes 25,000 messages. A modestly used HR bot in a 5,000-person company will burn through that in a week. Plan message budgets carefully or the bill scales sharply.

Security, governance and the rollout you didn't plan for

Most companies that have rolled out Microsoft 365 Copilot have, three months in, discovered a problem they didn't plan for. The model itself is fine. The data plumbing is fine. The licence is paid. What is not fine is the SharePoint sharing audit nobody did before launch.

The pattern: documents that were technically accessible to "everyone in the company" — old strategy decks, pre-redacted offer letters, unfiltered customer feedback — were practically hidden because nobody knew the URLs. Copilot makes those documents discoverable through natural language search. Suddenly an employee can ask "what's our average sales rep on-target earnings" and get a credible answer pulled from a 2022 spreadsheet that should have been Confidential and never was.

The fix is unfashionable but works. Before broad rollout, run Microsoft's SharePoint Advanced Management toolkit (a paid add-on to E5/Copilot tenants) to identify oversharing patterns. Either tighten the permissions on the surfaced sites, or apply sensitivity labels to the riskiest folders to scope Copilot out of them. Companies that skipped this step have spent the rest of their first year doing it reactively. Companies that did it before launch had no incidents.

The wider governance question is which Copilot variants are acceptable for which data. The pragmatic policy most enterprises landed on by mid-2025 looks like this:

The honest enforcement reality is that policy doesn't enforce itself. The technical complement is DLP rules in Microsoft Purview — block unsanctioned AI domains at the network layer, label confidential content so it can't leak via consumer tools, and audit Copilot interactions for unusual patterns. The combination of policy, DLP, and Copilot's enterprise data boundary is what makes the M365 Copilot deployment defensible to legal and compliance teams.

Certifications and AI-900 prep

Microsoft has built a certification ladder around Copilot and Azure AI. The entry rung is AI-900 (Azure AI Fundamentals), and it is the cert most office-track professionals should take if they want a credential to point at.

AI-900 is a 60-minute, 40 to 60 question multiple-choice exam covering AI workloads, machine learning fundamentals, computer vision, NLP, and generative AI. Pass mark is 700 out of 1000. The cost is around $99 USD. It is genuinely passable in 30 to 50 hours of study with no prior background.

Beyond AI-900, the ladder branches. AI-102 (Azure AI Engineer) is the practitioner cert for people building AI services — much more technical, requires Python and Azure familiarity. MS-900 + MS-100 series cover the M365 administration side. For the Copilot-specific certs as of 2026 (MS-721 Copilot Pro and the MS-7XX series for Studio), Microsoft has been iterating; check the official learning path before booking. We've laid out the full prep plan in our AI-900 30-day study guide.

Whether the cert opens doors depends on the role. For a sales rep, internal IT support, or finance analyst, AI-900 is a useful signal that you understand the vocabulary. For a software engineer, the cert says little your code samples don't already say. For consultants and trainers, it is essentially required to pitch yourself as a Copilot specialist.

Copilot vs ChatGPT — where to use which

The honest framing: Copilot and ChatGPT are increasingly the same models behind different doors. The model menu in M365 Copilot (GPT-4 class, GPT-5 once it lands enterprise) overlaps heavily with ChatGPT Enterprise. The interesting differences are about data, privacy, and integration — not raw model capability.

The mistake most companies make is treating these as alternatives. The companies that are getting the most out of AI are buying both — Copilot for the in-app, in-tenant work, and ChatGPT (or Claude) for the more flexible, web-connected work that doesn't touch sensitive data. The marginal cost of the second tool is the price of two coffees a month. The decision matrix is in Copilot vs ChatGPT.

The five rollout mistakes that kill Copilot projects

Two years of post-mortems on Copilot deployments at scale have surfaced the same five mistakes again and again. They have nothing to do with the technology and everything to do with how the rollout is framed.

Licensing everyone on day one. The single most common mistake. Companies that handed out 5,000 seats on the launch announcement reported 20 to 30% active usage at month three. Companies that piloted with 50 to 100 power users for three months and expanded based on what they actually used reported 60 to 70% active usage. The former feels democratic; the latter delivers ROI. The former is also harder to walk back when leadership asks why nobody is using it.

Treating it as software, not behaviour change. The licence is in the ribbon on day one. The habit takes 8 to 12 weeks. Companies that scheduled a single 60-minute training session and then went silent saw poor adoption. Companies that ran a 90-day cadence — onboarding session, daily prompts for three weeks, peer-share session at week 5, follow-up at week 10 — saw substantial adoption gains. The cost difference between these two programmes is roughly $50 per user. The output difference is roughly 4x.

Skipping the SharePoint audit. Already covered in the governance section above. The companies that ran the audit before rollout had zero "Copilot exposed our salary spreadsheet" incidents. The companies that skipped it had at least one and usually several. The audit is unglamorous IT work and it is the difference between a smooth rollout and a six-month firefight.

Choosing the wrong pilot users. The temptation is to give Copilot to executives so they advocate for it. The right answer is to give it to middle managers and individual contributors who actually do the work it is good at — drafting, summarising, meeting follow-up. Executives have assistants who do these things for them; middle managers don't. The user research on which roles get the most value from Copilot is unambiguous: project managers, executive assistants, finance analysts, and front-line sales lead the productivity gains. Pilot accordingly.

Failing to reallocate dormant licences. By month six, 20 to 30% of seats are routinely below five Copilot actions per week. Those are dormant. Reassigning them to people on the waitlist costs nothing and lifts the company-wide active-usage rate by 5 to 10 percentage points. Most enterprises forget to do this. The discipline of quarterly licence reallocation is what keeps the deployment from becoming dead weight on the budget.

Frequently asked questions

What is the difference between Copilot and Microsoft 365 Copilot?

Copilot is the free, web-based assistant that anyone with a Microsoft account can use — it replaces the old Bing Chat. Microsoft 365 Copilot is a separate paid licence ($30/user/month, billed annually) that adds Copilot inside Word, Excel, Outlook, Teams, PowerPoint, and OneNote, with access to your tenant's emails, files, and chats via Microsoft Graph. The two share a brand and a base model but are very different products. If a tutorial says "ask Copilot about your quarterly report," they almost certainly mean Microsoft 365 Copilot, not the free version.

Is Copilot included in my existing Microsoft 365 subscription?

No. Microsoft 365 Copilot is an add-on that requires an underlying M365 licence at Business Standard, Business Premium, E3 or E5 level — and is then $30 extra per user per month, billed annually. The free Copilot at copilot.microsoft.com is included with any Microsoft account but does not have access to your work files. Copilot Pro at $20/month is for individuals on personal Microsoft 365 plans; it does not work with business tenants.

Can Copilot see my company's confidential documents?

Microsoft 365 Copilot inherits your existing permissions. It can see anything you can see — and only that. If a SharePoint folder is restricted to the legal team and you aren't on it, Copilot won't surface its contents to you. The risk is the inverse: documents you do have access to but shouldn't (over-shared sites, legacy permissions) suddenly become discoverable through chat. Most rollout pain happens here. Audit SharePoint sharing before deploying.

Does my data get used to train Microsoft's models?

No. Microsoft's enterprise data protection terms for Microsoft 365 Copilot guarantee that prompts, responses, and Microsoft Graph data are not used to train the underlying foundation models. Data stays in your tenant region and is not shared with OpenAI. This is one of the strongest reasons enterprises choose M365 Copilot over the consumer-grade Copilot or third-party AI tools — the data boundary is contractually clear.

Does Copilot work offline?

No. Every Copilot variant requires an internet connection because the model runs in Microsoft's cloud, not on your device. Some lightweight features (like predictive autocomplete in Outlook) work offline because they use local models, but anything labelled Copilot in the ribbon needs network access. For sensitive air-gapped environments, the only option is on-prem hosting via Azure AI, not Copilot directly.

What models power Copilot in 2026?

Microsoft 365 Copilot uses a mix of OpenAI models (GPT-4 Turbo and GPT-4o as of early 2026, with GPT-5 rolling into enterprise tenants through 2026) routed by an internal orchestrator that picks the right model for the task. Microsoft has also added options for some Anthropic and proprietary Microsoft (Phi family) models for specific workloads. Most users never see the model name; the orchestrator picks based on prompt complexity and latency targets.

How long does Copilot rollout actually take?

The technical rollout — assigning licences and enabling the service — takes a day. The actual rollout that drives adoption takes 8 to 12 weeks. The pattern that works: pick 50 power users across functions, train them properly, give them three months, then expand based on what they actually used. Companies that hand out 5,000 seats on day one and assume people will figure it out report 20 to 30% active usage. Companies that train deliberately report 60 to 70%.

The bottom line

Microsoft Copilot in 2026 is not one product, it is a family. The licence you actually want — Microsoft 365 Copilot — is the one that grounds in your tenant's data, sits inside Word/Excel/Outlook/Teams, and costs $30 per user per month on top of your existing M365 plan. The free Copilot is fine for casual chat; Copilot Pro is fine for individuals; everything else is a specialist variant.

If you're deciding whether to buy: don't license everyone. License 50 people who are ready to learn it, train them, measure their before/after time on routine tasks, and expand from there. The companies that won the first wave were the ones that treated Copilot like a productivity programme, not a software roll-out.

If you're using it: spend 30 minutes learning the moves that actually save time. Drafting from a brief in Word, formula generation in Excel, thread summarisation in Outlook, meeting recaps in Teams. Those four cover 80% of the value. Everything else — Copilot Studio, custom agents, the certifications — comes after the core habits stick.

The longer-term frame to keep in mind is that Copilot is the surface area through which Microsoft is delivering AI to most knowledge workers in the world. Whatever model wars happen at the lab level, the user-facing reality for the median office worker through 2026 and 2027 is going to be a Copilot button in their existing applications. That makes Copilot fluency a baseline professional skill rather than a specialist niche. The same way Excel literacy became table-stakes by 2005, Copilot literacy is becoming table-stakes by 2027. The investment in learning it well now pays a much longer dividend than it might appear at first glance.

For organisations evaluating where to invest first, the order that consistently delivers the best ROI is: Outlook (everyone), Teams (managers and project leaders), Excel (analysts and finance), Word (anyone who drafts long-form), then PowerPoint and Studio for the specialist use cases. Anchoring training and rollout sequence to this order — instead of treating all apps as equally important from day one — is what separates the deployments that show measurable productivity gains by month six from the ones that are still trying to justify their licence cost at month twelve.

Browse all our Microsoft Copilot guides for app-by-app deep dives, or jump straight to the cert prep in AI-900 in 30 days.

Last updated: January 2026