Copilot deep dive — the AI already inside your Office, whether you noticed or not

It's four in the afternoon on a Thursday. The Teams meeting started late, ran long, and right when someone mentioned the client proposal your phone buzzed and you had to step out for fifteen minutes. When you come back, Copilot — a little button in the top right — has already written the summary: the three topics covered, who said what, and two action items assigned to you, with dates. You didn't ask. It just showed up.

That's Copilot for most people. Not a chatbot you go talk to. An assistant sitting inside the apps you use all day, offering help when it's useful.

So what is Copilot, exactly

"Copilot" is Microsoft's umbrella name for its AI. Under that umbrella there are several different products:

• Copilot web (copilot.microsoft.com): the free public chatbot, comparable to ChatGPT.
• Copilot in Windows 11: a sidebar for system tasks or general questions.
• Microsoft 365 Copilot: the one living inside Word, Excel, PowerPoint, Outlook, and Teams. The one writing your meeting summaries, drafting your emails, and building your slides.
• GitHub Copilot: the one auto-completing code in programming editors.

Four products, one name. For an office user, the one that matters is Microsoft 365 Copilot.

What it does inside Office

In Word, you ask "draft a first version of a proposal for X" or "rewrite this paragraph shorter" and it shows a suggested version next to yours for you to accept or discard. In Excel, you say "show me the ten rows with the biggest variation last quarter" or "add a chart for this trend" without touching a formula. In PowerPoint, you hand it a Word doc and it gives you a starting deck you refine from there. In Outlook, it drafts replies and summarizes long email threads. In Teams, it builds the meeting minutes and lets you ask questions about what was said.

Honest pricing

The web version is free. The one embedded in Office costs $30 per user per month, on top of your Microsoft 365 Business or Enterprise subscription. Not cheap. For a twenty-person shop that's $600 a month. Before buying, try a single license for a month and measure where it actually saves time — it's usually Outlook and Teams, less so PowerPoint.

Three things to take with you

1. Copilot isn't ChatGPT with Microsoft branding — it's an AI layer sitting inside the apps you already use.
2. The free web version is fine, but the value isn't there. The value is the integration with your mail, files, and calendar.
3. Before committing to M365 Copilot, give one license to a key user — someone who lives in Outlook and Teams — for 30 days. You'll know fast whether it's worth it.

The spreadsheet has 47,000 rows. Sales by region by product by month for the past three years. The financial analyst opens it in Excel, clicks the Copilot pane, and types: "Show me the ten biggest anomalies in Q3 2024 compared to Q3 2023, as a percentage, and add a column for the region."

Thirty seconds. A new table shows up on the side with the ten rows sorted by biggest variation, a chart attached. No formulas, no pivot table, no macros. Just a question in plain English and an answer.

What was a demo in 2023 and a preview in 2024 is the default Microsoft 365 Copilot experience in Excel in 2025-2026.

The strategy that won before anyone noticed

At Build 2024, Satya Nadella dropped the line that defines the strategy: "Copilot is the UI for AI." Don't race for the smartest chatbot. Race to be the interface people use to talk to AI. And the best way to be that interface is to live where people already work.

The numbers do the rest. Microsoft reported in its Q4 FY2024 earnings call over 400 million active Microsoft 365 seats. Selling Copilot doesn't mean convincing anyone to open a new app. It means convincing them to turn something on inside an app IT already pays for and Legal already approved. That's ten times less friction than any competitor faces.

Map of the Copilots that actually matter

The umbrella is wide. These are the ones most companies will touch:

• Microsoft 365 Copilot — the core: Word (draft, rewrite, summarize), Excel (natural-language analysis, chart generation, anomaly detection), PowerPoint (build decks from a Word doc, summarize), Outlook (draft replies, summarize threads, scheduling help), Teams (meeting summaries, in-meeting Q&A, automatic minutes).
• Copilot Pages — collaborative doc with multiple users plus AI, in the spirit of ChatGPT Canvas.
• GitHub Copilot — in-editor code completion plus a chat. Now multi-model (GPT, Claude, Gemini).
• Copilot Studio — low-code platform to build your own agents over corporate data (SharePoint, Dynamics, SQL). This is where the real future sits: not Microsoft's agents but agents each company builds on top of its own data.
• Copilot for Security / Sales / Service — vertical products. Security (alert triage), Sales (account summaries in Dynamics), Service (a co-pilot inside Dynamics 365 Customer Service).

All of them share the same base (Azure OpenAI mostly) but present as different interfaces per role.

The engine underneath

Copilot started tied to OpenAI — Microsoft invested more than $13 billion in OpenAI (figures reported across Bloomberg and Reuters coverage 2023-2025) and runs the models through Azure OpenAI Service. That's still the base.

But Microsoft diversified. It trains its own Phi family (small models optimized for efficiency). It signed deals with Anthropic and Mistral so Claude and other models show up inside Copilot Studio and GitHub Copilot. The message to the market: the engine isn't the product. The plumbing is.

Honest comparisons

Against ChatGPT Plus: standalone, ChatGPT wins on conversational UX, image generation (DALL-E), voice (Advanced Voice), the GPT Store. Copilot web loses that fight.

Against Claude: standalone, Claude wins on long coherent writing, PDF analysis, agentic coding. Copilot doesn't try to compete head-on — it redirects the user toward M365 Copilot inside Word/Excel.

Against Gemini in Workspace: this is the real fight. Google has Gmail, Docs, Sheets; Microsoft has Outlook, Word, Excel. The pick is about which platform your company lives on, not which AI is better. The two are roughly at parity on features, both below Claude or ChatGPT on raw conversation, both winning on deep integration.

What doesn't show up in the keynote

The UX across the different Copilots is uneven. Copilot in Excel is the most impressive when it works and the most frustrating when it misreads the question. PowerPoint Copilot generates slides that almost always need a rebuild. Outlook and Teams are the ones quietly delivering consistent value.

Pricing is also a thing. $30 per seat on top of an M365 subscription already priced at $22-$57 is a real buy decision. Forrester published a Microsoft-commissioned 2024 study showing ROI between 112% and 457% — numbers worth reading with the source caveat firmly in mind.

Open question to close

Where in your team's workflow is the friction of "opening another AI" the reason nobody actually uses it — and what would change if that AI were already inside Outlook? If you want more, start with Microsoft and Copilot — the AI inside your Office and then read The AI race.

On May 21, 2024, at Build, Satya Nadella spent the first ten minutes of the keynote explaining what sounded like a minor idea: "Copilot plus agents." It wasn't a product announcement. It was the public acknowledgment of a strategic pivot. For two years Microsoft had sold Copilot as "the chat inside Office." That day, officially, Copilot became something else: a platform for agents built on top of corporate data. The management lesson was written plainly: while the world watched who had the biggest model, Microsoft was building the layer those models would live inside.

The architecture that actually matters

There are four layers under every flavor of Copilot. Understanding them changes how you evaluate the strategy.

The first is Azure OpenAI Service, which runs the OpenAI models (GPT-4, GPT-4o, o1, o3, and whatever comes next) with tenant-level isolation. Not the same instance as ChatGPT — a dedicated copy per customer, with data residency guarantees and no training on customer prompts. Microsoft sold that story to compliance teams; OpenAI alone couldn't.

The second is Phi, Microsoft's own small-model family (3.8B to 14B parameters in Phi-3 and Phi-4). Optimized to run on-device or at the edge. Phi-4-multimodal shipped late 2024 with vision and audio. Not as capable as GPT-4, but cheap and local. Microsoft pushes Phi for scenarios where per-query cost matters.

The third is Microsoft Graph, the layer that supplies corporate context: mail, files, calendar, Teams, SharePoint, OneDrive. This is the structural advantage. No competitor has a graph this rich over Western office work.

The fourth is Copilot Studio, the environment where you build agents. Low-code but with technical hooks. An agent can read a SharePoint, talk to a SQL database over a connector, call an internal API, and live inside a Teams channel. The official line is "democratizing agent-building"; the real line is "get every company generating its own IP on top of data that lives inside Microsoft, and the switching cost climbs."

Multi-provider models: a strategic turn

For two years the story was simple: Microsoft = OpenAI. That shifted in 2024-2025. Microsoft signed with Anthropic to bring Claude into Copilot Studio and GitHub Copilot. It signed with Mistral for open models on Azure. It pushed its own Phi. And it started making clear at conferences that users pick the model inside the same product.

The diversification reads two ways. Defensive: reduce dependence on a single supplier, especially after OpenAI's governance volatility in November 2023. Strategic: deepen the moat around distribution rather than around the model itself. If Copilot runs on any model, the model becomes commodity and the integration layer becomes premium. It's the same move Microsoft ran twenty years ago with Windows on top of hardware from many manufacturers.

Vertical Copilots as enterprise Trojan horse

Copilot for Security (April 2024) is interesting less for the tech than for the distribution. Microsoft Defender has meaningful share in corporate endpoint security. Copilot for Security shows up inside the product IT already uses, triaging alerts and writing incident summaries. Priced separately, but the buying decision is made by a team that's already a customer.

Copilot for Sales lives inside Dynamics 365 CRM. It summarizes the account, preps the meeting, writes the follow-up. Competes against Salesforce Einstein. Less capable than Einstein on pure features, wins on pricing and Outlook integration.

Copilot for Service sits inside Dynamics 365 Customer Service. Assistant for human agents — reviews tickets, pulls knowledge-base articles, suggests replies.

The pattern is identical across the three: don't win with better AI, win with AI inside the product you're already in.

Technical and market comparison

Against Google Workspace + Gemini: the frontal fight. Gemini has stronger native multimodal capabilities (2M token context vs the 128K of GPT-4 Turbo that Copilot uses, more mature video/audio), but Workspace has less penetration in traditional enterprise than M365. Workspace made inroads in startups and small business in 2024; M365 still dominates the Fortune 2000.

Against Slack + Salesforce + ChatGPT Enterprise: a fragmented offer. More expensive in total. More flexibility but more friction. OpenAI added ChatGPT Enterprise to compete with Copilot web, but can't replicate the Outlook/Teams/Excel integration.

Against Claude Enterprise + ad-hoc integrations: Claude wins on conversational quality and on cases where the work is pure text (analysis, writing, code). Microsoft recommends (via Copilot Studio) integrating Claude when the use case calls for it. The sophisticated move: let Claude win specific cases inside Microsoft's wrapper.

Pricing is the real friction

$30 per seat per month on top of M365 Business Standard ($12.50) or M365 E3 ($36) is a cost doubling or more. For a 1,000-person company that's $360,000 a year. Forrester (a Microsoft-commissioned 2024 study) reports positive ROI in three categories: time saved on email drafting, accelerated reporting, meeting-time reduction. The 112%-457% three-year return figure comes from that study — and because it's a report funded by the selling company, those numbers should be read as a vendor claim, not independent evidence.

Academic work that's less tilted (Brynjolfsson et al. on a call center with an AI assistant, 2023) shows 14% average productivity gains and 34% for new hires. Not Copilot specifically, but a credible ceiling for what an integrated AI can do.

Editorial thesis: Microsoft doesn't need to win the model war

Microsoft's strategic bet fits in a single line: the model isn't the product; the integration layer is. If that thesis is right, Microsoft already won the enterprise fight by default — not because Copilot is the best AI, but because it's the AI sitting where work already happens. If the thesis is wrong and the future is standalone agents talking to each other without intermediation, then Microsoft paid $13 billion for OpenAI and built Copilot Studio to show up at the wrong party.

What we see so far suggests the thesis is at least partly right. AI is entering the products people already use faster than people discover new products. In that world, the differential is the Graph, not the model. Microsoft has the richest Graph in office work.

The main risk isn't technical. It's cultural. Copilot inherited Microsoft's historical sin: inconsistent UX across products. Copilot in Excel, Copilot in Word, Copilot in Teams, Copilot Studio, GitHub Copilot — each one with its own paradigm, shortcuts, and specific limitations. A user adopting M365 Copilot has to learn five different interfaces. Claude or ChatGPT, one paradigm.

The real question for the next three years isn't whether Microsoft will win the enterprise segment. It almost certainly will. The question is how long it takes to unify the experience — and whether, by the time it does, the agents built inside Copilot Studio will have generated enough lock-in that the conversation about "which model is best" stops mattering at all.