Copilot Can Now Act Inside Your Documents by Default. Here Is How to Stay in Charge of It.

For a long time, using AI at work was a choice you made. You opened a chatbot. You decided to try it. Now the choice is changing shape. You will open a document you already keep, and Copilot will be ready to act on it directly, not just suggest from the side. That is a different starting line, and it is worth understanding before it shows up on your screen.

What Microsoft actually announced

Build 2026 ran June 2 and 3 at the Fort Mason Center in San Francisco, with an online track alongside it. Two threads from that week are worth your attention as you decide where to point it.

The first thread is the one most coverage led with. Microsoft AI's Superintelligence Team, formed in November 2025 and headed by Microsoft AI CEO Mustafa Suleyman, used Build as its first big public showing. The team announced a family of seven in-house models it calls MAI. The lineup is led by MAI-Thinking-1, Microsoft's first reasoning model, and includes image, transcription, voice, and coding models. The coding model, MAI-Code-1-Flash, is tuned for VS Code and GitHub Copilot CLI. MAI-Transcribe-1.5 handles 43 languages. MAI-Thinking-1 is a Mixture-of-Experts model with 35 billion active parameters and a 256K context window, trained from scratch with zero distillation.

Part of the reason is economics. Owning more of the model stack can lower serving costs, reduce reliance on outside model providers, and give Microsoft more control over pricing and product integration. CNBC covered the launch in those terms: a move to lessen reliance on OpenAI and lower costs.

The second thread is the one that reaches your desk. Copilot's agentic capabilities in Word, Excel, and PowerPoint reached general availability on April 22, 2026. Microsoft says these capabilities are now generally available and the default experience for customers with Microsoft 365 Copilot and Microsoft 365 Premium subscriptions, and they are also available to Microsoft 365 Personal and Family users. The release describes AI that can take multi-step actions directly inside your documents, worksheets, and presentations while you stay in control, rather than handing you a suggestion to copy in yourself.

A note on the GPT-5.5 part

You may see this story told as if Microsoft released GPT-5.5 at Build. That is not what happened. GPT-5.5 is OpenAI's model. It went out to paid ChatGPT tiers on April 23, 2026, and became the default for all ChatGPT users, including the free tier, on May 5, 2026. At Build, GPT-5.5 appears only as the benchmark Microsoft measured its own models against. Keeping that straight is its own small skill, because how the story is told changes what you think you are buying.

The benchmark claim, and how to read it

Here is the line Microsoft used about its new models. According to the keynote transcript, Suleyman said: "When we tuned our models for McKinsey's tasks, MAI delivered the highest win rate, outperforming GPT-5.5 on quality, whilst being 10x lower on cost."

Read that slowly. The "10x lower cost" figure is not a measured production result. By Microsoft's own description, it is a projection built from public GPT pricing compared against MAI pricing, scaled by the difference in serving costs. Microsoft also reports that in its own blind evaluations, independent raters preferred MAI-Thinking-1 to Claude Sonnet 4.6, and that the model matches Claude Opus 4.6 on coding on the SWE-Bench Pro test. Those are Microsoft's own internal results, not numbers an outside auditor checked.

None of that makes the claims wrong. It makes them vendor claims. And learning to tell the difference between a measured result and a projected one is one of the most useful habits a professional can build right now. The question to keep in your pocket is short. Measured in real work, or projected from a price sheet?

Microsoft also cited its own engagement numbers for the agentic Copilot release: Word usage up 52 percent, Excel retention up 50 percent, and PowerPoint satisfaction up 25 percent, month over month. Useful to know. Also worth reading the same way, since those are the company's own figures about its own product.

Why this reaches you, even if you never asked for it

The practical change is the default. When a capability is opt-in, the people who use it are the ones who went looking. When it is the default, it shows up for everyone, including the professional who has been keeping a polite distance from all of this.

So the useful question is no longer whether to try AI. It becomes whether you know how to direct an agent that is already working inside your file, check what it did, and correct it when it is wrong.

That is a real skill, and it sits at a specific place in The 7 Levels of AI Proficiency. Knowing a tool exists is early-stage awareness. Telling whether its output is sound is Level 3, the Critical Thinker, the person who can look at what the AI produced and judge it. Setting up the work so the AI does the right thing in the first place is Level 4, the Commander, the person who gives an agent the context and the guardrails to act well. The default-Copilot moment asks for both at once.

You are no longer reviewing a suggestion. You are managing a worker that already moved.

Two reps you can run this week

You do not need a course to start building that skill. You need a few small, checkable reps inside work you already understand.

• The next time Copilot acts inside one of your documents, stop before you accept it. Ask it to explain, in plain language, what it changed and why. Then verify one piece yourself. You are practicing the Critical Thinker habit on a file where you already know what a right answer looks like.
• The next time you read a vendor benchmark, including this one, write down whether the headline number was measured in real use or projected from pricing. You will be surprised how often the answer is "projected," and how quickly the habit tightens your read on every AI pitch that crosses your desk.

The cost story is a leadership story

There is one more read here, aimed at anyone who signs off on software spend. Microsoft building its own models to undercut OpenAI's pricing tells you something about where this is heading. Model choice is becoming a budget decision, not only a capability decision. "Which model, at what cost, for which task" is going to be a recurring line in the conversation, the same way cloud spend became one.

You do not have to pick a side in the Microsoft-versus-OpenAI story to take the point. The point is that the price of the same kind of work is moving, and the leaders who track it will plan better than the ones who treat AI cost as a fixed number.

A next step

Open one document you keep current. Word, Excel, or a deck, whichever you use most. The next time Copilot offers to act on it, let it, then make it explain the change before you accept anything. That single habit, done on a file you already know, is how the move from watching AI to directing it actually gets built.

Related reading: Level 4: The Commander.

Sources

1. Microsoft Build 2026
2. Microsoft AI: Build 2026 MAI keynote transcript
3. CNBC: Microsoft unveils new AI models to lessen reliance on OpenAI, lower costs
4. Microsoft 365 Blog: Copilot's agentic capabilities in Word, Excel, and PowerPoint are generally available
5. Microsoft Blog: Build 2026, be yourself at work
6. OpenAI: Introducing GPT-5.5
7. OpenAI: GPT-5.5 Instant

Harrison Painter

Executive AI Advisor. Founder, LaunchReady.ai and AI Law Tracker.

Harrison is an Indiana AI Advisor who helps business owners and executives get their time back by building AI systems that run the work for them. Nearly 20 years in business and author of You Have Already Been Replaced by AI. Creator of The 7 Levels of AI Proficiency.

Connect on LinkedIn