Microsoft's Copilot Super App Meters the Agents

Microsoft did not ship another chatbot on September 25. It shipped the Copilot super app, a rewrite of the existing app and a new way to charge for the parts that act on their own. Fortune called it a super app aimed at business users. PYMNTS read the same launch as Microsoft leaving the personal chatbot race it walked into two years ago. Both descriptions fit. The product is one app. The bet is the enterprise seat, plus a meter.

We have been writing about agents that stall in pilot, including Collibra’s count of teams stuck there, and about labs gating models once agents showed up on a public hub in the Hugging Face security note. This launch is the vendor side of that mess: one interface, three modes, and a bill that splits “everyday” from “autonomous.”

What the Copilot super app actually contains

Fortune describes a major redesign of the existing Copilot app, meant to replace a sprawl of Copilot products that critics said left customers confused. The super app is the central interface. You chat with bots, use tools to finish tasks, write code, and reach Autopilot, the feature that can act on a task without a fresh prompt each time. A Microsoft executive in that piece said they now think of Copilot as a system that brings form and function together, not as one product or one form factor. That is a product-architecture sentence. It is also a warning that the name “Copilot” on a slide no longer tells you which surface you bought.

PYMNTS, citing GeekWire, organizes the new app around Home, Code, and Autopilot. Home combines chat with Cowork, the agent that takes a task and runs it to the end. Code lets employees build apps and automations by describing them. Autopilot lets people set up always-on agents that watch email and Teams channels and act without a new prompt. Full Word, Excel, and PowerPoint now run inside Copilot. In the demo PYMNTS describes, a product manager had Copilot pull data into an Excel workbook, accept tracked changes in a Word document, and turn both into a PowerPoint deck without leaving the app.

MediaPost puts the same three names on a Friday announcement and adds a calendar. The three tools roll out over the coming weeks. Autopilot expands to a private preview at the end of the month. Jared Spataro, Microsoft’s CMO of AI at Work, is the byline they cite. His example is the one a marketing team will recognize: ask Copilot for a launch plan in Word, then a budget in Excel, without switching apps. Home and Office, in that telling, can pull campaign data into Excel to look at return on investment, or generate a recap in PowerPoint. Code, for a marketer, means data pipelines, API integrations, or tracking scripts such as a custom Meta or Google pixel, without waiting on a dedicated engineering queue.

That last example is the one I would not paste into a launch email. A tracking pixel written by an agent inside a work tenant is still a tracking pixel. The launch does not say the script is reviewed. It says you no longer wait for engineering. Those are different claims. If your company already has a rule that marketing does not ship tags without review, Autopilot and Code do not repeal it. They make the violation faster.

MediaPost also mentions a secure digital vault inside Microsoft 365, so custom AI apps and tools can run with company data kept private, and a smarter way to track budget. Fortune had already reported the super app concept in May. September 25 is the day it stopped being a preview rumor and became a product with named modes.

Who pays, and who already isn’t

The pricing split is the part a CIO can use on Monday. Fortune says seats paid by an enterprise or by an individual cover “everyday AI”: Chat, Office integration, and standard routing. Usage-based billing can apply to Code, Autopilot, and access to the newest, most capable models. There is a free version of the super app. The full range needs a Microsoft subscription or a workplace license. Both enterprise customers and individual users pay extra, based on usage, for the most autonomous capabilities.

PYMNTS states the adoption number that makes the meter rational. Fewer than 7% of Microsoft’s 450 million commercial seats pay for Copilot today. The company is metering agent work by usage, on the bet that enterprise budgets will fund what consumer attention never did. I am not going to turn 7% of 450 million into a revenue forecast. The arithmetic people will do in the hallway is about 31 million paying seats, and that arithmetic is not in the article. What is in the article is the ratio, and the decision to stop pretending the consumer chatbot was going to carry this.

Folding consumer Copilot into a work product is the other half of that sentence. PYMNTS says Microsoft is walking away from the personal chatbot race it entered two years ago. The new Copilot puts the Office apps inside the work app and lets employees spin up always-on agents that live in the company’s Microsoft 365 tenant. Satya Nadella’s interest, as Fortune frames it, is subscriptions in a business that has become more important to Microsoft, and keeping the productivity suite from being sidelined by a new generation of AI rivals. The super app is how you stay in the document. The meter is how you get paid when the document is no longer the scarce thing.

If you are renewing a Copilot seat this quarter, ask which of Home, Code, and Autopilot the seat includes at the price you already pay, and which ones flip to usage the week Autopilot leaves private preview. Fortune’s split is the only price architecture I have. I do not have a per-action rate. Anyone who quotes you a dollar figure for an Autopilot run this weekend is ahead of the sources.

Office inside the app is the retention play

The demo is easy to shrug at. Excel, then Word, then PowerPoint, without leaving Copilot. People have been pasting between those three apps for decades. The change is that the paste is now a product surface Microsoft owns end to end, inside the app it wants you to open instead of a rival chat window.

That is why the “cedes the chatbot race” headline can be true and still not be a retreat. A personal chatbot that loses to ChatGPT or Gemini does not take Excel with it. A work app that contains Excel, and that can accept tracked changes and emit a deck, is a reason to keep the Microsoft 365 seat even if the employee prefers another model for weekend questions. Fortune says the app plugs into Office so a user can ask a bot to handle a spreadsheet task or change a Word layout. PYMNTS says the full versions of those apps now run inside Copilot, not a thin preview. I would want that distinction in a pilot. A thin preview that round-trips through the web apps is a demo. Full apps inside the shell are a migration.

The always-on piece is where pilots usually die, and where this launch is most specific. Autopilot watches email and Teams and acts without a fresh prompt. Home’s Cowork agent takes a task and runs it to the end. Those are two different autonomies. One is “I assigned this.” The other is “it noticed.” The second one is the one that will hit a retention policy, a records hold, or a manager who did not know an agent was replying in a channel. MediaPost’s private-preview date, end of the month, is the window to write that rule before the feature is a default someone enabled because the blog post sounded like a time saver.

We already know, from the Collibra piece, that a large share of agent projects never leave pilot because a human still has to check the work. Microsoft’s answer is not “the human goes away.” It is “the human’s check happens inside our tenant, and the autonomous slice is metered.” That can be a better procurement story than a model that lives in a browser tab. It is not evidence the check got cheaper. The 76% pilot figure in that earlier post was about human hours. This launch does not publish a new human-hour number. Do not borrow one.

Meta’s Muse is the comparison they invited

MediaPost sets the launch against Meta. At Meta Connect this week, Mark Zuckerberg laid out a vision for Muse that covers consumers and also enterprises, agencies, and brands. Some commerce platforms have started to integrate with Meta’s agent architecture. The gap MediaPost describes is operational: to launch a campaign with Muse, a marketer still has to bridge business planning and Meta Ad Manager by hand. Budget in one tool, strategy in another, then type it into Meta.

Microsoft’s counter, in that piece, is that Copilot can act across Word, Excel, and PowerPoint without that manual bridge. I would not score this as Microsoft winning ads. Meta still owns the ad system the campaign has to reach. A Word doc that becomes a budget does not become an approved campaign. What Microsoft can claim, on the evidence in these three pieces, is that the planning stack and the agent now share a tenant. What it cannot claim, because nobody published it, is a measured lift in campaign output or a lower error rate than Muse.

The competitive sentence that holds up is narrower. Microsoft is no longer trying to be the chatbot people open instead of a search box. It is trying to be the place work documents already live, with agents attached, billed in two layers. Meta is trying to put an agent in front of consumers and, this week, in front of brands. Those can both be true in the same September. A buyer who picks Copilot because a headline said it “challenges Meta” has picked a slogan. A buyer who picks it because Word and Excel are already the system of record, and the new meter is acceptable, has picked a stack.

What to ask before the preview ends

Autopilot’s private preview is the clock. MediaPost says it expands at the end of the month. The other two modes roll out over the coming weeks. That is enough time to ask five questions that the launch posts do not answer for you.

Which seats include Home at the price you pay now, and which actions inside Code or Autopilot generate a usage charge? Fortune drew the line at everyday AI versus Code, Autopilot, and the newest models. Your agreement may draw it somewhere else. Get the line in writing before someone enables Autopilot on a shared mailbox.

Where do the agents run, and what can they see? MediaPost’s vault claim is that custom AI apps can keep company data private inside Microsoft 365. “Private” in a launch blog is not a data-processing addendum. If you have a residency or a retention schedule, this is the week to match the vault language to the contract, not the week to discover the agent read a mailbox the schedule said to freeze.

Who reviews a pixel, a pipeline, or a reply the agent sent? Code’s marketing example is an integration built without engineering. That is a feature if engineering was the bottleneck and review still happens. It is an incident if review was the bottleneck and the agent skipped it. The Hugging Face note from earlier this month was about labs gating models because agents were already in the wild. The same instinct belongs inside a tenant. A model Microsoft hosts is not a model your change-control board has seen.

What does the 7% tell you about your own renewal? PYMNTS’s figure is Microsoft’s commercial base, not your department. If you are in the 7%, you are about to be asked to pay for autonomy on top of a seat you already defended. If you are in the other 93%, a sales motion will try to make the super app the reason to join. The product change does not answer whether your workflows survive Cowork running a task to the end. A two-week pilot on Home, with Code and Autopilot left off, will tell you more than the demo where a deck appears.

I would not budget the usage line at zero, and I would not budget it at a number these articles do not contain. I would name an owner for the meter, the same way you name an owner for cloud spend, before Autopilot leaves preview. The chatbot race PYMNTS says Microsoft ceded was a fight for a window. The super app is a fight for the document, with a second invoice attached. That second invoice is the part the September 25 posts actually changed.