Copilot's new bill starts where the agent stops waiting

On September 25, Microsoft rebuilt Copilot around three surfaces: Home, Code, and Autopilot. The announcement came from Jared Spataro, the company's chief marketing officer for AI at work. CEO Satya Nadella framed it bigger on X: "We're building Copilot as a new OS for work that spans every model, every form factor, and every task."

The product details are the least interesting part. The billing split is the story.

What happened

Home merges Chat and Cowork, with Word, Excel, and PowerPoint now embedded directly in the Copilot app. Code is a natural-language app builder for non-developers, running on the same underlying technology as GitHub Copilot, with apps hosted inside the customer's tenant through the new Copilot Managed Runtime (now in preview). Autopilot is the persistent cloud agent, previously called Scout: per Microsoft, you give it a name, a role, and a goal, and it keeps watching channels, following up on threads, and running recurring work "without waiting for a prompt." Each Autopilot gets its own identity, memory, computer, and workspace inside the tenant, with permissions, audit, and governance behind it.

Here is the billing line Microsoft drew. Everyday AI stays on the flat per-user subscription license (USL): Copilot in Chat and across Word, Excel, PowerPoint, Outlook, and Teams, covering "quick answers, first drafts, summaries and analysis." Everything agentic moves to usage-based billing (UBB): Cowork, Code, Autopilot, "new long-running agentic capabilities," and frontier models like Astra and Fable. Microsoft's justification, quoted verbatim in the announcement: "Agentic work requires a different type of spend."

Nothing here was hands-on tested; all billing mechanics are documentation-based research.

Why it matters

A seat was always a proxy for a human. A person reads email at human speed and drafts at human speed, and the seat captures that sloppily but adequately enough that nobody fights about it. That observation is FourWeekMBA's reading of the announcement, and it is the sharpest one in circulation: Autopilot, per Microsoft's own description, is "persistent, proactive and personal." It keeps working after you log off. Throughput is now the task queue and the model, not your attention. Once that is true, the seat is no longer a proxy for anything real.

Put numbers on it for the CFO. Microsoft-authored licensing material prices pay-as-you-go Copilot Credits at $0.01 each, against the standard Copilot seat at $30 per user per month billed annually. So one month of a seat buys exactly 3,000 credits. The arithmetic comes from The Weighted Average, and it deserves their caveat: this is a budget equivalence, not an allowance, a quota, or a prediction of what any user will burn. What it tells you is how quickly a heavy Autopilot deployment can cost you a seat a month per power user, and then keep going.

One mechanic genuinely favors buyers, with an asterisk. Per Unite.ai's reporting of Microsoft's pricing-model post, usage-based services stay off until an administrator creates a spending policy in the Microsoft 365 admin center, and nothing is billed before that point. Admins can set budgets at the tenant and group level, user-level caps, and alerts at spend thresholds, with Microsoft Graph API access for managing policies programmatically. From the same reporting: the per-user subscription is required to reach the metered capabilities at all, so UBB stacks on top of the seat rather than replacing it.

The asterisk is Auto. Microsoft's model router weighs "accuracy, speed and cost" per request and picks the model for the job. In other words, the system selects the consumption variable and you get the invoice. Meanwhile the announcement itself publishes no rates and no per-task costs for Code or Autopilot. The only consumption reference point in anything I read is Cowork's, reported by a Microsoft partner blog: light tasks at 100-300 credits, medium at 400-700, heavy at 700 and up, with consumption driven by model use, context retrieval, tool calls, and runtime. Whether Autopilot work burns at anything like those rates is unknown. Microsoft also says the subscription carries fair-use limits "set so the vast majority of users never reach one," with the heaviest users getting a warning and a choice: switch to Auto at no additional cost or move to credits. The shape is buyer-friendly. The numbers are still missing.

What to do before the preview expands

Autopilot expands to private preview at the end of September. Home and Code roll through the Frontier program over the coming weeks, and Code reaches Microsoft 365 Premium and Pro subscribers later this year. None of the three has a general-availability date. This is planning season, not purchasing season. Plan like this:

1. Create the spending policy before you grant the access. Nothing bills before a policy exists, which is the one generous default in this story. But once a policy exists, Microsoft's billing docs default new services to auto-apply under it, so audit existing policies rather than assuming Code and Autopilot stay uncovered. Keep spending gated until someone owns the budget line.
2. Budget two lines, not one. Seats are forecastable OpEx. Usage-based credits are variable and belong in their own line item with their own owner, the way cloud spend works. If your FY27 plan still treats Copilot as a single per-seat number, it is wrong.
3. Pilot on pay-as-you-go. Prepaid credit blocks don't roll over and the partner-blog-reported task tiers are the only consumption reference you have. Don't commit volume until you have measured yours.
4. Set tenant, group, and user caps plus threshold alerts on day one, and route credit requests through your existing approval workflows. Microsoft built those controls; using them is the point.
5. Decide who may create an Autopilot. Each one is a new identity in your tenant, per Microsoft. Treat creation like service-account provisioning, not a chat feature.
6. Demand task-level credit reporting per agent before anything touches production. If you cannot attribute the bill to the task and the agent that ran it, you have no handle on it.

Two things stay genuinely unknown: what a typical Autopilot week costs in credits, and whether seat-plus-meter saves or costs money versus the old all-in seat for heavy users. No rate is published for Code or Autopilot consumption, so no cost conclusion is possible in either direction. Microsoft Ignite, November 17-20 in San Francisco, is the next scheduled moment for those blanks to get filled.

The broader September context: coding-agent prices are falling across the board (see our price-war roundup), and Anthropic just cut its top-tier price (Opus 5.5). Microsoft is not cutting. It is re-anchoring: the seat covers the human, the meter covers everything the human no longer has to drive. Budget accordingly.