Microsoft 365 Copilot Cowork Usage Billing: Admin Spending Policies and Cost Controls

In a new feature, CNBC asks whether the 59-year-old CEO can repeat his cloud-era turnaround in the AI era. That's an interesting question for investors. For Windows and Microsoft 365 administrators it's a practical one, because the answer will show up on next year's software bill.

The scene: a CEO back in the spotlight​

CNBC's report opens in late September. About 50 executives from large Microsoft customers met Nadella at Seattle's Paccar Pavilion and could ask him anything. When one asked how Microsoft could help companies adapt to AI, Nadella talked about "diffusion inside the enterprise." In other words, AI only pays off if it's used widely across a company, not just by a few early adopters.

According to CNBC, Nadella has also been more visible in public AI debates lately. He posted on X about the need for "deliberate pacing" in model development and spoke about safety concerns at the All-In Summit in Los Angeles.

The criticism behind the story is blunt. CNBC says that despite a market value of nearly $4 trillion, Microsoft hasn't yet broken out as an AI leader:

• Chips: Nvidia dominates.
• Models: Anthropic and OpenAI have the leading ones.
• Consumer and business services: Google's Gemini-powered products are popular with both groups.
• Copilot: in CNBC's assessment, it isn't yet pervasive in the workplace, and Microsoft's own models remain niche.

Investors aren't excited either. CNBC reports the stock was up just 7% for the year as of Friday's close, behind its megacap peers and the Nasdaq. Microsoft hasn't beaten the S&P 500 in a calendar year since 2023.

Section summary: Microsoft is very large and makes a lot of money, but it hasn't yet convinced the market that it leads in AI.

Reinvention No. 1: the cloud playbook​

Nadella's track record is the reason anyone takes the comparison seriously. CNBC notes that he succeeded Steve Ballmer in 2014, after Microsoft's stock fell roughly 33% during Ballmer's 14 years in charge. Nadella then:

• moved customers away from perpetual Office licenses and toward subscriptions;
• built Azure into a major cloud business;
• let customers integrate products from rivals such as Oracle and Salesforce instead of insisting on an all-Microsoft setup;
• bought Mojang (the maker of Minecraft), LinkedIn and GitHub.

Microsoft's own fiscal 2014 annual report adds an important caveat. The cloud shift was already under way when he took over. That report already described Azure, Office 365 and Dynamics CRM Online as commercial cloud services. It also warned that heavy spending on datacenters "may decrease our operating margins." Nadella didn't invent Microsoft's cloud business. He made it the center of the company, which is arguably harder.

Reinvention No. 2: "per-seat-plus-consumption"​

The new pivot is stated in Microsoft's July 29, 2026 earnings call for the fourth quarter of fiscal 2026. Nadella said Microsoft is evolving "beyond per-seat to per-seat-plus-consumption." He said the change expands the company's total addressable market, meaning the total amount customers could potentially spend.

Jared Spataro, Microsoft's marketing chief for AI at work, explained the reasoning to CNBC. In January, Anthropic's Claude Code was gaining momentum, and Microsoft executives were impressed by how it combined many tools to get work done. Spataro said Microsoft realized it couldn't offer the best results within a fixed price per user per month. CNBC also cites a person familiar with the matter who said Microsoft was losing deals to Anthropic. Microsoft declined to comment on that.

Copilot executive Charles Lamanna made a similar point to Axios in June. As The Decoder reported, he said flat-rate pricing isn't sustainable because of "users who do hundreds of tasks a week," driving costs up quickly. The economics are simple: agentic AI, meaning tools that carry out multi-step tasks on their own, uses a lot of compute. A flat $30 seat doesn't cover a heavy user who runs dozens of long tasks every day.

Section summary: The model leaders forced Microsoft's hand. Copilot's fixed monthly price couldn't cover the compute cost of heavy agentic work.

GitHub Copilot was the test run​

GitHub is a useful case because it shows both Microsoft's early lead and the cost of falling behind. CNBC recounts that Microsoft and OpenAI released GitHub Copilot in 2021, a year before ChatGPT launched. Developers later moved to tools such as Cursor (now owned by SpaceX, according to CNBC) and Anthropic's Claude Code, and GitHub lost market share.

Then the pricing changed. Directions on Microsoft reports that as of June 1, Microsoft is charging GitHub Copilot users on a usage basis. On the July call, Microsoft said:

• GitHub Copilot has 50 million users.
• Copilot revenue grew more than 60% quarter over quarter.
• There was "significant consumption revenue" after the new pricing took effect.

CFO Amy Hood also said rising GitHub Copilot usage had hurt the Intelligent Cloud segment's gross margins, but that margins improved during the quarter after the June switch to usage-based pricing.

Read the 50 million figure carefully. Microsoft says "users," not paid seats. CNBC reports the figure was 26 million last October, but that comparison doesn't appear in the official material reviewed here. Kyle Daigle, GitHub's chief operating officer, told CNBC that the unit shared its lessons with other Microsoft teams considering usage-based pricing. In effect, GitHub was the pilot.

Copilot Cowork: where admins meet the new pricing​

The first Microsoft 365 product under Microsoft's usage-based billing (UBB) was Copilot Cowork. It's a cloud-hosted agent that takes on long-running tasks across several tools and returns a finished result. Microsoft made it generally available on June 16. On the July call, Nadella said "thousands of customers" were already paying for it and using it.

Microsoft's announcement and its admin documentation explain how the pricing works:

Directions on Microsoft confirms that the required license retails for $30 USD per user per month, with usage charges added on top. Futurum Group cites Microsoft's September 2026 Copilot Credits guide, which lists pay-as-you-go credits at $0.01 each, with pre-purchase discounts of up to 20%.

The approach has since grown beyond Cowork. At the September 25 event in Seattle, Microsoft introduced a unified Copilot app. TechPowerUp reported that the heavier agentic features, Cowork, Code, and Autopilot, [are] moving to usage-based billing that's off by default. Futurum notes that for enterprise customers, UBB services stay off until an admin creates a spending policy in the Microsoft 365 admin center.

Practical guidance for IT​

Microsoft Learn's Cowork governance page (last updated September 14, 2026) includes a detail that's easy to miss: a spending policy controls access, not just budget.

• Anyone included in a spending policy that selects Cowork can use Cowork. That's true even if the limit is a single credit.
• To keep someone out of Cowork, leave them out of every Cowork spending policy. A low limit doesn't block access.
• The old Frontier/Preview access control, set under Agents > All Agents, no longer does anything.

To move a previous allow list over, Microsoft's documented steps are:

1. Put the allowed users in a Microsoft Entra ID security group.
2. In the Microsoft 365 admin center, go to Copilot > Cost Management > Configuration and select + Add spending policy.
3. Under Scope user and group, choose Specific groups and select the security group.
4. Under Select agents and services, select Cowork.
5. Set limits, choose a billing method, and create the policy.

You've succeeded if users in the group can open Cowork and users outside every Cowork policy can't.

Microsoft also documents:

• spending limits at the tenant, group and user level;
• custom usage alerts;
• usage reporting broken down by user, group and feature.

Turning off the Anthropic model family changes which models users see. It doesn't take away Cowork access.

These controls address the main worry raised by David Berry, a vice president at Directions on Microsoft, in CNBC's story: once Cowork is switched on, employees could start "draining the bank" with no clear proof of benefit. Controls limit how much gets spent. They don't prove the spending was worthwhile. As Futurum puts it, the Copilot pricing model is still priced on effort, not results.

Section summary: Usage-based billing turns Copilot from a predictable per-seat cost into a metered one. Treat spending policies as access controls first.

Seats, adoption and the 450-million-user question​

Microsoft says it has more than 30 million paid Microsoft 365 Copilot seats, and that net new seats more than doubled from the previous quarter. CNBC compares that with a commercial Office 365 base of more than 450 million. That works out to roughly one in 15 commercial users paying for Copilot. Analyst Mary Jo Foley made a similar point earlier, noting via TechRadar that roughly 450 million commercial Microsoft 365 users have access to Copilot for free, meaning the paid base is comparatively tiny.

Investors are also unsure how pricing will settle. Tiffany Wade, a co-portfolio manager at Columbia Threadneedle, which owns Microsoft shares, told CNBC that Copilot needs to mature more before anyone knows the best solution or price. Cowen analyst Derrick Wood is more optimistic. He recommends buying the stock and sees a possible repeat of what happened when Claude Code moved to consumption pricing: customers' token bills rose sharply. Whether that's good news depends on whether you own Microsoft shares or pay its invoices.

OpenAI, Suleyman and the "every model" pitch​

The partnership that gave Microsoft its early AI lead has loosened. Microsoft's April 27 announcement confirmed four things:

• Microsoft remains OpenAI's primary cloud partner.
• OpenAI can now sell its products through any cloud provider.
• Microsoft keeps a license to OpenAI intellectual property through 2032, but it's now non-exclusive.
• Microsoft remains a major OpenAI shareholder.

The relationship has changed, but it hasn't ended.

Microsoft has also reorganized internally. On March 17, Microsoft announced that Jacob Andreou would lead the Copilot experience across consumer and commercial products. Mustafa Suleyman, CEO of Microsoft AI, would focus on Microsoft's "superintelligence" effort, meaning model development. CNBC reports that Suleyman's team has released models for coding, image generation, reasoning and transcription, but independent leaderboards don't yet rank them alongside Anthropic's or OpenAI's top models. Lamanna told CNBC that the Teams and Office group gets better transcription performance at a lower price from Suleyman's models.

Nadella's argument for why Microsoft doesn't need the top model is what he calls a "frontier ecosystem, not just a frontier model." At the September event, according to CNBC, he presented Copilot as connecting to "every model." The slide showed Anthropic, OpenAI, SpaceX and Microsoft AI logos. On the July call he put it in architectural terms: keep the "harness" separate from the model, so any model can be swapped in or out. In plain terms, Microsoft wants to own the platform that companies' AI agents run on, even if the models come from other companies.

The Windows angle​

PC users also have something to watch. CNBC reports that Microsoft is holding a Windows and Surface event in San Francisco this Wednesday. The invitation describes it as a conversation about how local AI, meaning AI that runs on the PC itself rather than in the cloud, "will shape the next chapter of the PC."

That sits awkwardly with the cloud-billing story. Local AI that runs on the device doesn't generate cloud usage charges. How Microsoft fits on-device AI alongside metered cloud agents could reveal more about its strategy than any earnings call.

The verdict, for now​

Berry told CNBC that Nadella "seems to have a plan, and he's working the plan." The evidence supports that much:

• Azure revenue topped $100 billion for the fiscal year.
• Microsoft reports more than 30 million paid Copilot seats.
• GitHub has a working usage-billing model.
• Cowork has detailed admin controls.

The evidence doesn't yet show that Copilot is everywhere, that Microsoft's own models can match the leaders, or that metered billing will lead to lasting value for customers rather than expensive experimentation.

The cloud transition succeeded partly because subscriptions made costs more predictable for customers. The AI transition asks customers to accept less predictable costs. For now, IT departments are in the middle of that experiment. Their best tools are spending policies, usage reports and clear success criteria agreed before Cowork is switched on for anyone.