The flat $30 seat is becoming an on-ramp, discounted by up to half for big buyers. Microsoft now needs metered agent work to make up the gap, and on the premium tier much of that revenue passes through to model suppliers. Shares rose about 3% on the launch. The margin story is weaker than that move suggests.
Microsoft launched a new Copilot on Friday and rewrote its billing at the same time. The product is a single app with three sections. Home runs chat and the Cowork agent, and full versions of Word, Excel and PowerPoint work inside it. Code lets non-programmers build dashboards and internal apps from a prompt. Autopilot is an always-on agent that runs recurring work without being asked. The shares rose about 3% in early trading, according to Reuters. Investors treated the launch as a distribution win. We see a concession: the $30 seat no longer carries the business case, and the replacement revenue will be harder to keep.
The billing change sits in the middle of Microsoft's own announcement. A per-user licence covers what Microsoft calls everyday AI: chat, Office inside Copilot, and an "Auto" router that chooses a model for each request by weighing accuracy, speed and cost. Cowork, Code, Autopilot and the frontier models are billed by usage. Jared Spataro, who runs marketing for AI at Work, described it as two kinds of AI at work with two ways to pay.
That split follows a sales instruction reported three days earlier. The Information said Microsoft had authorised discounts of about 30% for customers buying more than 1,000 seats and up to 50% above 10,000, starting as soon as October. The deepest discounts go to customers that also commit to usage fees. At list price a seat is $30 a month. Half off is $15.
The arithmetic is straightforward. Microsoft said at its July results that paid Microsoft 365 Copilot seats had passed 30 million, up from about 20 million three months earlier, and that net additions had more than doubled. At list, 30 million seats would bring in about $10.8 billion a year. GeekWire notes that this is roughly 7% of more than 450 million commercial seats, and that Microsoft does not break Copilot out of its $100.3 billion Microsoft 365 cloud revenue line. If the average large-account seat settles 40% below list, the same base yields about $6.5 billion. Seat growth could double from here and revenue per seat would still fall.
Microsoft is betting that the meter closes the gap. That explains Friday's rally. Azure worked because consumption pricing let Microsoft take the upside from its heaviest users rather than capping revenue at a licence fee. Agents that run all day are, in principle, the heaviest users software has ever had. An Autopilot that assembles supplier reviews each week or chases invoices through Outlook uses far more compute than a person asking for an email summary. If a few power users run agents constantly, metered revenue per customer could exceed what the seat ever earned.
The analogy breaks at the cost line. In Azure, Microsoft owns the data centres, the chips and the software it sells. On Copilot's premium tier, the models doing the expensive work belong to OpenAI and Anthropic; GeekWire identifies them as OpenAI's Astra and Anthropic's Claude Fable. Microsoft has IP rights to OpenAI's models, but inference still consumes GPU capacity that Azure could sell to paying tenants at a margin. For Anthropic's models, Microsoft is closer to a reseller. As an agentic task relies more on a frontier model, more of each metered dollar goes to the model maker. Metered revenue carries less margin than seat revenue.
Microsoft's launch material shows the pressure from customers. Alongside the new app it shipped "FinOps for AI": spending caps, policy controls and usage dashboards for administrators and staff. Vendors do not build budget brakes into a launch unless buyers have asked for them. Jacob Andreou, the executive vice-president for Copilot, told GeekWire that the autonomy that makes a long-running agent good "is the same thing that makes them absolutely terrifying to an IT admin." Chief financial officers will feel the same about an unbounded bill. During the first year of usage billing, the people least eager to raise the caps will control them.
Salesforce provides a precedent. It launched Agentforce in 2024 at $2 per conversation and has since added per-action credits and per-user plans because customers wanted predictable spend more than pay-as-you-go flexibility. Once a buyer can pick a flat plan, most will, and the vendor ends up discounting it. Microsoft is starting this process with a discounted seat already on the table.
The competitive logic is sound, which is why Microsoft is accepting the pricing trade-off. OpenAI launched its own super app in July. Anthropic's Cowork proved there was demand for an agent that works across files and inboxes, and Microsoft built its own Cowork in March partly from Anthropic's technology. Every month that knowledge workers spend in a rival's app is a month in which Word and Excel become interchangeable back ends. Putting Office inside Copilot keeps the user in Microsoft's app, which matters more to Microsoft than holding the price of one add-on. Satya Nadella's comment about using a private eval to swap one model for another points the same way: Microsoft wants to own the workplace interface and leave model makers competing to supply it.
The strategy keeps the franchise while lowering the price. Copilot is moving from a high-margin add-on to a distribution layer, with a discounted seat to secure the user and a meter to charge for heavy use. Part of the premium revenue will go to model suppliers. Seat counts will rise, while revenue per user and the gross margin on Copilot revenue lag them.
The rollout gives a timetable. Code moves to Microsoft's Frontier early-access programme and Autopilot enters private preview at the end of September, and the discounts take effect as soon as October. At the fiscal first-quarter results in late October, we expect Microsoft to report a seat count well above 30 million and again decline to disclose Copilot revenue. The next test comes at Ignite, from November 17 to 20. If the usage rates published there price a typical Autopilot deployment above the $30 list seat, Microsoft is using the meter to recover the discount. If they come in below it, the super app will grow Copilot's user base more than its profits.