Microsoft's AI Bill Has Three Meters
Ask a Microsoft customer what they spend on Copilot and the answer is usually a seat count multiplied by $30. That is one meter of three. Microsoft sells AI through Microsoft 365 Copilot seats, through Copilot Credits consumed by agents, and through Microsoft Foundry, where models are billed per token or per hour of provisioned capacity. Each meter has its own rules, its own report and, in most organizations, its own owner.
This post takes them one at a time, then puts them in one table.
Meter one: Copilot seats
The seat is the meter everyone knows. Microsoft 365 Copilot is $30 per user per month, paid yearly, on top of a qualifying Microsoft 365 plan. For smaller organizations, Microsoft 365 Copilot Business is $18 per user per month on an annual commitment under a promotion that runs from 1 July to 31 December 2026 (list $21), or $25.20 on a monthly subscription. Copilot Chat is included at no extra cost with eligible Microsoft 365 plans (list prices, accessed 23 October 2026).
A seat is billed when it is assigned, not when it is used. The gap between the two is the oldest problem in license management, and Copilot makes it expensive: at $360 a year, a hundred idle seats cost $36,000. NHS Copilot Rollout: Count Assigned Before Demand showed how quickly an assignment can run ahead of demand, and When a Copilot SKU Changes Without a Buying Decision how the product being paid for can change underneath the seat.
Microsoft exposes Copilot seat usage through the Microsoft Graph API. The getMicrosoft365CopilotUsageUserDetail report returns, for each enabled user, the last date of any Copilot activity and the last date of activity in each app: Teams, Word, Excel, PowerPoint, Outlook, OneNote, Loop and Copilot Chat, over 7, 30, 90 or 180 days. It is a beta endpoint, and Microsoft points new integrations to the equivalent report under the Copilot API path. Either way, the data to compare assigned seats with active ones exists. What is usually missing is the habit of comparing them before the renewal.
Meter two: Copilot Credits
Copilot Credits pay for agents built in Copilot Studio. Microsoft's billing rates page, updated on 15 October 2026, sets the rates per event:
| Agent event | Copilot Credits |
|---|---|
| Classic (authored) answer | 1 |
| Generative answer | 2 |
| Agent action | 5 |
| Tenant graph grounding | 10 |
| Agent flow actions, per 100 | 13 |
| Premium (reasoning) AI tools, per 1,000 tokens | 10 |
A single question to an agent can trigger several events. Microsoft's own example is a grounded answer that costs 12 credits: 10 for tenant graph grounding and 2 for the generative answer.
Credits are bought three ways. A capacity pack is $200 a month for 25,000 credits, pooled across the tenant. Pay-as-you-go bills consumption through an Azure subscription with no up-front commitment. Prepaid Commit Units save up to 20% (list prices, accessed 23 October 2026). At the pack rate, 25,000 credits cover about 2,000 of those 12-credit grounded answers a month.
Three rules decide the size of this meter:
- The licensed-user exemption. Employee-facing agent use is included at no charge when the person using the agent holds a Microsoft 365 Copilot license and the agent runs under that person's identity, within fair-use limits. Computer-using agents are not included, and agent flows qualify only when triggered by the agent calling the flow. The same agent can therefore cost nothing for one employee and 12 credits a question for the colleague at the next desk without a Copilot seat.
- The 125% cut-off. On prepaid capacity, Microsoft disables custom agents when a tenant reaches 125% of its prepaid credits. Conversations in progress finish; new ones are refused until capacity is added or resets. Agent flows are blocked as soon as prepaid capacity is used up. Environments on pay-as-you-go are not cut off; they keep billing.
- Per-agent caps. Administrators can set monthly credit limits for individual agents in the Power Platform admin center, which is also where credit consumption is reported, by environment.
The interaction between the first meter and the second is the point. A decision to remove idle Copilot seats can move the same people's agent use onto Copilot Credits. Seats and credits have to be read together.
Meter three: Microsoft Foundry
Microsoft Foundry, the platform formerly known as Azure OpenAI, is where applications call models directly. Its deployment types set the meter:
| Deployment type | Billing |
|---|---|
| Standard | Per token |
| Priority processing | Per token, at a priority rate |
| Batch | Per token, at a discounted rate |
| Provisioned | Per provisioned throughput unit (PTU) per hour, or through an Azure reservation |
Provisioned throughput is the one that behaves least like software. Microsoft's documentation states that provisioned deployments are billed hourly for the PTUs deployed, regardless of the number of tokens consumed, from the moment a deployment is created until it is deleted. Azure reservations of one month or one year lower the effective hourly rate, but they do not guarantee capacity: holding quota or a reservation does not mean the region has PTUs available when you deploy. When a provisioned deployment is full, spillover can route the overflow to a standard deployment, which bills per token. Cached input tokens do not consume PTU capacity, so the cache hit rate changes how many PTUs a workload needs.
In practice a PTU estate has a fixed cost that runs whether or not anyone calls the model, a variable cost from spillover, and a commitment that can outlast the project it was bought for. Those are licensing questions as much as cloud ones.
Foundry is also a marketplace for models from other vendors. Claude in Microsoft Foundry is billed through the Azure Marketplace in Claude Consumption Units, converted from Anthropic's standard per-token rates, and US Data Zone deployments carry the same 1.1 times multiplier as US-only inference on Anthropic's own API, according to Anthropic's pricing page. A Claude workload can therefore appear on a Microsoft invoice, under a Marketplace line, rather than on an Anthropic one.
Who owns which meter
| Meter | Bought through | Reported in | Usual owner |
|---|---|---|---|
| Copilot seats | Microsoft 365 licensing agreement or subscription | Microsoft 365 admin center; Graph Copilot usage report | Microsoft 365 or end-user computing team, with SAM |
| Copilot Credits | Capacity packs, Commit Units or Azure pay-as-you-go | Power Platform admin center; Azure Cost Management for pay-as-you-go | Power Platform administrators, often the team that built the agent |
| Foundry tokens and PTUs | Azure subscription and reservations, Azure Marketplace | Azure Cost Management | Cloud platform or FinOps team, and the application owner |
Three owners, three consoles, one vendor. Nobody in this table is doing anything wrong. The problem is that a decision in one row changes the cost in another: removing seats raises credit consumption, a new agent built on a Foundry model moves spend from credits to tokens, a reservation bought for one project quietly funds another. From Copilot to Coworker: Licensing the New Digital Workforce describes where this is heading as agents take on work that used to need a named user.
If you are planning a Microsoft renewal, the base license still comes first: Microsoft 365 E3/E5 Optimization: Clean Up Before E7 Lands covers that cleanup. The same seat-plus-meter pattern now applies to GitHub Copilot, as explained in Copilot Seat Prices Held. The Credit Meter Did Not.
The question this leaves open
Each of the three meters has a report. None of the reports shows the other two. Where do the seats, the credits and the tokens sit side by side, so that a Microsoft AI renewal starts from one number instead of three?
To see your own AI spend per person, Software Inventory Manager (SIM) now connects to Anthropic Console and Enterprise, OpenAI and ChatGPT Enterprise, Google Gemini, Microsoft 365 Copilot, GitHub Copilot and Cursor: read how on the AI cost management page. Related reading: Four Vendors, Four Meters: AI Pricing Compared and Tokens, Explained for the People Who Pay the AI Bill.
