The pilot looked free. A handful of licensed users asked an agent questions in Teams, nobody saw a bill, and the project got a green light to scale. Three months later, finance is asking why there is a Copilot Credit line that wasn’t in the budget. This is the most common AI cost conversation in government right now, and it is entirely avoidable.
Why the Pilot Felt Free
On September 1, 2025, Microsoft switched the agent billing unit from messages to Copilot Credits. One credit is one cent. If you hold a Microsoft 365 Copilot license and a licensed user talks to an internal agent in Teams, SharePoint, or Copilot Chat under their own identity, that usage is zero-rated; it does not draw down credits. That is the part everyone remembers, and it is exactly why pilots feel free.
The meter starts the moment you leave that lane. An agent serving an unlicensed user, an agent published to an external channel, or an agent that runs on its own without a person triggering it: all of those consume credits. The question that actually drives your bill is not which feature you turned on; it is who is using the agent, and under what identity.
Where the Meter Actually Runs
The rates are public, and they are not flat. A classic, scripted answer costs 1 credit. A generative answer costs 2. Grounding a response in your tenant-wide Microsoft Graph adds 10 credits on top, so Microsoft’s own example of a graph-grounded agent lands at 12 credits per response. Content processing runs 8 credits per page. The one that ends careers in Q4 is the autonomous trigger: 25 credits every time an agent fires on its own, always billed, with no exceptions and no inclusion in your Copilot license.
Interactive agents scale with users, so you can forecast them. Autonomous agents scale with events, and events spike. An agent that watches a mailbox or a records queue and acts on every item will generate cost with no human in the loop to notice. Stack a few of those, ground each response in the Graph, and a workflow that felt like a rounding error becomes a real number fast.
The question is not which feature you enabled. It is who is using the agent, and whether a human triggered it.
The Estimator Is Not Optional Anymore
Microsoft’s agent usage estimator forecasts monthly credit consumption before you deploy, and the April 2026 update extended it to cover Dynamics 365 agents like Sales Qualification and Customer Service alongside Copilot Studio agents. You model low, medium, and high volume, then export a PDF for procurement sign-off. Run it before you scale, not after the bill arrives. Microsoft’s own guidance is to buy 10 to 20 percent above your estimate as a buffer, because credit packs do not roll over; unused capacity expires on the first of every month, and if you blow past your purchased capacity, enforcement can deny service mid-workflow.
There are also costs the estimator does not show you. Bring your own Azure OpenAI model and those tokens bill separately on the Azure meter. Ground agents in documents through SharePoint and you may need SharePoint Premium. After November 2026, AI Builder capabilities fold into Copilot Credits too. None of this appears in the credit conversation until someone asks why the Azure invoice grew.
GCC Changes the Math and the Timeline
The Researcher and Analyst agents reached the Government Community Cloud (GCC) on April 2, 2026, so government shops can now enable exactly the kind of multi-step, autonomous reasoning that consumes credits in volume. That is the trap waiting for a lot of agencies: a deep-reasoning agent gets switched on without a forecast, and the meter does the rest. Government pricing is also a moving target; Microsoft has signaled pricing updates for Microsoft 365 Government in summer 2026, so do not anchor your budget to commercial per-user numbers. Confirm the gov SKU and model it against your own traffic.
The work that survives this math is the work that pays for itself. The license-reclamation and records-classification patterns I have engineered in production GCC environments recover real labor and real license spend, which is what justifies the credits you do spend. Cost control here is not a spreadsheet exercise; it is an architecture decision made before the first agent ships.
Who’s Writing This
I am a Navy veteran and an M365 and AI engineer who builds inside GCC constraints, and Puget Sound AI is a veteran-owned small business (VOSB; SBA VetCert in progress). No account managers, no license resale, no upsell. You talk to the person who does the forecasting and the building.
If you are scaling agents and want the credit math done before Q4 instead of during it, let’s talk.