There has been no such thing as a Copilot Studio message since 1 September 2025. The unit is the Copilot Credit, and half the material still circulating on agent costs counts messages. Some of it was written this year.
This looks like pedantry and is not. A message implied one flat thing that happened once. Credits are priced by what the agent actually did, and the published rates vary by an order of magnitude across the things a single question might trigger. A classic answer costs one credit. A generative answer costs two. An agent action costs five. Grounding a question in your own tenant’s Graph, which is the entire reason most people want the agent in the first place, costs ten. So a model that assumes a fixed number of interactions per user per month tells you almost nothing, because the expensive interactions are the ones you are building for.
The consequence of getting it wrong is not a slightly larger invoice. Prepaid capacity comes in packs of 25,000 credits a month, and overage enforcement fires at 125% of your prepaid capacity and disables custom agents. Agent flows get their new runs blocked instead. Underestimate badly enough and the thing you spent six months adopting stops on a Tuesday.
There is a large free allowance and people miss it in the other direction. Employee-facing agent usage carries no charge for Microsoft 365 Copilot licensed users, subject to fair use, when the agent runs under the licensed user’s identity. Two documented exceptions: Computer-Using Agents are not included in that licence, and agent flows are only free through the “When an agent calls the flow” trigger. Both exceptions are the sort of thing that turns up in month four.
Change the vocabulary and the questions in the room change with it. If nobody can tell you how many credits your agent burns answering one real question, you do not have a business case. You have a demo with a spreadsheet attached.