Microsoft renamed the billing unit from a message to a Copilot Credit in September 2025, and the meter underneath runs per capability rather than per chat. So a Copilot Studio answer draws a different amount depending on what the agent had to do to produce it, and a busy team burns a monthly allowance faster than the pricing page suggests.
At a glance: a Copilot Studio "message" is not one fixed unit, which is where surprise bills come from. A scripted answer costs about 1 credit, a generative one about 2, grounding it in your tenant's data adds roughly 10, a connector action about 5 more, and an autonomous event-triggered run 25 or more. They stack inside a single exchange, so one question can cost twenty times another. You can cap consumption per agent and get emailed before you hit the ceiling, in the Power Platform admin center. Credits were renamed from "messages" in September 2025. As at July 2026.
Quickly, on the pricing model itself, since a worked cost estimate belongs in its own post: Copilot Studio is licensed by usage rather than by seat, no per-user fee the way M365 Copilot has. You buy prepaid Copilot Credit packs or pay as you go through an Azure meter, and every agent interaction in your tenant draws from whichever pool you've set up. What eats those credits, and how you keep it under control, is the rest of this post.
What counts as a message, including the ones you didn't expect
Every reply a Copilot Studio agent sends draws at least one credit, and the amount depends on what the agent had to do to answer. A scripted topic, following a path you built in the authoring canvas, costs about 1 credit. A generative answer, composed by AI from a knowledge source rather than a script, costs about 2. Ground that answer in your organisation's own data through Microsoft Graph, SharePoint, Teams, Outlook, and it adds roughly 10 credits on top. Have the agent call a connector to look something up or complete an action elsewhere, and that's another 5 or so. Let it run autonomously, fired by an event rather than typed by a person, and one trigger can set off several actions in sequence, easily 25 credits or more before it's said anything back.
| Interaction | Approx. credits, as at July 2026 |
|---|---|
| Scripted (classic) topic | ~1 |
| Generative answer | ~2 |
| Generative answer grounded in tenant data | ~10-12 |
| Agent action calling a connector | +~5 each |
| Autonomous, event-triggered run | ~25+ |
The bit that catches people out is that these stack inside a single interaction. An answer that generates a response, grounds it in SharePoint, then calls a connector to raise a ticket costs the sum of everything it did along the way. Multiply that by however many times a day your team uses it, and the gap between what the pricing page implies and what the invoice says opens up fast.
Where capacity gets consumed fastest in a real deployment
In practice, three things drain a Copilot Credit pool faster than anyone budgets for. The first is a useful agent doing its job well. I've built ones like a support agent that cut troubleshooting time by about 80%, and once a team leans on something that heavily, pilot-week usage numbers stop meaning anything.
The second is grounding everything in tenant data by default. It's tempting to wire an agent to your whole SharePoint site so it can answer almost anything, but every one of those answers now costs roughly ten times a plain scripted one. Scope the grounding to the job in front of it.
The third is autonomous, event-triggered behaviour: an agent that fires off a new email landing, a form being submitted, or a schedule ticking over. Nobody's typing to rate-limit it, so it can burn through a chunk of the month's allowance overnight while everyone assumes the meter only moves when a chat window opens.
Pay as you go against prepaid capacity, and when each makes sense
As at July 2026, Copilot Studio offers two ways to pay for that consumption. Pay as you go runs through an Azure meter at roughly USD 0.01 per credit, no commitment, billed for whatever happened. A prepaid capacity pack buys a fixed block, commonly 25,000 credits for around USD 200 a month, roughly USD 0.008 per credit, about 20% cheaper. The catch: unused credits in a pack don't roll over, and you've paid for the block whether the month was busy or quiet.
Pay as you go suits a pilot, or an agent whose usage you can't predict yet, letting you see the real shape of consumption before committing. A capacity pack suits an agent that's settled into steady, predictable use. Once you know the number, size the pack with headroom above your average month rather than exactly on it. The month your busiest client goes quiet, or support tickets spike, is not the month you want to find out you've run out mid-week.
Setting a sensible cap before you find out the hard way
The Power Platform admin center lets you set a monthly consumption limit on an individual agent as well as the environment it lives in, and that limit applies whether you're on pay as you go or a capacity pack. Set it. An agent with no cap keeps answering, and keeps billing, well past the point anyone's actually watching.
Pair the cap with email alerts at a couple of thresholds, commonly 50% and 80%, sent to a group rather than one person, so someone reads it before month end. There's also a hard-stop option that switches the agent off at its limit, worth having on a low-stakes internal tool. Think harder before applying it to anything customer-facing: an agent that goes silent mid-shift because it hit an unreviewed number is its own kind of outage. Decide who gets paged before you turn the guardrail on.
Monitoring it without building a reporting project
You don't need a Power BI dashboard for this. Copilot Studio's own analytics already break billed consumption down by day and scenario, so you can see which agent, or which part of its job, is doing the spending. The admin center's capacity summary keeps roughly three months of daily consumption at the environment level, enough to spot a trend before it becomes a bill you weren't expecting.
A ten-minute look once a month, checking the trend line and the biggest driver that period, catches almost everything that matters for a business running one or two agents. Save custom reporting for the point where you're running enough agents across enough environments that a monthly glance stops being enough, and even then, export what the admin center already gives you.
If you're about to build one
If you're planning a Copilot Studio agent and want the consumption model thought through from the start rather than found on an invoice, that's part of what I help M365 firms get right on Power Platform and Copilot builds. Book a free 30-minute call and I'll give you a running cost alongside the build cost.