Copilot Studio does not charge per maker, per agent, or per user. It charges per billable event, in a unit called the Copilot Credit, drawn from one pool shared by every agent in the tenant.
The model is workable once you know it. The difficulty is that nobody designs a conversation while holding a per-event price list in their head, so an agent gets built, gets shipped, and the first serious invoice arrives with nothing attached to it that explains the number.
Here is the whole picture, with the arithmetic.
What a credit costs
There are two ways to pay.
Prepaid packs. $200 buys 25,000 credits a month, which works out at $0.008 a credit. Packs are pooled across the tenant, so one pack covers every agent you run. Unused credits do not carry over. Buy a pack you only half use and you have paid $200 for $125 of consumption.
Pay-as-you-go. Attach a billing policy to an environment and credits bill to an Azure subscription at $0.01 each. That is 25% more per credit, with no commitment, no forecast, and no cliff.
Both are US list prices. The rate card below comes from Copilot Studio billing rates and management on Microsoft Learn. Read it against your own agreement before you commit to a forecast: this is the model that replaced per-message billing, and it is young enough to move again.
The rate card
- Classic answer — 1 credit. A response from a topic you authored by hand.
- Generative answer — 2 credits. A model-generated response over a knowledge source.
- Agent action — 5 credits. Triggers, topic transitions, connector calls, deep reasoning steps. Everything that lands on the activity map.
- Tenant graph grounding — 10 credits. Grounding an answer in your Microsoft 365 tenant graph.
- Agent flow actions — 13 credits per 100 actions. Per hundred, not per action. Power Automate cloud flows are not billed here at all — they stay on Power Automate licensing.
- Content processing — 8 credits per page. Per page, not per document. A 40-page contract is 320 credits on its own.
- Prompt and AI tools — 1, 15 or 100 credits per 10 responses, by tier. The premium tier is also what a reasoning model bills on top of its normal feature rate.
- Voice — 10, 35 or 75 credits per minute for classic, generative and premium generative voice. That per-minute rate already includes classic answers, generative answers and agent actions, so do not count those twice. Grounding, agent flows, AI tools and content processing are still billed on top.
Example one: a public support agent
A website agent for external customers. Four hundred conversations a day, seven days a week, so 12,000 a month. An average conversation runs four generative answers, and one in four escalates to an order lookup — an agent action plus a four-step agent flow.
- Generative answers: 12,000 x 4 x 2 = 96,000 credits
- Agent actions: 3,000 x 5 = 15,000 credits
- Agent flow actions: 3,000 x 4 = 12,000 actions, billed as 120 hundreds x 13 = 1,560 credits
Now the part that actually saves money. The top five questions are the same five questions every day. Author them as classic topics, and if that takes one and a half of the four answers away from the model on an average conversation, you save 18,000 credits a month. The total drops to 94,560, which fits inside four packs. Five authored topics, written once, are worth $200 a month for as long as the agent runs.
One thing this agent can never do is get cheaper through licensing. External users never qualify for the Microsoft 365 Copilot inclusion, because the person talking to the agent is not one of your licensed users. Run your own conversation shape through the credit estimator before you decide the model should answer everything.
Example two: an internal policy agent
An HR and policy agent for employees. A hundred conversations a working day, 21 working days, so 2,100 a month. Three generative answers per conversation, all grounded in the tenant graph so the agent can read the real policy library, plus one agent action to look something up in Dataverse.
- Grounded generative answers: 2,100 x 3 x 12 = 75,600 credits
- Agent actions: 2,100 x 5 = 10,500 credits
The grounding line alone is 63,000 of those 86,100 credits. Seventy-three percent of the bill, for one switch. Point the same agent at a specific SharePoint library as its knowledge source instead, with tenant graph grounding off, and the identical traffic costs 23,100 credits — a single pack, $200 a month.
The answers are not identical, and it would be dishonest to pretend otherwise. Grounding is what lets an agent reason across everything in Microsoft 365 that the user can already see. The question is whether this agent needs that. For a policy agent pointed at one document library, it usually does not, and the difference is $7,200 a year.
The two decisions that set the bill
Whether tenant graph grounding is on. It is a per-agent switch, it fires at 10 credits a time, and it is easy to leave on because it makes a demo better. Every agent should have to justify it in writing.
Whether the people using an internal agent hold Microsoft 365 Copilot licences. For an authenticated employee with a Microsoft 365 Copilot licence, classic answers, generative answers, agent actions, tenant graph grounding, agent flows on the agent-called trigger, and AI tools are all zero-rated, subject to fair use. Voice is not covered. Computer-using agents are not covered. Nobody outside your tenant is covered.
Do not buy Microsoft 365 Copilot at $30 a user to save credits. That arithmetic almost never works. Do use it in the forecast in the other direction: if 40% of the traffic in the example above comes from people who already hold the licence, the billable total drops to about 51,700 credits, and the right way to buy changes with it.
It also changes which agents are expensive. An agent for the finance team, who all hold Copilot licences already, is close to free. The same agent rolled out to frontline staff who do not hold them is billed in full. That is a rollout decision, not a licensing one, and it is normally made months before anyone runs the numbers.
Why the big pack is the wrong instinct
Unused prepaid credits expire at the end of the month. That single rule decides the buying strategy.
A pack costs $200 and holds 25,000 credits. The same credits on the meter cost $250. So a pack only saves money if you burn at least 20,000 of it — 80% — every month. Buy for your average month and you overpay in every below-average one.
Take the mixed example above, at 51,660 billable credits:
- Three packs: $600, with 23,340 credits expiring unused.
- Pay-as-you-go alone: $516.60.
- Two packs plus a billing policy: $400 for the first 50,000, then 1,660 credits at a cent each. $416.60.
So: packs to cover the floor you are confident of hitting every month, a billing policy behind them for everything above it, and a per-agent cap in the admin center so one runaway agent cannot drain the pool the others share.
Price the conversation before you design it
The difference between a $200-a-month agent and an $800-a-month agent is usually three decisions taken in the first week — whether grounding is on, whether the top questions are authored topics or model output, and who the agent is for. All three are cheap to change then and awkward to change later.
The credit estimator prices your own conversation shape against the published rates, prepaid and pay-as-you-go side by side.