One of the questions I hear quite often around Copilot Studio is this:
How much will this agent cost?
That sounds like it should be a simple question, but in practice it is not always that easy to answer. The reason is that Copilot Studio is not really priced like a traditional fixed per-user product. Instead, Microsoft measures usage in Copilot Credits, and the number of credits used depends on what the agent actually does. That includes things like classic answers, generative answers, agent actions, Graph grounding, agent flows, and AI tools.
So, if you are planning to build a Copilot Studio agent and would like to understand the cost before you go live, the best approach is usually not to look for one simple number. What you want instead is a reasonable way to estimate the likely usage and then turn that into an estimated cost.
Let’s look at how you can do that.
Why Copilot Studio cost can be difficult to estimate
The main reason this can be difficult is that the cost is tied to usage pattern, not just to the existence of the agent itself. Microsoft states that the number of Copilot Credits an agent consumes depends on the design of the agent, how often people interact with it, and which features are used during each interaction.
That means two agents can have the same number of users but very different costs.
For example, a simple internal FAQ agent that mostly returns predefined or lightweight responses can have a very different cost profile from an agent that uses generative answers, Microsoft Graph grounding, actions, and flows in the same run. Microsoft explicitly says that one interaction can consume multiple billing items at the same time.
So, rather than asking “what does one agent cost?”, it is usually better to ask:
- who will use the agent?
- where will they use it?
- what does the agent typically do in one interaction?
- is the usage internal, external, autonomous, or a mix?
Those questions get you much closer to a useful estimate.
First, understand the commercial model
Microsoft currently documents four main ways to pay for Copilot Studio usage:
- Prepaid Copilot Credit packs
- Pay-as-you-go (PAYG) through Azure
- A Copilot Credits Pre-Purchase Plan for annual prepaid usage
- Microsoft Agent Pre-Purchase Plan, Copilot Studio and Microsoft Foundry at tiered discounts
For PAYG, Microsoft’s Azure pricing page currently lists $0.01 per Copilot Credit. Microsoft also notes that prices shown on the Azure pricing page are estimates and that the actual amount can vary depending on agreement, currency, and purchase date.
Microsoft also states that a prepaid pack contains 25,000 Copilot Credits and that prepaid capacity is pooled across the tenant. Microsoft further states that unused prepaid credits do not roll over to the next month.
That is useful to know, because it means there is a big difference between:
- an agent with uncertain or low early adoption, where PAYG may make sense, and
- an agent with predictable monthly usage, where prepaid capacity may be easier to plan for.
What actually consumes Copilot Credits
Microsoft publishes a billing table for the main Copilot Studio capabilities. The current, as of march 15th 2026, published billing rates include:
- Classic answer = 1 Copilot Credit
- Generative answer = 2 Copilot Credits
- Agent action = 5 Copilot Credits
- Tenant Graph grounding for messages = 10 Copilot Credits
- Agent flow actions = 13 Copilot Credits per 100 actions
- Text and generative AI tools (basic) = 1 Copilot Credit per 10 responses
- Text and generative AI tools (standard) = 15 Copilot Credits per 10 responses
- Text and generative AI tools (premium) = 100 Copilot Credits per 10 responses
- Content processing tools = 8 Copilot Credits per page
The important point here is that these are not mutually exclusive. Microsoft states that a single interaction may use several feature types at the same time. Microsoft even gives the example that a single tenant-graph-grounded generative response can cost 12 Copilot Credits for one prompt: 10 for Graph grounding and 2 for the generative answer.
Billing rates and management – Microsoft Copilot Studio | Microsoft Learn
So, when you estimate cost, it helps to think in terms of what building blocks are used in an average interaction rather than just counting conversations.
Microsoft 365 Copilot licensing can change the picture a lot
This is one of the most important parts of the calculation.
Microsoft states that if a user has a Microsoft 365 Copilot license, Microsoft shows no charge for Microsoft 365 Copilot licensed users for classic answers, generative answers, agent actions, tenant Graph grounding, agent flow actions, and the listed AI tools in those employee-facing usage scenarios.
Microsoft also specifically states that when Microsoft 365 Copilot licensed users use agents in Copilot Chat, Teams, or SharePoint, then classic answers, generative answers, and Microsoft Graph grounding do not count against the Copilot Studio meter or prepaid pack.
That means one of the first questions you should ask when estimating cost is this:
Are the intended users already licensed for Microsoft 365 Copilot, and are they using the agent in a Microsoft 365 employee-facing scenario?
A practical way to estimate the cost
So, how do we turn this into something useful?
This is the model I would use.
1. Define the agent scenario
Start by deciding what kind of agent you are building:
- internal employee-facing
- external/customer-facing
- autonomous
- or mixed
That matters because these scenarios can be billed very differently.
2. Estimate the expected usage volume
I thinks this step often is the most complicated. You can often figure out the cost of one interaction with your agent (depends on the complexity of the agent of course), but can you figure out how often your users will interact with the agent? This is the part were we will do some qualified guess work 🙂
- number of users
- number of interactions per user per day or month
- total sessions or requests per month
- expected adoption growth after launch
It is often a good idea to create three scenarios:
- low
- expected
- high
That gives you a more useful forecast than trying to guess one exact number from day one.
3. Estimate the average credit mix per interaction
For a typical interaction, ask:
- will the agent give a classic answer or a generative answer?
- will it use tenant Graph grounding?
- will it trigger one or more agent actions?
- will it run agent flow actions?
- will it use AI tools, premium reasoning, or content processing?
That gives you a much better idea of how many credits one average interaction may consume.
4. Split the usage by user type
Now split your usage into groups such as:
- Microsoft 365 Copilot licensed internal users
- unlicensed internal users
- external users
- autonomous runs or triggers
This is important because the same interaction pattern can have a very different commercial outcome depending on who is using it and where.
So when you have gathered the above information you can use the Microsoft agent usage estimator to calculate the cost of your agent.
A couple of useful examples
Microsoft includes a few billing examples in the documentation that are useful as reference points.
Microsoft describes a customer support agent scenario where the estimated cost per day is based on a combination of classic and generative answers, and Microsoft’s example lands at 7,200 Copilot Credits per day.
Microsoft also describes a sales performance agent using tenant Graph grounding for unlicensed users, where the example comes to 4,800 Copilot Credits per day.
And Microsoft includes an order processing agent example for an autonomously triggered internal-facing agent, where the example run comes to 20 Copilot Credits for the set of actions performed.
These are of course just examples, but they are helpful because they show how quickly the numbers can grow once an agent starts using multiple billable features in the same run.
Conclusion
Calculating the cost of an agent has two parts:
- Calculate the cost for each interaction with the agent, for a simple agent with only one or few actions this is easy. For a more complex agent with a lot of different actions, prompts and connected Agent flows or Child/connected agents this becomes more difficult.
- Estimate how people will actually use the agent. This is usually the hardest part to get right before the agent is live. We can count users, but predicting how often they’ll interact with the agent is mostly educated guesswork. At this stage, that’s fine. One approach is to work with three scenarios: low usage, expected usage, and high usage. The important thing to remember is that these numbers are not final. Once the agent is published and starts being used, real data replaces assumptions. At that point, you can measure actual usage, adjust your estimates, and finally understand what the agent really costs in practice.
- When you have your estimated numbers, run them through the Microsoft agent usage estimator to get a forecast of how much Copilot credits your agent will consume