Copilot Studio licensing: who is talking decides the bill

The usual question about Copilot Studio licensing is what an agent costs to run. It is the wrong shape of question, and it produces business cases that are wrong by an order of magnitude in either direction.

The same agent, answering the same question, in the same way, costs either twelve credits or nothing at all. The variable is not the agent. It is whether the person talking to it holds a Microsoft 365 Copilot licence.

Microsoft’s own worked example makes the point better than any argument. A tenant graph grounded agent used by 50 Microsoft 365 Copilot licensed users and 100 unlicensed users is costed at 4,800 Copilot Credits per day. The arithmetic runs across 100 users, not 150. The licensed half of the audience does not appear in the calculation, because it does not appear on the bill.

That example is on the optimistic side. In the builds we are asked to price, the unlicensed share is not a third of the audience. It is most of it.

Copilot Credits, and what each thing costs

Copilot Credits are the common currency across Copilot Studio. They replaced *messages* on 1 September 2025, with no change to the quantity in a prepaid pack or the pay-as-you-go rate, so anyone working from a 2025 model is not out of date on price, only on vocabulary.

The rate card is short enough to hold in your head, which is unusual for Microsoft licensing and worth exploiting.

Voice is metered per minute rather than per interaction: 10 credits for classic voice, 35 for generative voice, and 75 for premium generative voice, with core agent activity included in that rate.

Note the spread. A classic answer and a premium AI tool response differ by a factor of a hundred. Two agents can be described identically in a project plan and differ by two orders of magnitude in running cost, entirely because of design decisions that nobody costed at the time.

The grounding line is the one that catches people. An agent grounded in your tenant graph spends 10 credits on the grounding and 2 on the generative answer, so a single complex prompt costs 12. Tenant graph grounding is optional and can be switched on or off per agent, which makes it one of the few genuinely large cost levers available after an agent is built.

The zero-rating rule, and its four conditions

“No charge” down an entire column looks like the whole answer. It is a conditional, and the conditions are where estates get caught.

The inclusion applies when the usage is employee facing, the user holds a Microsoft 365 Copilot licence, the agent operates under that authenticated user’s identity, and usage stays inside fair use limits that Microsoft reserves the right to change.

Three exclusions matter in practice:

Computer-Using Agents are not included. CUA usage is billed at the agent action rate regardless of the licence, at 5 credits.

Agent flows are only included on one trigger. The no-charge inclusion covers agent flows started by the “When an agent calls the flow” trigger. An agent flow on any other trigger consumes credits at the standard rate, even for a licensed user. A scheduled or event-triggered flow is not covered by anyone’s Copilot licence.

Generative answers are only free in one build path. Agents incur generative answer charges unless the agent was created in Agent Builder in Microsoft 365 and the response does not use tenant graph grounding.

There is also a distinction that saves real money once it is understood. Power Automate cloud flows use Power Automate licensing, not Copilot Credits. They are not billed as Copilot Studio usage and are not subject to Copilot Studio enforcement. Only agent flows managed in Copilot Studio are. Estates conflating the two either over-buy credits or panic about an enforcement risk that does not apply to them.

Most agents are built for people who do not hold the licence

This is the part that decides the business case, and it runs against the way agent projects are usually pitched.

In the vast majority of client builds, the intended audience does not hold Microsoft 365 Copilot licences. Organisations build agents for broad operational groups: frontline staff, HR and IT service desk requestors, external users. Buying Microsoft 365 Copilot Enterprise seats at $30 per user per month across that entire base was never budgeted, and was never going to be.

In our engagements, 70 to 80% of the target audience typically lacks a Microsoft 365 Copilot seat.

That creates an immediate architectural divide, and the two sides of it barely resemble each other:

That 12 to 17 figure is not a worst case. It is what a single multi-step turn costs once you combine a generative answer at 2 credits, an agent action at 5, and grounding at 10 or more. Microsoft’s own documented example of a grounded prompt is 12. Add one action and you are at 17.

So the failure is rarely a bad agent. It is treating agent interactions as free internal bot features, on the strength of a Microsoft 365 Copilot licence that most of the audience does not hold. That assumption is what turns a successful pilot into an unbudgeted Azure consumption bill.

The three ways to buy

Credits arrive by one of three routes, and the differences are operational rather than commercial.

Prepaid Copilot Credit pack subscriptions. Bought through the Microsoft 365 admin centre. Worth knowing that the Copilot Studio user licence is free of charge, but the tenant needs a prepaid pack subscription before that free user licence can be issued at all.

Pay-as-you-go. Billed to an Azure subscription through a billing policy created in the Power Platform admin centre, with environments linked to it. No commitment, no upfront purchase, and you pay for the credits actually consumed. Environments can be unlinked at any time, after which usage stops being billed.

The Copilot Credits prepurchase plan. A one-year prepaid pool of Copilot Credit Commit Units, purchased in the Azure portal, usable across eligible Microsoft products rather than locked to Copilot Studio.

Two access routes are not purchases at all. A Microsoft 365 Copilot licence lets you extend Microsoft 365 Copilot with agents built in Copilot Studio. And there is a trial licence with a restriction that catches technical evaluations: you can build agents and test them in the test chat panel, but you cannot publish. A proof of concept run entirely on trial licences proves the agent works and proves nothing about deployment.

Enforcement is not a warning, it is an off switch

This is the part that belongs in the risk section of any agent rollout, and it rarely is.

Capacity is enforced monthly and unused credits do not carry over. Microsoft permits a grace band above your purchased capacity, but enforcement triggers when a tenant reaches 125% of prepaid capacity, and custom agents are disabled. An in-flight conversation finishes. Every subsequent attempt to invoke the agent is rejected until capacity is increased or the month resets.

Your users see one of two messages: that there is a billing issue, or that the agent has reached its usage limit and is unavailable. Both are visible to whoever was relying on the agent, including customers if the agent is customer facing.

Agent flow enforcement behaves differently and the difference is favourable. When prepaid capacity is fully consumed, new agent flow runs are blocked, but the parent agent keeps working normally for classic answers, generative answers and agent actions. Runs already in progress complete. So flows degrade before the agent does.

Two structural protections are worth designing in from the start:

Allocate capacity at environment level. Microsoft’s own example is instructive. A tenant with 25,000 credits allocates 10,000 to environment A and leaves the rest pooled. If the pooled environments burn through 125% of their share, enforcement hits them, and environment A carries on untouched while it still has allocation left. Environment-level allocation is how you stop an experiment in a development environment taking down a production agent.

Pay-as-you-go environments are exempt. Once a pay-as-you-go environment passes its allocation, the meter takes over and enforcement does not apply. For anything genuinely business critical, that is a reasonable insurance premium.

There is also a per-agent cap available in the Power Platform admin centre, under Licensing, Copilot Studio, Manage Agents. It limits monthly consumption for an individual agent before tenant enforcement is ever reached, which is the control you want on any agent exposed to the public.

The trap in reasoning models

Reasoning-capable models bill twice, and the second meter is the expensive one.

An operation that uses a reasoning model is charged the feature rate for what it did, plus the premium text and generative AI tools rate for the reasoning tokens, at 10 credits per 1,000 tokens. A generative answer with a reasoning model is therefore 2 credits plus the token charge, not 2 credits.

This is defensible pricing, because deep reasoning genuinely costs more to serve. It is also completely invisible in a design review, where “use the better model” reads as a quality decision rather than a pricing one. Bring-your-own-model configurations, including Azure Foundry models, are billed separately again.

Our position: design the audience before the agent

In our view the first decision in any agent project is not what the agent does. It is who it is for, and specifically whether those people hold Microsoft 365 Copilot.

Two agents with identical scope can differ by their entire running cost on that question alone. An internal agent aimed at a licensed population is close to free to operate. The same agent pointed at contractors, frontline staff, or customers is a consumption line that grows with adoption, which is to say it gets more expensive precisely as it gets more successful.

We would not build the second kind without a per-agent consumption cap and an environment-level allocation in place first. In our engagements the failure is never that an agent was too expensive per interaction. It is that nobody modelled what happened when it worked.

Put more strongly: audience licensing must precede functional scoping. Designing an agent’s orchestration depth, its grounding sources and its tool calls before you have established the audience’s licence profile is building in the dark. An architecture that is cost-free for 100 Copilot-licensed knowledge workers will blow through tenant credit allocations if it is delivered to 1,000 unlicensed operational staff. Same design, same quality, entirely different outcome.

Sizing a rollout from a pilot

We do not recommend sizing from pilot consumption without adjusting it, because a pilot is run by a small, licensed, enthusiastic, internal population, which is the cheapest audience the agent will ever have.

Our benchmark for what happens next comes in two phases, and the first one alarms people who were not expecting it:

The curiosity spike is temporary and it is not waste. People probing the edges of a new agent are learning what it is for. But it lands in weeks one to four, which is exactly when a tenant is most likely to hit its 125% enforcement threshold, and an agent that gets disabled during its launch month rarely recovers its reputation internally.

The steady-state figure is the one that belongs in the business case. Note that it does not settle back to the pilot rate. It settles two and a half to three and a half times above it, because real use is conversational rather than transactional, and because automated triggers keep consuming after the human has stopped typing.

Where this connects to the rest of the Copilot estate

Three things land on the same conversation and it helps to see them together.

If you are weighing bundles, the Microsoft 365 E7 and Copilot licensing question determines how much of your agent usage is zero rated before you buy a single credit. That is the same decision as this one, viewed from the other end.

If you are wondering whether your estate is ready to get value from any of it, Copilot readiness is the prior question.

And AI Builder credits are a different currency entirely, on a different deadline. The names are close enough that estates confuse them. They are not interchangeable and one does not top up the other.

Where to start

Copilot Studio licensing becomes tractable once you answer four questions about the agent, and all four come before design.

Who is the audience, and what proportion hold Microsoft 365 Copilot. Does it need tenant graph grounding, or is that a default nobody chose. Does it use agent flows, and on which trigger. Does it use a reasoning model.

Those four answers produce a cost estimate accurate enough to make a decision with. Microsoft publishes an agent usage estimator that will do the arithmetic once you have them. Apply the rollout multipliers above to whatever it returns, because the estimator prices the interactions you describe, not the ones a curious new user will actually try in week two.

Veratas designs and licenses Copilot and AI deployments, including the capacity model underneath them.

If you are sizing an agent rollout and the numbers you are being shown vary wildly, talk to our team. The variance is usually one unanswered question about the audience.

Frequently asked questions

What are Copilot Credits? They are the common currency for Copilot Studio usage, replacing messages from 1 September 2025. The rename did not change pack quantities or the pay-as-you-go rate.

Is Copilot Studio free for Microsoft 365 Copilot users? Largely, for employee-facing scenarios where the agent runs under the licensed user’s authenticated identity, within fair use limits. Computer-Using Agents are excluded, and agent flows are only included on the “When an agent calls the flow” trigger.

What happens when we run out of Copilot Credits? Enforcement triggers at 125% of prepaid capacity and custom agents are disabled. Existing conversations finish, then further invocations are rejected until capacity is increased or the month resets. Agent flows are blocked earlier, at full consumption, without disabling the agent.

How do we stop one environment taking down another? Allocate capacity at environment level. An environment with its own allocation is unaffected by tenant-level overage until that allocation is spent. Pay-as-you-go environments are exempt from enforcement entirely.

Do Power Automate cloud flows consume Copilot Credits? No. Cloud flows use Power Automate licensing and are not subject to Copilot Studio billing or enforcement. Only agent flows managed in Copilot Studio consume credits.

How much does consumption grow from pilot to full rollout? Expect 5x to 8x the pilot’s per-user consumption during the first four weeks, as unfamiliar users explore the agent’s limits, then a steady state around 2.5x to 3.5x the baseline pilot query rate per active user. It does not fall back to pilot levels, because real use is multi-turn and automated triggers keep consuming after the conversation ends.

How many of our users will actually be covered by Microsoft 365 Copilot? Fewer than most business cases assume. Agents are usually built for frontline staff, service desk requestors and external users, and in our engagements 70 to 80% of that audience typically holds no Microsoft 365 Copilot seat, because seats at $30 per user per month were never budgeted across an operational base.

Do reasoning models cost more? Yes, and they bill on two meters. You pay the feature rate for the operation plus the premium AI tools rate for the reasoning tokens, at 10 credits per 1,000 tokens.

Can we run a proof of concept on trial licences? You can build and test agents, but you cannot publish them on a trial licence, so a trial proves the build and nothing about deployment.

Manish Kumar

Senior Business Intelligence Architect with 22 years of experience designing enterprise analytics and data platforms. Focus areas include Power BI, real-time analytics, and large-scale BI architecture across the Microsoft data stack.