Agentmarketplace
All posts

Copilot Credits: Cost, Copilot Studio Pricing, and What Dynamics 365 AI Agents Actually Consume

Marcus Reyes, Editorial · Aug 1, 2026 · Last updated August 1, 2026 · 9 min read

What it does

Connects to
security-scanned
model-agnostic scoped access human-in-the-loop
Deployed · Live
Demo

Copilot Credits are the single currency Microsoft uses to bill every Copilot Studio agent and every first-party Dynamics 365 agent. They replaced Copilot Studio "messages" on September 1, 2025, with no change to the pack quantity or the pay-as-you-go rate. Credits are sold as a prepaid tenant pack, commonly quoted at $200 a month for 25,000 credits, which works out to $0.008 a credit, or metered through an Azure subscription at roughly $0.01 a credit. Microsoft publishes the consumption rates rather than the dollars: a classic answer costs 1 credit, a generative answer 2, an agent action 5, a tenant graph grounded response 10, 100 agent flow actions 13, a page of content processing 8, and a minute of voice 10, 35, or 75 depending on tier. Two rules decide your bill more than any of those rates. Employee-facing use by someone holding a Microsoft 365 Copilot licence is charged at zero, and unused credits never roll over, with Microsoft disabling custom agents outright once a tenant hits 125% of its prepaid capacity.

This guide converts every published rate into cents, works through the three consumption examples Microsoft documents itself, and explains the two traps that turn a $200 line item into a four-figure one. Rates below are from Microsoft Copilot Studio billing documentation as updated in July 2026.

Copilot Credits billing rates, converted to dollars

This is the table to keep. The credit column is what Microsoft publishes. The cents column is the same number at the $0.008 prepaid rate.

What the agent doesCopilot CreditsCost at $0.008 a credit
Classic answer (scripted response)1$0.008
Generative answer2$0.016
Agent action (trigger, deep reasoning, topic transition, computer use)5$0.040
Tenant graph grounded response10$0.080
Agent flow actions, per 10013$0.104
Content processing, per page8$0.064
Text and generative AI tools (basic), per 10 responses1$0.008
Text and generative AI tools (standard), per 10 responses15$0.120
Text and generative AI tools (premium), per 10 responses100$0.800
Classic voice, per minute10$0.080
Generative AI voice, per minute35$0.280
Premium generative AI voice, per minute75$0.600

The first thing that jumps out of that table is voice. A six minute call on premium generative voice consumes 450 credits, about $3.60. A typical text conversation of four scripted answers and two generative ones consumes 8 credits, about $0.064. Voice costs roughly 56 times a text conversation of comparable length. If someone in your organization is modelling a voice agent off text-agent economics, that is the number to show them.

The second is that agent flows are almost free. At 13 credits per 100 actions, a single flow action costs about a tenth of a cent, roughly forty times less than an agent action. Deterministic automation that does not require the model to reason at each step is the cheapest thing you can run in Copilot Studio, by a wide margin. Note that this applies to agent flows managed in Copilot Studio, not to Power Automate cloud flows, which run on Power Automate licensing and never touch Copilot Credits at all.

What are Copilot Credits?

A Copilot Credit is Microsoft unit of agent work. It measures the time and effort an agent spends retrieving information, responding to a prompt, or invoking an action, and the number consumed depends on the complexity of the task rather than on a fixed input size. Credits are pooled at tenant level, so every agent anyone in your organization builds draws on the same capacity, then allocated down to individual environments if you choose.

You get them three ways. A prepaid Copilot Credit pack subscription bought through the Microsoft 365 admin center is the standard route, and you need one before Microsoft will issue the otherwise free Copilot Studio user licence. Pay-as-you-go links an environment to an Azure subscription and bills actual monthly consumption with no commitment. A Copilot Credit prepurchase plan is a one-year prepaid pool of Copilot Credit Commit Units bought in the Azure portal and usable across eligible Microsoft products.

Copilot Credits vs tokens: what is the difference?

They measure different things and it matters. Tokens are units of text a language model reads and writes. Credits are units of billable agent activity, and most of the rate card is priced per event, not per token: an agent action is 5 credits whether the reasoning behind it was short or long.

There is one place the two meet. When an agent uses a reasoning-capable model, Microsoft bills on two meters at once: the normal feature rate for the operation, plus the premium text and generative AI tools meter at 10 credits per 1,000 tokens, about $0.08. So a generative answer produced by a reasoning model costs 2 credits for the answer plus 10 credits for every thousand tokens of reasoning. This is the one line on the rate card where a verbose agent design costs you directly, and it is the easiest place to overspend without noticing.

How much do Copilot Credits cost in USD?

Microsoft documents the credit rates and leaves the dollar conversion to its licensing guide, which is why this question is asked so often. The widely reported and consistently quoted figures are $200 a month for a 25,000 credit prepaid pack, giving $0.008 a credit, and about $0.01 a credit on the pay-as-you-go Azure meter. Microsoft own enforcement documentation uses a 25,000 credit tenant as its worked example, which lines up with that pack size.

Pay-as-you-go therefore carries roughly a 25% premium per credit. It is still frequently the better buy, for a reason covered further down: prepaid capacity can get your agents switched off, and metered capacity cannot.

How do I calculate Copilot Credit consumption?

Microsoft publishes three worked examples in its billing documentation. They are the most useful thing on the page because they show how differently the same rate card treats different agent shapes. Here they are with the monthly dollar cost added at the prepaid rate.

AgentPer runVolumeCredits per dayCost per month at $0.008
Customer support agent on a website4 classic answers + 2 generative answers = 8 credits900 customers a day7,200about $1,730
Tenant graph grounded sales agent in Copilot Chat4 generative answers + 4 grounded responses = 48 credits100 unlicensed users a day (50 licensed users billed at zero)4,800about $1,150
Autonomous order processing agent4 agent actions = 20 creditsper ordervaries$0.16 an order

The pattern is clear once you see the three side by side. Conversations cost money and back-office automation is nearly free. A support agent serving the public burns more than eight 25,000-credit packs a month. An autonomous agent that processes an order end to end costs sixteen cents a run, which means you could process ten thousand orders a month for less than the support agent costs in a week.

That asymmetry should shape which agent you deploy first. Internal, event-driven, back-office work is where the credit model is generous. Anything with a customer typing at the other end is where it is not. It is the same conclusion the honest deflection numbers point to, which we set out in the analysis of AI ticket deflection rates: the vendor-advertised 30% to 60% range does not survive independent measurement, and a support agent that costs $1,700 a month and deflects 22% is a different business case from the one in the deck.

One consumption line catches finance teams out. Content processing bills 8 credits, about six cents, per page. If your invoices, statements, or contracts arrive as scans rather than structured e-documents, every page runs through that meter before any agent logic happens at all, and a month of vendor invoices adds up quickly. Teams processing document volume at scale usually get better economics running the extraction step through dedicated document data extraction software and handing the agent structured fields, rather than paying a per-page rate inside the agent platform.

The Microsoft 365 Copilot rule that zeroes most of the bill

This is the most valuable line in the whole rate card and it sits in a footnote.

Every rate above is charged at zero when the person interacting with the agent holds a Microsoft 365 Copilot licence and the agent runs under their authenticated identity. Microsoft calls this business-to-employee usage and includes it in the Microsoft 365 Copilot user subscription licence: classic answers, generative answers, agent actions, tenant graph grounding, and agent flows all become no charge, subject to fair usage limits Microsoft reserves the right to change.

Look again at the sales agent example above. Fifty licensed users and a hundred unlicensed ones, and only the hundred generate a bill. The agent is identical. The licence on the other end is what decides the price.

Three exceptions are worth knowing before you plan around this:

  • Agent flows are only free on one trigger. The no-charge inclusion applies to flows started by "When an agent calls the flow". A flow on any other trigger consumes credits at the standard rate even for a licensed user.
  • Computer-using agents are never included. They bill at the 5-credit agent action rate regardless of licensing.
  • Customers are never licensed. Anyone outside your tenant is by definition an unlicensed user, so every public-facing agent bills in full.

If you already pay for Microsoft 365 Copilot, the internal agent tier you are entitled to is probably the most underused thing in your Microsoft estate. The full picture of what that licence costs and what changed on July 1, 2026 is in our breakdown of Microsoft 365 Copilot pricing.

What happens when you run out of Copilot Credits?

Your agents stop. Not gradually, and not with a warning to the person using them.

Microsoft allows a grace band above your purchased capacity, then enforces at 125% of prepaid capacity by disabling custom agents. A conversation already in progress finishes normally. After that, anyone who tries to use the agent gets one of two messages: "There is a billing issue" or "This agent is currently unavailable. It has reached its usage limit." An email goes to the tenant administrator and a notice appears in the Power Platform admin center, but the customer facing your agent sees only the failure.

Agent flows are cut off earlier and more narrowly. Once prepaid capacity is fully consumed, at 100% rather than 125%, new agent flow runs are blocked while the parent agent keeps working normally for classic answers, generative answers, and agent actions. Runs already in progress complete, test runs are exempt, and flow authors see a design-time warning.

The escape hatch is worth knowing: environments linked to a pay-as-you-go Azure meter are exempt from enforcement entirely, because overage bills to Azure instead of hitting a wall. That is why the 25% per-credit premium on pay-as-you-go is often the cheaper option in practice. You are buying insurance against your customer-facing agent going dark mid-month.

You can also cap spend deliberately rather than accidentally. In the Power Platform admin center, under Licensing then Copilot Studio then Manage Agents, you can set a monthly consumption limit per agent, which stops one runaway agent from taking the whole tenant to the enforcement threshold.

Do Copilot Credits roll over to the next month?

No. Copilot Studio enforces purchased capacity monthly and unused credits do not carry over. That makes sizing a prepaid pack to your busiest month an expensive habit, because you pay for the peak in all eleven other months as well. The common pattern that works is to hold a prepaid pack sized to your typical month and link the environment to pay-as-you-go for the spikes, so you get the cheaper rate on baseline volume and no enforcement risk on the surges.

Copilot Credits pay as you go vs prepaid: which to choose

Prepaid packPay-as-you-go
Rate per creditAbout $0.008About $0.01
CommitmentMonthly subscription, sized in advanceNone, billed on actual use via Azure
Unused capacityExpires monthly, no rolloverNothing to waste
What happens on overageAgents disabled at 125% of capacityExempt from enforcement, overage bills to Azure
Best forPredictable internal workloads you have measuredCustomer-facing agents, new deployments, spiky volume

The practical answer for most teams is both. Put your measured, steady, internal consumption on a prepaid pack at the cheaper rate. Put anything customer-facing or new on pay-as-you-go until you have a month of real data, because the cost of guessing wrong on a prepaid pack is not a bigger invoice, it is an agent that stops answering your customers.

Where this leaves a Dynamics 365 buyer

Two things follow from the rate card if you are evaluating Microsoft first-party agents.

First, the licence you already bought does not include them. Copilot features are increasingly bundled into the Dynamics 365 app licences, but agents draw on this separate, tenant-wide credit pool. Budget for it as a line item, not as something that came free with your seats.

Second, price the agent by who talks to it, not by what it does. A Dynamics 365 Finance reconciliation agent used by controllers who hold Microsoft 365 Copilot licences is close to free. The same platform running a public Contact Center agent is the most expensive thing on this list. We set out the full inventory of which agents exist in each app, which are generally available and which are still production ready preview, on the Dynamics 365 AI agents page.

And if the process you need automated is not on Microsoft list yet, which for Finance, Supply Chain, and Field Service is still the common case, a ready-made agent on a flat monthly subscription sidesteps the credit model altogether. Browse the catalog and deploy one against a single process from $29 a month, or compare the wider market first in our guide to AI agent pricing.

Find your agent

Browse vetted, ready-made AI agents, deploy one in a click on your own stack, and run it on any model. No lock-in.

Browse agents