Copilot Studio Credits: How the Meter Actually Works
Audience: M365 / Power Platform admins and architects who already run Copilot Studio agents in production and need to explain, forecast, or defend a credit bill.
- A Copilot Credit is not "one message". One user turn can hit several meters at once, and the expensive ones (tenant graph grounding, premium AI tools, reasoning models) stack on top of the base answer rate.
- Microsoft 365 Copilot licensed users are zero-rated for most agent features, but only under specific conditions. Their usage still appears in Power Platform admin center (PPAC) as consumed credits, tagged nonbillable. A large credit figure is not automatically a large bill.
- Capacity is pooled at the tenant and optionally allocated per environment. Overage handling and enforcement are decided per environment, and pay-as-you-go (PAYG) is attached to environments through a billing policy, not to the tenant.
- Enforcement is not one rule. Custom agents are disabled at 125% of prepaid capacity; agent flows are blocked at 100%; PAYG-linked environments are exempt from both.
1. The unit of measure
| Fact | Detail |
|---|---|
| Currency | Copilot Credits (renamed from messages on 1 September 2025; pack size and PAYG rate unchanged) |
| Prepaid pack | 25,000 credits per month, pooled across the tenant, must be assigned to an environment to light up features there |
| PAYG meter | USD 0.01 per credit, billed to an Azure subscription via a PPAC billing policy |
| Carry-over | None. Unused prepaid credits expire monthly |
| Prepurchase plan | One-year Copilot Credit Commit Units (CCCUs), bought in Azure, still needs a billing policy in PPAC |
| Scope of this article | The standard harness. Agents on the GitHub Copilot harness use a different, fully usage-based model that meters LLM tokens, tools, and the harness itself, and starts charging at build time, not at publish |
2. The rate card
| Agent feature | Credits | Zero-rated for M365 Copilot licensed user? |
|---|---|---|
| Classic answer | 1 | Yes |
| Generative answer | 2 | Yes |
| Agent action (triggers, deep reasoning, topic transitions, Computer-Using Agent steps) | 5 | Yes, except CUA |
| Tenant graph grounding (per message) | 10 | Yes |
| Agent flow actions (per 100 actions) | 13 | Yes, only via the When an agent calls the flow trigger |
| Text & generative AI tools, basic | 0.1 per 1K tokens (1 per 10 responses) | Yes |
| Text & generative AI tools, standard | 1.5 per 1K tokens (15 per 10 responses) | Yes |
| Text & generative AI tools, premium | 10 per 1K tokens (100 per 10 responses) | Yes |
| Content processing tools | 8 per page / image | Yes |
| Voice, classic orchestration | 10 per minute | Core activity included |
| Voice, GenAI | 35 per minute | Core activity included |
| Voice, premium GenAI | 75 per minute | Core activity included |
Two rows deserve a second look:
Agent flow actions are billed in Copilot Credits. Power Automate cloud flows are not; they stay on Power Automate licensing. Converting a classic flow to an agent flow moves its AI Builder prompts and models onto the Copilot Credit meter exclusively, and prompts inside agent flows consume credits even from the designer test panel.
Reasoning models are billed on two meters simultaneously. The feature rate (e.g. 2 for a generative answer) always applies, and the premium AI tools rate (10 per 1K tokens) is added for the reasoning model's token usage.
cost = feature_rate(operation) + 10 × (reasoning_tokens / 1000)
3. Meters stack per turn
The number on the PPAC dashboard is the sum of every meter each turn touched. Microsoft's own examples make the point:
| Scenario | Per-turn arithmetic | Daily total |
|---|---|---|
| Website support agent: 4 classic + 2 generative answers per conversation, 900 conversations/day | (4×1) + (2×2) = 8 | 7,200 |
| Copilot Chat agent with tenant graph grounding: 4 generative + 4 grounded messages, 100 unlicensed users | (4×2) + (4×10) = 48 | 4,800 |
| Autonomous order agent: 4 actions per trigger, generative orchestration | 4×5 = 20 per run | 20 per order |
Note the second row. Fifty M365 Copilot licensed users were in that population too; they contributed zero to the billed figure because the same interactions are zero-rated for them.
A single "what's the leave policy?" turn against a graph-grounded agent costs 12 credits (10 grounding + 2 generative) for an unlicensed user and 0 for a licensed one. Same agent, same prompt, same tokens. This is the single most important fact for reading a consumption report.
4. The Microsoft 365 Copilot inclusion, precisely
The "No charge" column is real but conditional. All of the following must hold:
- The scenario is employee-facing (B2E). External / customer channels are always billed.
- The agent runs under the authenticated identity of a user who holds a Microsoft 365 Copilot license. Anonymous or unauthenticated channels lose the inclusion.
- For agent flows, the run is triggered by When an agent calls the flow. A scheduled, HTTP, or Dataverse trigger inside the same flow bills at 13 per 100 actions regardless of who owns it.
- Not a Computer-Using Agent action.
- Within fair-use limits, which Microsoft reserves the right to change.
Where the inclusion applies, PPAC still records the credits. The Copilot credit consumption details grid on the Environments tab splits each agent/feature row into billed and nonbillable counts. If your tenant has a large M365 Copilot footprint, expect nonbillable to dominate.
Separately, in Microsoft 365 Copilot Chat (the no-add-on experience), declarative agents grounded only in instructions and public websites cost nothing. Agents that read shared tenant data (SharePoint, Copilot connectors) are metered in Copilot Credits for users without an M365 Copilot license, and those metered agents are off by default until an admin sets up a Copilot Studio subscription or PAYG policy.
5. Capacity topology and what happens at the edges
Think of it as three layers, resolved per environment:
Tenant pool (sum of prepaid packs)
└─ Environment allocation (optional, admin-assigned)
└─ Overage policy per environment:
a) draw from remaining tenant pool, or
b) bill to a linked PAYG billing policy
Microsoft's worked example, restated as a table:
| Environment | Allocation | Overage source | Enforcement exposure |
|---|---|---|---|
| A | 10,000 | Tenant pool after its own 10,000 is spent | Protected while its own allocation lasts; then subject to tenant-level 125% rule |
| B | none | Tenant pool | Subject to 125% rule on the remaining pool |
| C | none | Tenant pool | Same as B |
| D | 500 | PAYG billing policy | Exempt. After 500, the Azure meter takes over |
Total purchased: 25,000. Remaining unallocated pool: 14,500. If B + C alone push the pool past 125% of 14,500, enforcement fires for B and C, and for A once A has exhausted its private 10,000. D is untouched.
Two different enforcement mechanisms
| Custom agents | Agent flows | |
|---|---|---|
| Trigger point | Tenant reaches 125% of prepaid capacity | Prepaid capacity fully consumed (100%) |
| Effect | Agent is disabled after the current conversation ends; new invocations rejected with "billing issue" / "usage limit reached" | New flow runs rejected; in-flight runs complete; the parent agent keeps working for classic/generative answers and actions |
| Notification | Email to tenant admin + PPAC notice | Design-time warning in the Copilot Studio designer |
| Reset | When capacity is increased or reset | Monthly, when prepaid credits renew |
| Exemptions | PAYG-linked environments | M365 Copilot licensed users via the agent trigger; test runs from designer or test chat |
The practical consequence: a tenant can be well under 125% and still have agent flows silently failing, because the flow threshold is lower than the agent threshold.
6. Worked scenario: a large number in the Default environment with nothing to bill it
A pattern that shows up regularly, described generically:
| Attribute | Value |
|---|---|
| Prepaid Copilot Studio packs | None |
| Microsoft 365 Copilot licenses | Widely deployed, most seats assigned |
| PAYG billing policy | Linked to a single non-default environment |
| Observation | A six-figure Copilot Credit total reported against the Default environment for the month |
With no packs and no PAYG policy on Default, where does the number come from, and is anyone paying for it?
Diagnosis path, in PPAC:
- Licensing → Copilot Studio → Summary. The Pay-as-you-go credits card shows active billing plans and month-to-date billed credits. If Default is not in a plan, its usage cannot appear here.
- Capacity consumption by product. Consumption from Copilot Chat surfaces under the Microsoft 365 Copilot Chat product line, not the Copilot Studio line. This separates "agents in Teams/Copilot Chat" from "agents on a website".
- Environments tab → select Default. The consumption grid lists each agent, its product, the feature name (generative answer, tenant graph grounding, agent flow actions, and so on), and billed vs nonbillable counts.
- Also select the pseudo-environment Microsoft 365 Copilot Chat in the same picker. Copilot Chat agent usage is reported there.
In this configuration the grid typically attributes the volume to M365 Copilot licensed users interacting with agents surfaced in Copilot Chat and Teams: the credits are counted, tagged nonbillable, and zero-rated under the USL inclusion. Billed credits on Default are nil because nothing is configured to bill them. The number is a usage metric, not a cost metric.
The exposure that does exist in this configuration:
- Any unlicensed user hitting a tenant-data-grounded agent in Copilot Chat will bill at 10 credits per grounded message plus 2 per generative answer. With metered agents off by default for Copilot Chat that path is closed until an admin opens it, but if a Copilot Studio subscription is later added, it becomes live.
- A maker who publishes an agent to an external channel (website, custom app) from Default converts the whole population to billable, and Default has neither allocation nor PAYG. That agent runs on the tenant pool, which is zero, and the flow-enforcement threshold ("fully consumed") is met immediately.
- Agent flows on any trigger other than When an agent calls the flow bill regardless of the user's license.
7. Controls worth turning on before the next invoice
| Control | Where | What it does |
|---|---|---|
| Per-agent monthly limit | PPAC → Licensing → Copilot Studio → Manage Agents | Caps credits per agent, with Nearing limit / Over limit status, notifications to environment and tenant admins, and an optional hard stop that disables the agent at the cap. Works for both prepaid and PAYG environments |
| Environment overage choice | Manage Copilot Credits → environment | Decide whether overage draws the tenant pool or bills PAYG. Make this explicit for every production environment |
| Copilot author security group | PPAC → Settings → Copilot Author | Restricts who can create agents at all, which is the real control on Default |
| Move production agents off Default | Solution export/import | Default is the environment every licensed user can create in. Production agents belong in an environment with a deliberate allocation or PAYG policy |
| Agent usage estimator | microsoft.github.io/copilot-studio-estimator | Model credits by agent type, orchestration, knowledge sources, tools, and traffic before you publish |
| AI Builder Activity page | Power Automate portal | Real-time view of each AI tools / AI Builder consumption event, in AI Builder credits or Copilot Credits |
| Daily consumption data | PPAC Summary → Capacity summary | Daily granularity for the current month plus two prior months; monthly totals for 12 months. Export it before the window rolls |
8. Gotchas that generate support tickets
| Symptom | Cause |
|---|---|
| "We have no packs and no PAYG, yet PPAC shows a very large credit total" | Nonbillable consumption by M365 Copilot licensed users. Check the billed column, not the total |
| "The agent works but its flow doesn't" | Agent flow enforcement at 100% of prepaid capacity, while agent enforcement is at 125% |
| "The flow was free last month, now it bills" | Trigger changed away from When an agent calls the flow, or the flow is now invoked by unlicensed users |
| "Credits doubled after we switched models" | Reasoning model adds the premium AI tools meter per 1K tokens on top of the feature rate |
| "Costs spiked after converting a cloud flow to an agent flow" | Its AI Builder prompts/models now consume Copilot Credits only, and even designer test runs of prompts bill |
| "Testing the agent is consuming credits" | Prompts in topics from the test panel are free; prompts in agent flows are not, even from the designer |
| "Bring-your-own-model should be cheaper but the bill went up" | BYOM removes the AI tools charge but each invocation is still an agent action (5 credits), and Foundry token costs land on the Azure bill separately |
| "Agent was working, now says billing issue" | 125% enforcement on the tenant pool. Reallocate, buy packs, or link PAYG |
Sources (Microsoft Learn, accessed September 2026)
- Billing rates and management:
learn.microsoft.com/microsoft-copilot-studio/requirements-messages-management - Licensing for agents powered by the standard harness:
learn.microsoft.com/microsoft-copilot-studio/billing-licensing - Manage Copilot Credits and capacity (PPAC):
learn.microsoft.com/power-platform/admin/manage-copilot-studio-messages-capacity - AI tools / AI Builder licensing and Copilot Credits:
learn.microsoft.com/ai-builder/message-management - Using agents in Microsoft Copilot Chat:
learn.microsoft.com/copilot/agents - Licensing and cost considerations for Copilot extensibility:
learn.microsoft.com/microsoft-365/copilot/extensibility/cost-considerations - Copilot Studio prepaid capacity packs for Copilot Chat and SharePoint agents:
learn.microsoft.com/microsoft-365/copilot/pay-as-you-go/copilot-capacity-packs - Meters for Microsoft Copilot pay-as-you-go:
learn.microsoft.com/microsoft-365/copilot/pay-as-you-go/meters - Billing overview for the GitHub Copilot harness:
learn.microsoft.com/microsoft-copilot-studio/agents-experience/billing-credit-overview - Microsoft Copilot Studio Licensing Guide (PDF):
go.microsoft.com/fwlink/?linkid=2320995
Tags: CopilotStudio, CopilotCredits, PowerPlatform, PowerPlatformAdmin, Microsoft365Copilot, AgentFlows, PayAsYouGo, Licensing, FinOps, M365Admin, AIAgents, PPAC
No comments:
Post a Comment