Content reviewed: 13 October 2026
Budget Copilot Studio from billed operations, not conversation counts. One prompt can trigger an answer, tenant grounding, an action and extra model processing. Consumption depends on agent design, harness and user identity.
Separate author access, audience entitlements, payment sources and operational limits before ordering. A credit pack funds consumption; it neither assigns source-data permissions nor establishes a universal spending cap.
Author access and the audience
The current standard-harness licensing guide lists several access routes: a free Copilot Studio User License tied to a tenant prepaid subscription, the configured Copilot Studio authors group, applicable Microsoft 365 Copilot use and a trial. Do not assume the same route for every authoring model. Check the selected configuration and environment permissions.
PAYG lets you start through Azure without buying fixed capacity upfront. It does not itself assign authors-group membership or environment roles. A trial supports creation and testing in the test panel, but not publication. Assign creation, approval and operation to specific people.
A published agent’s audience does not automatically need the authoring Copilot Studio User License. Usage and billing still depend on channel, feature and authentication. Assess the licences of applications or systems the agent accesses separately.
What credits measure
| Billed feature | Copilot Credits | Note |
|---|---|---|
| Classic answer | 1 | Static response; can include a billed conversation start |
| Generative answer | 2 | Other features used can add consumption |
| Agent action | 5 | Count actual metered events, not arbitrary flow lines |
| Tenant graph grounding | 10 | Added to an answer that uses it |
| Agent flow actions | 13 per 100 actions | A distinct flow type in Copilot Studio |
| Reasoning model | 10 per 1,000 billed tokens | In addition to the main feature rate |
This summarises current standard-harness rates, not the entire rate card. Documentation has further AI and voice items. Bring-your-own-model configurations, including your own Azure Foundry models, are billed separately. Do not automatically apply these rates to a GitHub Copilot harness agent.
A reasoning model combines the feature rate and additional token consumption. If an answer uses generative answering and reasoning, include both meters. A completed customer task has no universal fixed credit price.
Consumption can occur without a customer reply: the FAQ describes billed proactive greetings and agent readiness. Engaged conversations depend on user response. Zero engagement does not establish zero billing.
When a user licence covers usage
Included standard employee scenarios under Microsoft Copilot USL require a licensed user and execution under that authenticated identity, subject to fair use and feature conditions. For agent flows, inclusion is limited to the When an agent calls the flow trigger. Other triggers are not included merely because the author is licensed. The rate-table notes specify these conditions.
Computer use is outside ordinary Copilot USL inclusion. GitHub Copilot harness agents have their own consumption model and can use credits during building, testing and evaluation. Microsoft distinguishes these agents. A blanket “testing is free” claim is insufficient: standard agent-flow tests do not meter their flow actions, while another harness can be billed.
Packs, PAYG or annual prepurchase
| Model | What you acquire | Decision factors |
|---|---|---|
| Prepaid monthly pack | 25,000 credits per pack | Allocation, usable capacity and further draw sources |
| PAYG | Actual metered consumption through Azure | Variable demand, active billing and operational controls |
| Copilot Credits prepurchase plan | A one-year prepaid CCCU pool in Azure | Eligible products and specific commercial terms |
A monthly capacity pack and annual prepurchase are different products. Standard harness licensing requires an Azure billing policy for the annual plan. PAYG can therefore also be the route for consuming a prepurchase pool; the PAYG label does not always mean entirely uncommitted consumption.
Unused monthly capacity does not carry forward. Check reset timing: the Studio FAQ specifies the first calendar day for environment usage, while the Copilot Chat pack guide describes replenishment at the start of the monthly billing period. Confirm the cycle for the actual service and reconcile it with the report; purchase date is not a safe universal assumption.
Our 26,000-credit calculation
A hypothetical standard agent on a customer website completes 1,000 tasks per month. Each task produces exactly two billed generative answers and one metered agent action. One of those answers also uses reasoning with 1,500 billed tokens. Outside completed tasks, assume another 2,000 billed classic answers. The audience has no included employee Copilot USL usage.
| Model component | Calculation | Credits |
|---|---|---|
| Main features of completed tasks | 1,000 × (2 × 2 + 1 × 5) | 9,000 |
| Additional classic answers | 2,000 × 1 | 2,000 |
| Reasoning on one answer per task | 1,000 × (1,500 / 1,000) × 10 | 15,000 |
| Total | 9,000 + 2,000 + 15,000 | 26,000 |
This is our transparent model, not customer measurement or a forecast for every agent. It excludes additional graph grounding, retries, voice, computer use, AI tools and external services. Replace assumptions with actual billed events and tokens after a pilot.
Two packs provide 50,000 credits, leaving 24,000 unused in this model. A supported combination of one pack and PAYG would have 1,000 credits beyond prepaid capacity. Compare another pack’s price with the price of those 1,000 credits and order conditions. This does not automatically recommend PAYG: continuity and limits still matter if demand rises or a process fails.
What actually stops consumption
| Control | Effect | Boundary |
|---|---|---|
| Environment allocation | Reserves prepaid capacity | Does not alone block tenant-pool draw or PAYG |
| Draw from tenant capacity disabled | Prevents unallocated tenant-pool draw | Does not stop linked PAYG |
| Agent limit with Stop usage | Can stop that agent at the limit | Not an aggregate cap for the environment |
| Azure budget and alert | Notifies owners about cost status | Does not automatically stop the agent |
Check these controls in PPAC under Licensing → Copilot Studio and the relevant environment or agent settings. Microsoft separates allocation, tenant draw, PAYG and Stop usage. An environment-group rule can make a local draw setting read-only. Plan an unavailable-agent fallback and assign an operational contact.
The agent answers but its flow fails
For standard custom agents using prepaid capacity, documentation describes general enforcement at 125% of the relevant prepaid capacity. The ongoing conversation is not interrupted, but later invocations are blocked. Agent flows have a different threshold: when prepaid capacity is fully consumed, new runs are blocked, ongoing runs finish and the parent agent can continue non-flow responses. Billing documentation states the difference and exemptions.
125% is not purchased extra capacity or a recommended operational budget. Investigate allocation, available tenant pool, linked PAYG and the actual billed type. Ordinary Power Automate cloud flows use Power Automate licensing; they are different from Copilot Studio agent flows.
Copilot Chat has its own setup
Microsoft describes a standalone Copilot Chat credit policy without Azure/PAYG, or a policy paired with PAYG. User groups and the relevant environment allocation must be configured. The current note says credit policies are available only for Copilot Chat; SharePoint agents should continue their PAYG setup. Do not treat the capacity-pack guide as a universal configuration for every channel.
What to measure before ordering
- Correctly completed tasks and repeated attempts.
- Billed answers, actions, grounding, reasoning tokens and other tools.
- Test, preview-feature and production behaviour in the actual harness.
- Agent and environment usage, report freshness and payment source.
- Behaviour at exhausted capacity, failed flows and unavailable sources.
- Licences and costs of connected applications, data and external models.
Measure errors and preparation as well as a successful demonstration prompt. PPAC consumption data is aggregated and can be delayed. Reconcile costs with actual operational tests, and approve expansion only after explaining differences between estimates and observed usage.
Frequently asked questions
Is one conversation one Copilot Credit?
No. One interaction can combine several billed features and additional token usage.
Are building and testing always free?
No. Conditions vary by harness and feature. GitHub Copilot harness can use credits during building and evaluation.
Does an Azure budget stop further billing?
Not automatically. It is a notification. Check an available agent limit with Stop usage and other capacity sources.
Does every audience member need an authoring licence?
Not automatically. Author access is separate from published-agent usage and its consumption model.
Why does chat work while an agent flow will not start?
Enforcement can block new flow runs after prepaid capacity is exhausted while ordinary agent responses continue. Also check the trigger and any included usage.
Can a pack work without PAYG?
Yes with a supported Copilot Chat credit policy. Do not assume the same availability for SharePoint or every channel.
Prepare the harness, channel, audience, measured event mix and availability requirement before ordering. Axeti can check the relevant licensing and commercial options. See Copilot Chat: agent licences and usage for individual user scenarios.








