Some Copilot agents cost extra; others do not. Agents built on instructions and public websites are included with Copilot Chat, while agents that reach shared company data consume Copilot Credits unless the user's Microsoft 365 Copilot licence covers it. Prepaid packs or pay-as-you-go billing fund the rest; your policy decides whether excess use stops or reaches an Azure bill.
The crucial question is what the agent does for each user, not whether its icon says “Copilot”. A policy-answering agent that reads a shared library is a different billing case from an agent using public websites. Check the live licensing and billing screen before a rollout, because the same agent can produce a different bill for differently licensed users.
Start with the agent's audience and sources
Make an inventory with four columns: who will use the agent, what it can read, what it can do, and which licence each user has. This takes less time than trying to interpret a credit report after launch. Microsoft distinguishes included declarative agents grounded in instructions and public websites from pay-as-you-go agents that access shared tenant content for users on its Basic Copilot access. It also says that users with a Microsoft 365 Copilot licence (Microsoft now also calls it Microsoft Copilot) are not charged for classic answers, generative answers or tenant graph grounding when an agent in Copilot Chat, Teams or SharePoint runs with their own signed-in identity. Do not turn that into a blanket promise that every possible flow or connected service is free: agent flows started by other triggers, and agents that run on their own after an event, still consume credits.
For a furniture maker, a public agent answering “What are the published care instructions for this finish?” is a candidate for included access. The answer lives on a public care page. Change the question to “What did we promise this customer in their private quotation?” and the agent must look in a shared business system. That source change is the point at which an owner should ask the administrator for the applicable licence and billing policy. The visual similarity of the two chat windows conceals a material cost difference.
Write the inventory before building: “Sales staff: six; source: shared quotation library; actions: read and draft only; current licence: record per staff member; expected questions: 12 per person per working day.” If the source owner cannot say who has permission to see each quotation, pause the knowledge connection. An agent should respect the source's permissions, but poor existing permissions can still expose the wrong document to someone who was already allowed to open it. The SharePoint permissions tutorial explains that preparation in more detail.
The Complete Excel Course with AI Integration
From formulas and pivot tables to Power Query, DAX, VBA and AI with Copilot, ChatGPT and Claude — hands-on from the first lesson.
By Sagnik Bhattacharya and Virginia Thorn, Primetime Emmy Award winner
Buy nowWhat one Copilot Credit measures
A Copilot Credit is a consumption unit, not a user seat and not necessarily one whole conversation. Microsoft's current Copilot Studio table lists a classic answer at one credit, a generative answer at two, an agent action at five, and tenant graph grounding for a message at ten. A single interaction may combine features. The documentation itself illustrates a response using ten credits for graph grounding and two for a generative answer. A bot that performs several actions before replying can consume more than a simple answer bot.
| Agent activity | Published credit rate | Owner's implication |
|---|---|---|
| Classic answer | 1 credit | Useful baseline for a narrowly scripted response. |
| Generative answer | 2 credits | Budget for each answer, including follow-up questions. |
| Agent action | 5 credits | Count an action as well as the reply around it. |
| Tenant graph grounding | 10 credits | A source-rich answer may combine this with other charges. |
| Agent flow actions | 13 credits per 100 actions | High volumes can matter even if each action looks small. |
Those are listed feature rates, not a quote for every prompt. A business can change the agent's instructions, add a knowledge source, or enable an action, and the consumption pattern changes. Microsoft also documents separate tool rates, and when a reasoning model orchestrates the agent it adds a premium charge of 10 credits per 1,000 billed tokens on top of the feature rates, so a single long answer can run to hundreds of credits. For the pilot, record the actual credit total for a representative set of published interactions rather than multiplying “one chat” by a guessed universal rate. The Copilot Studio test chat is useful for checking behaviour, but Microsoft's billing FAQ says agent activity there does not consume credits and does not enter consumption reports. Use a controlled published test for a billing measurement.
Another illustrative case: a specialist food importer sets an agent to answer 30 staff questions a day. If every answer used only the two-credit generative-answer feature, 22 working days would produce 1,320 credits. If half the requests also invoked ten-credit tenant grounding, the estimate becomes 4,620 credits: 1,320 plus 330 grounded requests times ten. That is a planning calculation, not a prediction of what the published agent will meter. Keep an example of a grounded answer and compare the measured credit entry before approving the monthly budget.
A six-person pilot with numbers
Consider an illustrative six-person furniture maker that wants an internal agent to find the latest care instructions and draft responses to trade customers. Each person expects ten questions on each of 20 working days: 1,200 questions per month. The owner tests 40 representative published requests. Suppose the observed average is four Copilot Credits per request, including the features actually invoked. At that measured rate, the forecast is 4,800 credits. Add a 25% allowance for retries, onboarding and unusually busy days: 6,000 credits. The allowance is a management choice, not a Microsoft requirement.
Microsoft's pay-as-you-go meter for Copilot Chat and SharePoint agents is listed at $0.01 per message, and Microsoft says the switch from messages to Copilot Credits did not change the pay-as-you-go rate, so 6,000 billable credits would be about $60. Confirm the rate in the tenant's current billing view before using that dollar number in an approval, because local currency and agreements can differ. A 25,000-credit prepaid capacity pack is much larger than this illustration's forecast; read its price and terms at checkout rather than inferring them from its size. Pack credits reset each month and unused capacity does not carry over, so buying one simply because the organisation has six people could leave most of it unused.
The owner can put an approval rule on one page: “Trial audience: six named staff; trial agent: care-instruction search; approved source: current care library; target volume: 1,200 requests monthly; planning allowance: 6,000 credits; weekly review: administrator plus owner; stop condition: source error or unapproved action; expansion: only after four weeks of measured use.” This filled-in example is deliberately more useful than a bare “set a budget” instruction. It gives someone a reason to check the report and a criterion for changing scope.
Suppose the published test instead averages 12 credits because it uses tenant grounding and a generated answer on every request. The monthly forecast becomes 14,400 before the allowance and 18,000 after it. That remains within one 25,000-credit pack, but is three times the first forecast. A second source or action could change it again. The business should first ask whether every question needs that grounding, then see whether a narrower source set can answer the actual queries accurately. Do not strip useful context merely to make a cost line look smaller; measure answer quality at the same time.
Three ways to pay for metered agents
Microsoft's capacity-pack guidance describes a prepaid pack of 25,000 Copilot Credits each month. For Microsoft Copilot Chat, an administrator can connect those credits through a Copilot credit policy without a pay-as-you-go policy or Azure subscription. The credits can also be paired with pay-as-you-go overage. The third option is pay-as-you-go only, where eligible use reaches the linked Azure subscription. Microsoft says Copilot credit policies are currently available only for Copilot Chat; SharePoint agents still use pay-as-you-go billing, so do not assume the Copilot Chat credit-policy screen applies everywhere.
| Funding choice | What happens at the allocation edge | When it may suit a small firm |
|---|---|---|
| Prepaid policy only | Covered Copilot Chat use stops until credits replenish or more capacity is added. | Experiment where a firm prefers a firm usage boundary. |
| Prepaid plus overage | Eligible excess is billed to the linked Azure subscription. | Service needs continuity and someone watches the bill. |
| Pay-as-you-go only | Billable use is metered against Azure from the start. | Small or uncertain volume, with controls and reporting. |
An Azure subscription is therefore needed for the pay-as-you-go arrangements, but not for the documented prepaid-only Copilot Chat policy. That is an important correction to older advice that says any shared-data agent requires Azure. The administrator still needs the right role to buy and allocate packs, and the pack is a pooled resource rather than a personal allowance for each employee. Microsoft says a Global Administrator or Billing Administrator can buy packs, but only a Global Administrator can enable pack usage; Billing and AI administrators can set up pay-as-you-go but cannot activate a pack. Let the administrator handle those access choices instead of sharing a powerful admin login with the person who writes agent instructions.
For an online clothing store, a temporary product-launch support agent could be a poor fit for a year-long assumption about capacity. Start from expected queries during the launch, measure an actual published sample, and compare the available pay-as-you-go and prepaid terms at checkout. If the business needs the agent every day for staff operations, a prepaid-only stop could frustrate the team when credits run out. That operational cost belongs in the decision beside the dollar amount. The Azure budget setup tutorial deals with the configuration step after this choice.
Budget alerts are signals, not spending limits
There are several separate controls, and they answer different questions. A Copilot credit policy scopes who may consume a prepaid pool. The Power Platform admin centre shows allocated and remaining capacity. Its overage setting determines whether the environment can draw from other capacity or bill a pay-as-you-go plan. Microsoft also describes a notification when capacity nears a selected percentage. Azure cost budgets can notify people about usage against a threshold, but Microsoft's budget documentation says resources are not affected and consumption is not stopped, and cost data typically arrives 8 to 24 hours late. An alert is evidence that someone should investigate; it does not by itself guarantee that requests stop at the figure.
The control that can actually stop a Copilot Studio agent is a monthly credit limit set for that agent in the Power Platform admin centre, under Licensing, Copilot Studio, Manage Agents. Once a limit is set, administrators can receive alerts as the agent nears it and can choose a hard stop that turns the agent off at the limit. Microsoft adds two caveats: limits are not applied instantly, so an agent can keep answering for a while after reaching one, and an agent's limit does not stop the agent flows it calls, which are checked against the environment's capacity separately. For the furniture maker, a 6,000-credit limit with the hard stop on would turn the forecast into a ceiling, give or take the delay Microsoft warns about.
Use an illustrative alert ladder for the furniture maker: notify the owner and administrator at 50% of the approved monthly allowance, review at 75%, and require a decision before broadening the pilot at 90%. If a prepaid-only policy governs the trial, explain to staff that service can pause when capacity is exhausted. If the agent supports a time-sensitive quotation desk, the owner may choose continuity with a pay-as-you-go connection, but that choice needs a named reviewer and a weekly look at the Azure charge. Record which person can turn the billing policy off and who can remove the agent from the audience.
One failure is treating a 25,000-credit pack as if it belonged solely to the new agent. Capacity is pooled, and another environment or agent may use it depending on allocation. For example, an agent that summarises internal documents is introduced midway through the month and consumes 8,000 credits. The care-instruction pilot's original 18,000-credit high estimate plus that new use exceeds 25,000. The relevant check is remaining capacity in the environment and the consumption by agent, not the number on the original purchase order. Give each material agent a label in the report and note the change that caused its usage to rise.
Find the charge before approving more users
Ask the administrator to show the current Copilot billing policy, the allocated environment and the Azure subscription, if one is linked. Then run a small published test with a named user from each licence group. Record the time, question, knowledge source, answer type and reported credits. A test with only the fully licensed administrator can give a falsely reassuring picture of costs for unlicensed staff. A test with only generic public questions can miss the charge triggered by shared business content. Keep the questions representative of what staff will actually ask.
Here is a copyable test record in prose: “Question 07, sales staff licence group, ‘Find the approved care note for the walnut dining table and draft a short reply’; source returned: current care library; answer checked against approved note: yes; action taken: none; reported credits: four; issue: answer omitted drying time.” The owner should fix the missing information in the source or instructions, rerun the test, and log the new result. Credit efficiency is pointless if the answer requires a human to repair every time.
A second test should be deliberately adversarial but ordinary: ask for an obsolete care sheet or a quotation the user lacks permission to view. The safe result is an honest inability to retrieve it or a link to the authorised current item. If the agent presents a stale sheet as current, locate that item in the knowledge source and remove or archive it. If a restricted quotation appears, investigate the underlying permissions before opening access to more people. The agent knowledge-sources tutorial helps decide what should be connected at all.
Microsoft has reporting in the Power Platform admin centre for Copilot Studio capacity, including downloadable consumption reports by environment, agent or user, and a Cost Management page in the Microsoft 365 admin centre for pay-as-you-go services. Azure then shows charges for a connected pay-as-you-go policy. These views may have different timing and grouping. Reconcile the test over a defined period rather than declaring a discrepancy from one instant refresh. Save a screenshot or export for the approval record, showing the date and the agent version tested. Repeat the measurement after a new connector, tool or flow is added, because the earlier four-credit average will no longer be a safe assumption.
Questions that change the decision
First, is this an agent that staff could replace with a plain searchable page? For a short care guide with ten stable answers, a well-maintained document or FAQ may be faster to operate. An agent becomes more useful when people ask varied questions across changing material and need a sourced answer. The Copilot Studio tutorial covers the broader build-or-buy decision; here the only concern is whether the resulting activity requires metered funding.
Second, does the firm really need an action? Reading a source and drafting a proposed response is easier to test than sending an email, changing a record or triggering another workflow. Suppose a craft brewery supplier wants an agent to locate a current delivery policy and then automatically alter a customer order. The first step is an information retrieval job; the second can create operational harm and may add billable action activity. Split the workflow so a person confirms the change. This also produces a clearer credit trace when the owner asks why an interaction cost more than expected.
Third, is the agent available to everybody or a defined pilot group? A firm with 20 staff does not need 20 users on day one if four experienced staff can cover the main questions. Start with real users across different permissions, then expand only after the answer checks and credit forecast are stable. If a staff member has an eligible Microsoft 365 Copilot licence, verify whether the specific use is included for that person; do not count an included interaction as evidence that everyone else will be free. Nor is the licence automatically the cheaper route: Microsoft lists the Copilot Business add-on at $21 per user a month on a yearly plan ($18 for the first year on yearly plans bought by 31 December 2026, or $25.20 on a monthly subscription), which is the same money as 2,100 credits at the $0.01 pay-as-you-go rate. If agents are the only reason to buy it, compare that figure with the person's measured monthly credits; the licence also brings Copilot in Word, Excel and Outlook, which may justify it on other grounds. Licence status and agent design both belong in the measurement sheet.
Finally, decide whether a service pause is tolerable. A prepaid-only policy can provide a clear boundary for a non-critical pilot. A customer-facing operational agent may need continuity, which can justify pay-as-you-go overage, but only with monitoring, ownership and a tested route to disable the agent. Ask the owner to sign off on that trade-off in one sentence. “The agent may pause when capacity is exhausted” and “The agent may continue and create an Azure charge” are materially different permissions. Staff should know which one the business chose.
After a month, compare measured requests, measured credits, answer correction rate and time spent maintaining sources with the forecast. If use is low, avoid buying a larger pool merely to make a dashboard look tidy. If use is high and the answers are dependable, review the pack and licence options with the current vendor terms. The bill is one piece of a working agent; the useful output, human review effort and exposure from connected actions determine whether the spend earns its place.
Copilot agent billing questions
Does creating an agent itself incur a Copilot Credit charge?
A draft alone is not a useful cost unit. Microsoft says activity in the Copilot Studio test chat does not consume Copilot Credits, while published agent activity can. Check the author licence, the knowledge sources, the audience and the billing policy before allowing real users in.
Can prepaid Copilot Credits work without Azure?
Yes, for the documented Microsoft Copilot Chat credit-policy-only configuration. A capacity pack supplies 25,000 credits per month and can be scoped to users without an Azure pay-as-you-go connection. When the allocated credits run out, covered use stops until replenishment unless you enable paid overage. Microsoft says credit policies currently cover Copilot Chat only; SharePoint agents still use pay-as-you-go.
Does an Azure budget stop Copilot agent spending?
No. Microsoft's budget documentation says a budget sends notifications when a threshold is crossed but does not stop consumption. To limit a Copilot Studio agent, an administrator can set a monthly credit limit for it in the Power Platform admin centre, with notifications and an optional hard stop, and can control what each environment may draw. Limits are not applied instantly, so leave a margin.
Further reads
- Microsoft 365 Copilot Pricing: Business vs Enterprise Licences — Compare seat licences with usage charges.
- Copilot Studio Pricing: What a Small Business Customer FAQ Agent Costs — What a customer FAQ agent costs in Copilot Studio: credits per chat, pack versus pay-as-you-go break-even, and cheaper ways to answer.
- AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.
Sources: Microsoft Learn: What is Microsoft Copilot (licence comparison); Billing rates and management for Microsoft Copilot Studio; Copilot Studio licensing; Use capacity packs vs pay-as-you-go billing for Microsoft Copilot Chat; FAQ for Copilot Studio billing and licensing; Meters for Microsoft Copilot pay-as-you-go services; Manage Copilot Credits and capacity for Copilot Studio; Azure Cost Management budgets. Microsoft Copilot business plans and pricing page. All read 5 November 2026.