Copilot Studio is one of the few AI agent platforms where you can work out the bill before you build anything. The pricing is public, the billing modes are documented, and the meter that matters — Copilot Credits — is the same regardless of how many people in your business talk to the agent. That is genuinely useful for an Australian SME trying to decide whether an internal service desk agent or a quoting assistant is worth committing to. It is also where most of the confusion sits, because the credit meter is not the only meter, and a Microsoft 365 licence does not pay for any of it.
Here is what the numbers actually are, what sits outside them, and what to watch when you model it.
The three ways to buy
Microsoft sells Copilot Studio capacity at the tenant level rather than per seat. You buy a pool of Copilot Credits, agents draw down on the pool, and everyone in the tenant shares it. There are three commercial routes into that pool.
Capacity packs
The headline number on Microsoft's Copilot Studio pricing page is 25,000 Copilot Credits for US$200 per month, billed annually. Converted at roughly US$0.65 to the Australian dollar, that is about A$310 a month, or A$3,700 a year per pack — rates move, so treat it as an indicative figure rather than a quote. Packs stack: run out of credits and you add another pack.
This is the mode most SMEs land on, and it behaves well for predictable internal workloads. You know the annual commitment on day one, and the finance conversation is a single line item rather than a variable cloud bill.
Pay-as-you-go
The second route is metered billing through an Azure subscription at US$0.01 per Copilot Credit, documented in Microsoft's billing and licensing reference. That is roughly 1.5 Australian cents per credit at the same conversion.
Do the arithmetic and the two modes do not price the same. A capacity pack works out at US$0.008 per credit; the identical 25,000 credits bought on pay-as-you-go at US$0.01 each would cost US$250, not US$200. The pack is roughly twenty per cent cheaper per credit, and that discount is what you are paid for committing annually.
The difference beyond the unit rate is risk direction. Pay-as-you-go costs nothing when nobody uses the agent, and costs whatever it costs when they do. Capacity packs cap the spend and waste the remainder. For a pilot where you genuinely cannot predict volume, the higher unit rate is usually worth paying — you learn the real consumption shape in a month and switch. For a production agent with steady traffic, packs are both cheaper and free of surprises.
Pre-purchase commit units
The third route, described in Microsoft's Copilot Studio licensing guide, is an annual pre-purchase using Copilot Credit Commit Units. You commit to a volume up front and draw it down over the year. This suits organisations that already buy Azure on commitment and want agent spend on the same paper. For a twenty-person business, it is usually more contract than the workload justifies.
The meters that sit outside the credit pool
Copilot Credits are not the whole bill. Dataverse capacity is charged separately, and it has more than one price depending on how you buy it. The August 2026 Power Platform licensing guide lists the capacity add-ons at US$40 per GB per month for database capacity and US$2 per GB per month for file capacity — about A$62 and A$3.10 per GB per month. Consumed through the separate pay-as-you-go meters rather than bought as add-ons, the same capacity runs US$48 and US$2.40 per GB per month, roughly A$74 and A$3.70.
Those are very different numbers, and the difference matters. Database capacity at A$62 to A$74 per gigabyte per month is expensive storage by any measure, and file capacity is a small fraction of it. Most tenants have a Dataverse entitlement already through existing Power Platform licensing, and a well-designed agent may not add much to it. But an agent that logs every conversation turn, stores transcripts, writes case records and keeps attachments in Dataverse rather than SharePoint can grow that line quietly. The design decision about where an agent's data lands is a cost decision, not just an architecture one.
This is the single most common gap between a Copilot Studio estimate and a Copilot Studio invoice. The credit maths gets modelled carefully; the storage sitting under it does not get modelled at all.
What your Microsoft 365 licences do not cover
The assumption that breaks budgets is "we already pay for Microsoft 365, so agents are included". They are not. Copilot Studio capacity is purchased separately from Microsoft 365 seats, and the credit consumption an agent generates is billed against that capacity regardless of how many M365 licences the tenant holds.
A Microsoft 365 Copilot seat and a Copilot Studio agent are different products solving different problems. The seat gives an individual an assistant inside Word, Outlook and Teams. Copilot Studio gives the business a purpose-built agent that answers from your own knowledge, calls your own systems and follows rules you set. If the plan is an agent that handles supplier enquiries or triages internal IT tickets, that is Copilot Studio, and it has its own meter.
The second-order version of the same trap: credit consumption is not uniform. Different agent actions draw different numbers of credits, so a conversation that just answers from a knowledge base and a conversation that triggers an autonomous action across three systems are not equivalent line items. Before you divide 25,000 by your expected conversation count, read the consumption table in the licensing guide against the specific actions your agent will actually perform.
The Australian residency question
For Australian tenants, data stays within the Australia datacentre region, built on Microsoft's Sydney and Melbourne region pair. For most SMEs — professional services firms, healthcare practices, anyone with a client contract that specifies onshore storage — that resolves the question that usually stalls the project.
It does not resolve every version of it. If your obligation is stricter than "in Australia" — a government contract requiring assessed sovereign hosting, or a data classification that cannot touch a hyperscaler at all — Copilot Studio is the wrong starting point and a self-hosted or Australian-hosted agent stack is the conversation to have instead. Knowing which of those two situations you are in before you scope the build saves a rebuild later.
What Copilot Studio is good at, and where it stops
It is strongest when the work lives inside Microsoft. If your documents are in SharePoint, your identity is in Entra, your users are in Teams and your records are in Dynamics or Dataverse, the connectors, the authentication and the governance are already there. An agent that answers policy questions from SharePoint content, respecting the permissions each user already has, is close to the platform's ideal use case.
It gets weaker as the work moves outward. Multi-step orchestration across a dozen non-Microsoft SaaS tools, heavy data transformation, or scheduled batch processing against a Postgres database are jobs where a workflow engine does better. That is why real deployments are usually mixed: Copilot Studio as the conversational front door and Microsoft-native surface, n8n or similar handling the integration plumbing behind it, and Retool where a human needs a screen to review or approve what the agent produced. Choosing the right split — and knowing when Copilot Studio alone is genuinely enough — is most of the value in scoping a Copilot Studio build properly before you commit to an annual capacity pack.
The cost line nobody models
Licensing is the part you can look up. The part that gets underestimated is everything after the first working agent: what happens when a connector's credentials expire, who is notified when the agent starts answering wrong because a SharePoint library was restructured, how you separate a test agent from the production one, how you log what the agent told a customer, and who owns the failure at 6am. Platform pricing tiers hint at this across the market — governance features like environments, version control and audit logging consistently sit in the upper plans, because that is where operational reality starts. Budget for the operating model, not just the meter.
The first move
Before you price anything, count. Pick one candidate workload and estimate its monthly interaction volume — enquiries handled, tickets triaged, documents summarised — then map each interaction to the specific actions the agent would take. Check that against the credit consumption table in the current licensing guide. If a single pack's 25,000 credits comfortably covers it, the commercial case is straightforward. If it does not, you have learned something important before signing an annual commitment.
If you would rather have that modelled properly, we will build the estimate with you: expected credit consumption for your actual workload, the Dataverse capacity implication of the data design, an honest read on whether Copilot Studio is the right platform or whether the job belongs in a workflow engine, and a two-to-four-week build scope with a number attached. Start with the return-on-investment calculator and we will come back with the figures and a plain answer on whether it is worth doing.