Copilot Studio has a published Australian price and a published quota sheet, and the second one is the document that decides what your agent can be. Microsoft's Australian pricing page lists Copilot Studio as a tenant-wide licence with 25,000 Copilot Credits per capacity pack at AU$299.30 per pack per month, with an Azure subscription required to run agents. That number is easy to budget against. What catches teams out is not the pack price — it is discovering, three weeks into a build, that the agent they designed wants to behave like a transactional application and the platform will not let it.
Designed the other way round, Copilot Studio is excellent. It sits inside the tenant you already pay for, it inherits Entra ID identity, it answers in Teams where people already are, and it reaches Dataverse and SharePoint without a bespoke integration layer. The trick is knowing which work belongs inside the agent and which work should be handed to an execution engine that counts differently.
What a credit pack actually buys
Credits are consumed by what agents do, not by how many agents exist. Conversations, generative answers, actions and automated runs each draw down the same pool. That makes the useful budgeting exercise a usage model rather than a headcount: estimate monthly conversations per agent, estimate how many of those involve a generative answer against knowledge sources versus a deterministic topic, and estimate how many trigger a connector action. Multiplying agents by a flat fee will give you a number that bears no relationship to the invoice.
Worth noting before you build a business case on a figure you found in a US blog post: Microsoft's own licensing guidance and US-facing material quote US$200 per pack per month billed annually, which does not reconcile cleanly with the AU$299.30 shown on the Australian page as at 6 October 2026. Regional pricing, tax treatment and billing terms may all be in play. For an Australian budget, use the Australian page, note the date you checked it, and confirm in writing whether the figure is GST-inclusive, whether an annual commitment applies, and what the Azure dependency adds on top.
The quota sheet is the design document
Microsoft publishes agent quotas and service limits in a single reference that was last updated on 6 October 2026, with the page originally dated 10 January 2026. A selection of the documented limits:
| Limit | Documented ceiling |
|---|---|
| Knowledge sources per agent | 500 |
| Skills per agent | 100 |
| Topics per agent (Dataverse environments) | 1,000 |
| Trigger phrases per topic | 200 |
| Agents per team | 50 |
| Sessions per user per 24 hours, all agents in a tenant | 10 (in the relevant quota category) |
| Power Platform requests per 24 hours | 250,000 in one category, 6,000 in another |
Read that list as an SME and most of it looks generous. Nobody is writing 1,000 topics. Nobody is attaching 500 knowledge sources. The two rows that matter are the last two, and they are the ones that get skimmed.
Sessions and requests are the limits that bite
A session cap measured per user per 24 hours across all agents in the tenant is a usage shape constraint, not a scale constraint. It is fine for an agent people consult a few times a day. It is not fine for an agent you have quietly turned into the interface for a repetitive operational task — logging a delivery exception, checking a job status, requesting a purchase order number — where the same dispatcher might hit it fifteen times before lunch. The agent does not fail because of credits. It fails because of how it is being used.
Power Platform request limits behave the same way. Every connector call, every Dataverse operation, every flow step consumes requests from a daily allowance. An agent that answers questions from indexed documents barely touches them. An agent that loops through a list of records and writes back to a line-of-business system can burn through an allowance in a way that is invisible until a Tuesday when volume is high and things start queuing.
The design conclusion is unglamorous and reliable: Copilot Studio should own conversation, intent, identity and retrieval. It should not own bulk transactional work.
Where the agent should stop and something else should start
The clean boundary is this. The agent interprets what the person wants, confirms the specifics, checks that this person is allowed to ask, and then calls out to one orchestration endpoint. Everything after that call — the fifteen API hops, the retries, the pagination, the error branches, the writes to three systems — happens outside the agent and reports back.
That is exactly the work n8n is built for, and its accounting model is the reason. Under n8n's pricing, an execution is one complete workflow run rather than one node or step, so a workflow with twenty nodes still counts as a single execution when it fires once. The August 2025 pricing change removed active-workflow limits and moved plans to unlimited workflows, steps and users, with volume counted in executions. The Business plan was listed at €667 per month billed annually including 40,000 executions when checked on 6 October 2026, and is currently offered for self-hosted n8n rather than hosted n8n; the self-hosted Community and Registered Community editions are available at no software charge, with paid capabilities gated behind a licence key.
Push the twenty-step integration into n8n and it costs you one execution. Leave it inside the agent as a chain of connector actions and it costs you credits, Power Platform requests and a chunk of your skills budget. Same business outcome, two very different consumption curves.
The watch-out on the n8n side is the mirror image: executions rise through schedules and polling, not through complexity. A workflow on a five-minute schedule is roughly 8,640 executions a month before anyone uses the system. Event-driven webhooks from the agent are cheap. Polling is not.
Retool for the screens the agent should not become
The third failure shape is an agent that has become a form. Once a conversation needs five fields, a date picker, a line-item table and a supervisor approval, chat is the wrong surface regardless of quota. That work belongs in an internal application with a real data model and real permissions. Retool currently offers cloud, self-hosted and air-gapped deployment options, which matters for Australian organisations with network or data-location constraints — though the deployment label alone does not settle residency for licensing, telemetry, logs, backups, support access or any model calls. Establish those in the contract, per service, rather than assuming them from the word "self-hosted".
In practice the pattern that holds up is: agent in Teams for the ask, n8n for the orchestration, Retool for anything that needs a screen, approvals and an audit trail. We spend most of our time on designing and governing Copilot Studio agents inside existing Microsoft 365 tenants, and the architecture conversation is almost always about where the agent hands off rather than how clever the agent is.
Designing for the limit you will actually hit
Three habits make the difference between an agent that scales and one that becomes a support ticket.
Instrument before you launch. Track sessions per user per day, credit consumption per agent per week, and Power Platform request volume. Without those three series you cannot tell a usage spike from a design fault.
Give every limit a defined behaviour. Decide now what the agent says and does when it cannot complete a request — a clear message, a handoff to a person, a queued job, a logged exception. The default behaviour of an agent at a limit is to be confusing, and confusing agents get abandoned.
Keep the credit-expensive path narrow. Generative answers against a large knowledge corpus are the costly part. Common, high-volume, unambiguous requests should be deterministic topics that resolve without a generative call. Reserve the model for the genuinely open-ended questions.
The first move you can make without involving anyone: open the quotas and service limits page alongside a list of the agent use cases you have in mind, and mark each one against the session limit and the request limits specifically. Anything a single person would plausibly do more than ten times a day, or that writes to records in bulk, is not an agent feature — it is an orchestration job wearing an agent's clothes.
If you would rather have that assessed than guess at it, we will model your expected credit consumption against the AU$299.30 pack price, map which parts of your workflows belong in Copilot Studio and which should sit in n8n or Retool, and tell you plainly whether the result is worth building — including the cases where it is not. Book an assessment and you get a scoped architecture, a monthly running-cost number you can take to a budget meeting, and a delivery timeline measured in weeks.