There is plenty written about how to build a Copilot Studio agent, and very little written about what one actually costs to run each month. That gap matters, because the build is a one-off and the bill is forever. Most Australian buyers we work with can describe the agent they want in a sentence, but cannot tell you whether it will cost $200 or $2,000 a month to operate. This guide closes that gap.
We will walk through the two ways Microsoft charges for Copilot Studio consumption, the credit multipliers that quietly inflate the bill, and a simple method for estimating a chatbot's monthly cost before you commit. All figures are USD list unless stated, because Microsoft prices Copilot Studio capacity in US dollars; we have flagged where the Australian dollar and GST change the picture.
The thing nobody budgets for: message consumption
When people say "Copilot Studio pricing" they usually mean the licence. But the licence is the cheap part. The real running cost is consumption — every time your agent answers a question, takes an action, or reaches into your data, it burns capacity. Microsoft now meters this in Copilot Credits (the model that replaced the older flat "message" counting), and a single user question can consume anywhere from 1 to well over 20 credits depending on what the agent does to answer it.
This is the part finance never sees in the demo. The agent looks free because, during a pilot, volume is tiny. Scale it to a customer-facing chatbot handling thousands of conversations and the consumption line becomes the dominant cost. Budgeting for the build and ignoring the meter is the single most common mistake we see.
Option one: the 25,000-credit capacity pack
The headline commitment option is a prepaid capacity pack: 25,000 Copilot Credits for an indicative US$200 per pack per month (indicative USD list — confirm at purchase). You can buy multiple packs and they pool at the tenant level, so one purchase covers every agent built across your Microsoft tenant rather than being tied to a single bot.
In Australian dollars, US$200 lands somewhere in the low-to-mid $300s AUD ex GST at typical 2026 exchange rates (indicative AUD list — confirm at purchase, as Microsoft bills in USD and the converted figure moves with the rate). The practical point: a pack is a predictable, fixed monthly block of capacity, which makes it the right choice once you know your volume.
Option two: pay-as-you-go at a cent a message
The alternative is a consumption meter billed to an Azure subscription at roughly US$0.01 per message-equivalent credit (indicative USD list — confirm at purchase). You pay only for what you consume, settled at the end of the monthly billing period, with no upfront commitment.
Pay-as-you-go is the sensible default for pilots and anything with unpredictable volume. You are not guessing at pack sizing, and you are not paying for capacity you might not touch. The trade-off is that at steady, high volume the per-credit meter usually works out dearer than buying packs — so the smart pattern is to start on pay-as-you-go, watch the actual consumption for a month or two, then switch to packs once the run rate is known.
The multiplier trap: not every answer costs the same
Here is where naive estimates fall apart. People assume one question equals one message. It does not. Copilot Studio charges by what the agent does to produce the answer, and those costs stack inside a single response.
- A classic answer (a static, manually authored reply) is the cheapest and can be free for licensed users.
- A generative answer — the AI writing a response — consumes 2 credits.
- Grounding the answer in your Microsoft 365 tenant graph adds around 10 credits on top.
- Each action the agent calls (a connector, a flow, an API call) adds roughly 5 credits.
Stack those together and a single "smart" interaction gets expensive fast. A generative answer that grounds in your tenant data and calls two actions is 2 + 10 + 5 + 5 = 22 credits for one user question. At that rate, a 25,000-credit pack is exhausted by a little over 1,100 such interactions — not 25,000. That is the trap: the more useful and connected you make the agent, the more each answer costs.
How to estimate a chatbot's monthly bill
You do not need a spreadsheet model with twenty tabs. You need four numbers and a multiplication. Here is the method we actually use with clients.
- 1Estimate monthly conversations — how many distinct user sessions the agent will handle.
- 2Estimate turns per conversation — most support-style chats run 4 to 8 back-and-forth turns.
- 3Estimate credits per turn — be honest about the architecture. A plain generative agent is ~2 credits a turn; one grounded in your data with actions is realistically 12 to 22.
- 4Multiply the three, then divide by 25,000 to see how many packs you need (or multiply by ~US$0.01 to price it pay-as-you-go).
Worked example: 2,000 conversations a month, 5 turns each, at 12 credits a turn is 2,000 × 5 × 12 = 120,000 credits. That is roughly five capacity packs, an indicative US$1,000 per month (indicative USD list — confirm at purchase), or close to $1,600–$1,700 AUD ex GST depending on the exchange rate (indicative AUD list — confirm at purchase). If your real architecture is leaner — fewer actions, lighter grounding — the number drops sharply, which is the whole point of estimating before you build.
What we'd actually do
Our standard approach is deliberately boring, because boring keeps bills predictable. We start every new agent on pay-as-you-go so the first month's consumption is measured rather than guessed. We instrument the agent to log credits per interaction from day one, so the multiplier reality is visible instead of a month-end surprise. Then, once a fortnight of real traffic shows the run rate, we either commit to the right number of capacity packs or keep it on the meter if volume stays low and spiky.
We also design for cost, not just capability. If a frequently asked question can be answered with a classic, authored response instead of a generative one grounded in the tenant graph, that is the difference between near-zero and a dozen credits every single time it is asked. On a high-traffic chatbot, choosing the cheaper answer path for the top twenty questions can halve the bill without any user noticing. The goal is an agent that is genuinely useful where it counts and quietly frugal everywhere else.
Finally, a note on the numbers in this article. Microsoft adjusts Copilot Studio pricing and credit weightings periodically, and the figures here are USD list at the time of writing. Before you sign anything, confirm the current per-pack price, the live AUD conversion, and GST treatment for your purchasing path — buying through an Australian partner or reseller can change both the price and how it appears on your invoice.