Copilot Studio pricing for NZ firms, 2026

How does the credit model work?

Until September 2025, Microsoft called this unit a message. It's now a Copilot Credit, and the rename matters less than what didn't change underneath it: the pricing is still consumption-based, and it's still pooled at the tenant level rather than handed out per user. Your business holds a stock of credits and every agent, in every environment, draws from the same pool.

You get two ways to fill that pool. Buy a prepaid capacity pack, 25,000 Copilot Credits for $200 USD a month, which works out to $0.008 a credit. Or run pay-as-you-go through an Azure meter at $0.01 a credit, with no upfront commitment, billed on what you used. Pay-as-you-go costs about a quarter more per credit than the pack. That premium buys you the option of not having to guess your volume before you've built anything.

The part that trips people up is that credits are not all worth the same. A scripted reply costs about one. A generative answer costs about two. Grounding that answer in your own SharePoint or Dataverse adds roughly ten on top, and an autonomous action costs more again. They stack inside a single exchange, so one question can cost twenty times what the cheapest one does. That is the difference between a $200 month and a $2,000 month, and it is worth understanding before you build rather than after. I break down exactly what counts as a billable interaction, and where a real deployment burns through an allowance fastest, in the credit meter breakdown.

How much would a small internal support agent cost?

Say you're building an agent for a 25-person team, answering questions about one internal system. Not everyone will use it daily. Assume 15 of the 25 actually ask it something, and each of those 15 asks around 4 questions a week, across 48 working weeks a year.

15 × 4 × 48 = 2,880 questions a year, or 240 a month. That's your interaction count. Now weight it by type. A realistic internal support agent probably runs something like 40% simple scripted lookups (1 credit), 50% generative answers composed from documentation (2 credits), and 10% harder questions needing tenant-wide grounding (call that 12 credits, stacking a generative answer with the grounding cost).

(0.40 × 1) + (0.50 × 2) + (0.10 × 12) = 0.4 + 1.0 + 1.2 = 2.6 credits per question, on average.

240 questions × 2.6 credits = 624 credits a month. At pay-as-you-go pricing, that's $6.24 USD. Push the mix harder: say the agent starts resolving some tickets itself and 15% of interactions become autonomous actions at a floor of 25 credits each. The weighted average climbs to about 5.6 credits a question, and the monthly bill to roughly $75, and that is a floor rather than a forecast. For a 25-person internal tool, that's still nowhere near the $200 capacity pack. Pay-as-you-go is the right call here and the bill barely moves.

Re-run this with your own numbers: your headcount, your realistic weekly question count, your best guess at the mix of simple, generative, grounded and autonomous. Multiply through the same way. The arithmetic doesn't change, only your inputs do. Microsoft's per-credit price is the one number in this whole exercise you should go and check before you rely on it.

Which three licences are easy to confuse?

Copilot Studio credits are only one of three things Microsoft sells you under a name that starts with Copilot or Power.

Power Platform premium licensing is the oldest of the three and has nothing to do with AI. Power Apps Premium (around $20 USD a user a month, as at July 2026) and Power Automate Premium (around $15 USD a user a month) give the people who hold them Dataverse, premium connectors and unlimited custom apps and flows. This is the plumbing layer: the app your staff open, the flow that moves data between systems. No credits involved anywhere in it.

Microsoft 365 Copilot is the per-seat assistant that lives inside Word, Excel, Outlook and Teams. Microsoft's NZ list price for the Business add-on is NZ$34 a user a month, paid yearly, on top of an existing Microsoft 365 subscription (as at 1 October 2026). It's licensed by headcount, not usage. I've gone through where this specifically gets confused with Copilot Studio in Copilot and a Copilot Studio agent are not the same thing, and whether the per-seat price pays for itself is worked through seat by seat in is Copilot worth it for a small NZ firm.

Copilot Studio sits apart from both: the tenant-wide credit pool this whole post is about. There's one genuine interaction between it and M365 Copilot worth knowing. If a staff member already holds an M365 Copilot seat, their use of an agent inside Copilot Chat, Teams or SharePoint for classic answers, generative answers or tenant-Graph grounding draws on their seat rather than the shared credit pool. Autonomous actions don't get that break, and neither does anything reaching an external customer instead of an internal staff member. It's a real saving where it applies. Don't buy Copilot seats to get it.

A business can hold all three licences, two, one, or none. Owning Power Automate Premium tells you nothing about your Copilot Studio bill, and owning M365 Copilot seats tells you nothing about either of the other two. Budget them separately, because Microsoft sells them separately.

When can a plain Power Automate flow do the same job for less?

A lot of what gets pitched as an AI agent is a lookup, a routing decision or an approval step, and Power Automate has done all three for years without touching a credit meter.

The test I use: can you write the logic as an if/then flowchart without the system needing to interpret free text or judge intent? "If a ticket's status equals X, reply with Y." "If a form total is over $500, route to a manager." "If an invoice lands in this inbox, pull the fields and push them to Xero." None of that needs a language model. It needs a trigger, a condition and an action, and Power Automate Premium covers unlimited flows like that for a flat licence instead of a bill that moves with usage.

Go back to the 240-question support agent above. In practice, a good chunk of "support questions" are really status checks: where's my ticket, has this been approved, what version am I on. Route those through a flow that reads a record and replies directly, and they never touch the agent at all. What's left for Copilot Studio is the harder half, the open-ended questions that need interpreting. The credit bill drops because the interaction count drops, and the agent gets to be good at fewer, harder things.

There's a second reason to make this split, beyond cost. A flow either runs its logic correctly or it doesn't. There's no ambiguity to manage. An agent answering a status question in natural language can occasionally get creative with something that should have been a plain database read. Keep the deterministic work deterministic and save the credits for the parts that need judgement.

How are costs estimated before either option is recommended?

This is most of what a Workflow Review is, on the Power Platform side of my work. I pull two to four weeks of real questions or tickets, the actual ones your team fielded, not a guess at what people might ask. I sort them into two piles: things a flow can answer with a lookup, and things that genuinely need an agent to interpret free text or judge intent.

Then I model both. What a flow-only build costs to run, mostly a Power Automate Premium licence, against what the agent side would cost in Copilot Credits, using the same working as the estimate above but with your real volume and your real mix instead of assumed numbers. Usually a flow covers most of it for a fraction of the cost and the agent earns the hard remainder. Occasionally the questions are open-ended enough that only an agent does the job. Either way you get the number first.

If it needs more than a call can cover, a paid Workflow Review gets you a written model of both options, credits and all, before anyone signs anything. And if you want the full picture of where Copilot Studio fits alongside the rest of the Power Platform, that's on the services page. For what the prepaid tenant pack itself buys before you get to any of this arithmetic, see Copilot Studio capacity: what 25,000 messages buys. And once the agent works, where it actually gets published, Teams, the web, M365 Copilot or elsewhere, changes both its reach and its credit draw, covered in Copilot Studio channels: Teams, M365, web and more.