Copilot Studio for New Zealand businesses.

What Copilot Studio is, and when NZ organisations use it

Copilot Studio is Microsoft’s builder for AI agents: assistants that answer from chosen knowledge sources, follow authored conversation topics, call actions and flows, and appear in Teams, Microsoft Copilot or on a website. It is aimed at organisations already on Microsoft 365 that want something more specific than general Copilot Chat: an HR policy helper, an IT help desk, a sales agent grounded in CRM data, or a customer-facing agent on the website.

Two licensing facts shape every decision. Microsoft says Microsoft Copilot (formerly Microsoft 365 Copilot) includes agent building for use within your organisation, but publishing to outside channels needs a standalone Copilot Studio plan. And agents used by staff with a Microsoft Copilot licence, under their own identity, carry no Copilot Credit charge. Everyone else is metered.

What it is good and bad at

Copilot Studio is good when the knowledge and the users are already inside Microsoft 365. Pointing an agent at a SharePoint site, publishing it to Teams and controlling it with existing Entra ID groups takes days rather than months, and the admin, security and compliance tooling is the same as the rest of your tenant. Agents can also use tools from MCP servers, and Microsoft runs a preview certification for them; our guide to MCP for business systems covers when that route makes sense.

It is weaker when:

• Costs depend on design. The same traffic can cost 1 or 12 credits a message, so a careless build is expensive.
• The core systems are not Microsoft. Xero, ServiceM8 or a bespoke job system need connectors, flows or MCP servers, and Microsoft says data sent to non-Microsoft systems is the agent maker’s responsibility.
• Behaviour must be precise. Generative answers vary. Use authored “classic” answers where wording matters.

What Copilot Studio costs

Microsoft’s NZ pricing page showed “NZ$323.60 license” where the US page showed “$200.00 license”. That probably is the NZD capacity pack price, but the page does not say so; confirm with Microsoft or a partner. Copilot Chat users without a licence can use your custom agents only on a metered basis.

The billing rates per event:

Illustrative example (USD, no conversion): a Wellington retailer’s website agent handles 200 conversations a day. A typical conversation uses two classic answers and three generative answers: (2 × 1) + (3 × 2) = 8 credits. That is 1,600 credits a day, or 48,000 in a 30-day month. Pay-as-you-go costs US$480. One capacity pack plus pay-as-you-go for the remaining 23,000 credits costs US$200 + US$230 = US$430. Ground the same agent in your tenant’s Graph and each message adds 10 credits, roughly tripling the bill. Microsoft’s agent usage estimator lets you test a design before building.

One licensing change to note: from 1 November 2026, new or renewed Power Platform licences get no seeded AI Builder credits, and agents and agent flows consume only Copilot Credits.

Where agent data is stored and processed

Copilot Studio stores customer data in the Azure geography of the environment, and Microsoft’s data location table lists Australia (Australia East and Australia Southeast) but not New Zealand. Microsoft’s note says that if a tenant’s location is not listed, data is stored in the United States. Some activity data, including conversation logs, is stored in the region of the user’s Exchange mailbox. Microsoft also says Copilot Studio can be configured to use generative AI features in other regions, including when capacity is constrained.

For the Privacy Act 2020, the difference between internal and external agents is the point. A staff agent mostly handles information you already hold. A website agent collects new personal information from the public, so IPP 1 limits it to what you need and IPP 3 expects people to know it is being collected, for what purpose, and by whom. Say clearly that the customer is talking to an AI agent, link your privacy statement, and do not ask for details the agent does not need.

Starter project: an internal staff help desk in Teams

This replaces a generic tutorial with an agent that costs no credits for licensed staff and teaches you how Copilot Studio behaves before anything faces the public.

1. Pick one SharePoint site. Choose HR or IT guidance that is current. Remove old versions and check the site’s permissions, because the agent can only be as tidy as its sources. More in AI and SharePoint.
2. Create the agent. Add the site as the knowledge source and paste the instructions below.
3. Author two classic topics. For “report a privacy breach” and “report a safety incident”, write fixed answers. Every agency must have a privacy officer under the Privacy Act, so name yours.
4. Test 20 real questions. Use the test pane and the activity map. Mark each answer correct, partly correct or wrong, and include three questions the site cannot answer.
5. Publish to Teams for a pilot group of licensed staff, and tell them it is an AI agent that can be wrong.
6. Review after two weeks. Read the conversation history, fix sources that caused wrong answers, and only then consider external use.

Agent name: Staff help desk (pilot)

Instructions:
You answer staff questions about [Organisation]'s HR, IT and health and safety guidance.
Use only the SharePoint knowledge source. If the answer is not there, say so and give the contact: [HR email] or [IT email].
Always name the document you used and link to it.
Use New Zealand English, plain language, under 150 words.
Never advise on an individual's pay, performance or employment situation; refer those questions to HR.
If someone mentions an injury or safety incident, tell them to follow the incident reporting process and contact their manager now.

Our lesson on testing a business AI assistant has a fuller test plan, and AI and Microsoft Teams covers publishing options.

Copilot Studio or a custom agent?

Choose Copilot Studio when users and knowledge are in Microsoft 365 and the agent mostly answers questions with a few actions. Choose a custom-built agent when you need fine control of model choice, evaluation, cost per conversation, or deep integration with non-Microsoft systems. Copilot Studio versus custom agents sets out the trade-off, and our Microsoft Copilot guide covers the licence itself.

Limits and failure modes

• Capacity runs out. When prepaid capacity is exhausted, agent flows stop starting new runs. Microsoft allows some overage as a grace period; do not rely on it.
• Grounding cost surprises. Tenant Graph grounding at 10 credits a message is the most common cause of a bill several times the estimate.
• Unlicensed users. An internal agent shared with staff who lack a Copilot licence is metered, not free.
• Connector reach. An agent with write actions into another system can do damage at speed. Start read-only; Xero’s own developer MCP server, for example, does not ask for confirmation before writing.

If the agent needs actions in Xero or other non-Microsoft systems with approvals and monitoring, that is where our AI agent development work fits.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.