Copilot Studio vs custom AI agents for New Zealand businesses.

How do Copilot Studio and custom agents compare?

Copilot Studio is Microsoft’s low-code tool for building agents that run on Microsoft’s platform and connectors. A custom agent is software your team or a partner writes against model APIs from OpenAI, Anthropic or Google, with your own hosting, tools and logs.

Where is each one stronger?

Copilot Studio is stronger inside a Microsoft 365 organisation. An HR policy agent or an IT help agent can be built by a capable staff member, sits in Teams, respects existing permissions, and costs nothing in credits for staff who hold Copilot licences. Agents can also call tools from MCP servers. See our Copilot Studio guide.

Custom agents are stronger when the agent serves customers at volume, acts across several systems or must meet a specific processing requirement. Since 26 March 2026, Amazon Bedrock’s Australian profile can keep Claude inference within Australia and New Zealand, which no Microsoft Copilot option promises. You also choose and pin the model, so a model change happens when you decide. Our AI agent development page describes how we build them.

How do you test both on the same agent?

Build the narrowest useful version in each and run the same cases. Use our AI agent acceptance checklist as the pass mark.

1. Internal policy agent. Collect 40 real staff questions with answers you have checked against your policies in SharePoint. Score correct answers, whether it names the right policy, and whether it says “ask HR” when the policies are silent.
2. External enquiry agent. Replay 50 past website enquiries. Score answers, count credits or tokens per conversation, and include three messages that try to make the agent ignore its instructions.
3. One gated action. Have the agent create a job in ServiceM8 or a ticket in your helpdesk, but only after a person approves. Check that the log shows what was proposed, who approved it and what was written.

What do they cost over a year?

Worked examples at prices listed on 19 October 2026, in US dollars, before GST and any partner fees. Copilot Studio credit rates come from Microsoft’s billing rates page.

Customer-facing enquiry agent: 3,000 conversations a month, four answers each, so 12,000 answers.

• Copilot Studio, answers from public website content: 12,000 × 2 credits = 24,000 credits. Pay-as-you-go is US$240 a month; one capacity pack at US$200 a month covers it, or US$2,400 a year.
• Copilot Studio, answers grounded in tenant data: 12,000 × 12 credits (Microsoft’s own example) = 144,000 credits, or US$1,440 a month pay-as-you-go.
• Custom agent on Claude Sonnet 5.5: assume 6,000 input and 400 output tokens per answer. That is 6,000 × US$2 ÷ 1,000,000 + 400 × US$10 ÷ 1,000,000 = US$0.016 an answer, or US$192 a month. On Haiku 5.5 it is US$0.0008 an answer, or US$9.60 a month, if quality holds in your tests.

Model fees are only part of a custom agent. With build cost B, monthly hosting and maintenance M, custom running cost C and Copilot Studio spend S, the build pays back in B ÷ (S − C − M) months. If S − C − M is zero or negative, a custom agent never pays back on cost and is only worth it for control, region or capability.

Internal agent for unlicensed staff: 40 people on free Copilot Chat asking 25 questions a month each, grounded in SharePoint, is 1,000 × 12 = 12,000 credits, or US$120 a month metered. Staff holding Copilot licences would not be charged credits, but each licence costs NZ$34.00 a month excluding tax.

Later costs on both sides:

• Platform changes. Microsoft stops seeding AI Builder credits on new or renewed licences from 1 November 2026. On the custom side, OpenAI shut down its Assistants API on 26 August 2026, and Anthropic retires Claude Sonnet 4.5 from its API on 30 November 2026. Budget for migrations; see AI model migration.
• Reasoning models. Copilot Studio adds 10 credits per 1,000 tokens for reasoning models; custom agents pay more per token on larger models. Measure before choosing one.

Which should you choose?

Choose Copilot Studio if

• You are a Microsoft 365 organisation and the agent serves your own staff.
• Most users already hold Copilot licences.
• A capable staff member or partner will maintain it without engineers.
• Standard connectors and MCP tools reach the systems you need.

Choose a custom agent if

• The agent serves customers at volume, where per-answer credits add up.
• You need a specific model, or inference kept in Australia and New Zealand.
• It must act across several systems with approval gates and full audit logs.
• You are not on Microsoft 365, or the agent is core to a product you sell.

Where do both fall short?

• Stale knowledge. An agent grounded in an outdated policy library answers confidently from it. Assign an owner to the sources.
• Too much access. NCSC NZ’s agentic AI guidance, co-authored with Five Eyes partners, says never to give agents broad or unrestricted access and to put human approval on high-impact actions. Our guide to agent permissions and approvals turns that into design rules.
• No evaluation. Without a fixed test set, nobody notices when answers degrade. See AI agent evaluations.
• Offshore processing. Microsoft says customers outside the EU may have Copilot queries processed in the US, the EU or other regions, and a custom agent on a first-party API processes offshore too unless you choose a regional route.

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