In a small business that runs on Microsoft 365, the question always comes up at the same point in a conversation about AI agents: why not just build it in Copilot Studio? It's a fair question. Microsoft has put agent-building tools inside the Copilot license, and for some needs they're enough.
To answer it honestly, you have to read Microsoft's billing table rather than its feature list. As long as the agent serves an employee who holds a Copilot license and stays inside Microsoft 365, it costs almost nothing extra. The moment it serves someone else, runs on its own, or reasons through several steps, a meter starts. That pricing line maps almost exactly onto the line between the two kinds of projects.
The short answer, for the busy
Microsoft offers two ways to build an agent. Agent Builder, inside the Microsoft 365 Copilot app, lets you set up a question-and-answer agent in plain language for yourself or a small team, drawing on SharePoint, email or files. Copilot Studio is the full portal: multi-step flows, connectors, publishing to a website, governance through separate environments. With a Microsoft 365 Copilot license (CAD $28.50 per user per month on an annual plan at the regular rate, for the Business edition), internal use by employees who hold that license is included. Outside that frame, everything is paid in Copilot Credits: CAD $271.40 a month for a pack of 25,000 credits, or about one US cent per credit pay-as-you-go. A Microsoft agent is a very good fit for answering employees from your own documents. A custom agent becomes the better call when the agent serves your customers, touches software with no connector, or handles a volume that makes the credit bill unpredictable.
Several tools under one name
"Copilot Studio" means several things, and Microsoft renames them as it goes.
Agent Builder lives in the Microsoft 365 Copilot app and in Teams. You describe the agent in plain language, give it sources (a SharePoint site, files, email) and share it. The pricing page also calls it Copilot Studio Lite.
Copilot Studio, the full portal, is a Power Platform product. It targets a department, the whole organization or external customers, with multi-step flows, integrations and separate development, test and production environments.
Copilot Studio now runs agents on three "harnesses", which are different execution engines. The Copilot Chat harness extends the assistant with your internal knowledge. The standard harness follows the topics and paths you define. The GitHub Copilot harness, the most capable, takes a goal, breaks it into steps, works with Word, Excel and PDF files, and recovers when a step fails.
What the Microsoft agent does well
The ideal case: an employee asks a question, and the answer is somewhere in your Microsoft 365 documents. An onboarding agent wired to HR's SharePoint site, or an agent that answers questions about your internal procedures.
For that job, an outside vendor will have a hard time beating Microsoft. The agent creates no new privileges: it respects existing permissions and only shows an employee what they could already open, under the audit logs and retention policies already configured in your tenant. And it costs nothing extra when the user holds a Microsoft 365 Copilot license.
In that case, with the license already paid for, keep your money: clean up your SharePoint sharing, build the agent in Agent Builder, and see whether people use it. We compared the subscriptions themselves in our piece on Copilot, ChatGPT and Claude; here we assume the license is already bought.
The line that matters: who holds the license
The "free" part comes with conditions, written in the fine print of the billing documentation. Usage is included when three things are true at once: the person using the agent is an employee, they hold a Microsoft 365 Copilot license, and the agent acts under their authenticated identity. For flows the agent triggers, the inclusion only covers the "When an agent calls the flow" trigger; a flow that starts on its own, say when an order comes in, consumes credits at the standard rate. And all of it is subject to fair-use limits that Microsoft reserves the right to adjust.
So three situations push a project onto the meter. The agent serves people without a Copilot license: your customers, but also the shop-floor staff you never bought a license for. The agent is published outside Microsoft 365, on your website or in an app: the pricing page reserves that for the standalone Copilot Studio license. Or the agent runs on the GitHub Copilot harness, the one built for long processes: Microsoft says it requires credits even for licensed users, and that it bills from the moment you start building, testing and evaluations included, whereas the standard harness only bills after publishing.
Put differently, the agents most useful to a small business, the ones that answer customers or move a process forward without being asked, are exactly the ones the license doesn't cover.
The credit meter, in numbers
The unit has been called the Copilot Credit since September 1, 2025 (before that, the "message"). There are three ways to buy them. The prepaid pack: 25,000 credits a month for CAD $271.40 on Microsoft's Canadian page. Pay-as-you-go through an Azure subscription: USD $0.01 per credit on the Azure pricing page. And an annual pre-purchase plan, with up to 20% off.
Every move the agent makes has a price. A pre-written answer costs 1 credit, a generated answer 2, an agent action 5 and 100 flow actions 13. A search across your whole Microsoft 365 data costs 10 on its own. Reasoning models add 10 credits per 1,000 tokens.
Microsoft gives a telling example of its own: a support agent on a website, 900 customers a day, four canned answers and two generated answers per conversation, for 7,200 credits a day. Over 30 days, that's 216,000 credits. In packs, you need nine, about CAD $2,443 a month. Pay-as-you-go, USD $2,160. And that agent only answers questions, without taking a single action in your systems.
Two rules are worth knowing before you sign. Unused credits don't roll over to the next month. And when consumption reaches 125% of prepaid capacity, Microsoft disables custom agents: your customers then see a message like "This agent is currently unavailable. It has reached its usage limit." Pay-as-you-go billing avoids the cutoff, and a monthly per-agent cap limits what any single agent can burn through, provided someone sets them up.
Where the Microsoft agent tops out
Systems outside Microsoft. Copilot Studio shares more than 1,000 prebuilt connectors with Power Automate, and it accepts MCP servers. For software that isn't on the list, Microsoft offers the custom connector, which it defines as a wrapper around a REST API. That's the catch: if your accounting software, your manufacturing ERP or your clinic management system doesn't expose an API, no connector is possible, and someone has to build the bridge.
Multi-step processes. The standard harness follows the paths someone drew. To judge, follow up and handle exceptions, you need the GitHub Copilot harness, billed from the first minute of building. You're back to the questions from our no-code vs. custom comparison: who builds it and who fixes it, now with a meter running.
Where the computing happens. According to the Power Platform documentation, for an environment hosted in Canada, the Azure OpenAI service behind the generative features is hosted in the United States, and several features require an admin to tick "Move data across regions". In practice, that changes little: section 17 of Quebec's Law 25 covers any communication "outside Québec", so Toronto counts as much as the United States. As soon as personal information leaves, you need a privacy impact assessment and a written agreement whether Microsoft or we built the agent. The real question is which data goes to the model, which is where anonymizing before the AI earns its keep.
When custom is the better call
Stay with Microsoft if the agent answers licensed employees, from Microsoft 365 documents, in Teams or Copilot. That's the ground where the tool is good, already paid for and governed by your admin.
Look at a custom agent in four cases. The agent serves your customers or unlicensed people, and the volume makes the credit bill hard to predict. The process touches software outside Microsoft, especially software with no public API. You want to choose the model, control what gets sent to it and strip personal information before the call. Or your team doesn't have the Copilot license: at CAD $28.50 per person per month at the regular rate (CAD $24.43 for the first year under a promotion that runs until December 31, 2026), buying it for 30 employees just to power one agent comes to CAD $855 a month before the first credit.
A custom agent costs money too, obviously. Ours is a fixed monthly subscription with no upfront development fee, quoted after a first call, and we break down what drives the price in our article on what an AI agent costs. What differs is the shape of the price: a fixed rate you know in advance on one side, a meter that follows your traffic on the other.
Often the two coexist. Agent Builder for internal questions, a custom agent for the process that brings in the money.
Write to us with what the agent needs to do, for whom and in which software. We'll tell you plainly whether Copilot Studio is enough, including when it is.
Prices, product names and billing rules were read on Microsoft's official pages on September 23, 2026; Microsoft changes them often. This article also explains a Law 25 obligation in plain language and is not legal advice.
Written byXavier Peich