Copilot Studio & agents

Your first Copilot Studio agent: pick one boring task and know what it costs

By Matteo Garoli 7 min read

A white outline friendly bot head with two sparkles on a frosted glass tile, floating over soft amber and ice blue light on a near-black background.

The first agents I see go wrong mostly go wrong on ordinary things. The task was too big, two documents disagreed, nobody owned the agent, or the first bill surprised someone.

I work as a Microsoft 365 Copilot Cloud Solution Architect, and I build agents. The billing details below come from Microsoft’s pages as they stand on 22 October 2026. Microsoft’s licensing guide logs regular changes since September 2025, so check the date on any page you rely on.

Which task should your first Copilot Studio agent take on?

Pick a task that one person repeats every week, that follows the same steps each time, and where a human still reads the result before it goes anywhere. A wrong answer should cost a colleague a minute of fixing, not a wrong payment. Boring helps, because you can tell when the output is right.

An internal agent I built helps people write performance reviews. The writer dictates, the agent rewrites the text, and the writer pastes it into the form. The agent connects to no other system and never suggests a rating. In an early test, a whole review took 1 hour 40 minutes, against the half day to a full day it usually takes. That is one test, so treat it as a first signal.

What should be in place before you build?

One source of truth, one owner, and one person who decides when the agent is good enough.

An agent quotes whichever document it finds, so if the current procedure sits in two places it may use either. Keep facts in the knowledge sources and rules in the instructions. Microsoft’s declarative-agent page warns against moving instructions into SharePoint files to get past the 8,000-character limit, because anyone who can edit the file can then change what the agent does. Copilot Studio’s quotas page lists the same limit, and I assume the warning applies there too.

The owner is a named person who updates the sources and reads the consumption report. The decider is whoever does the task today.

How do you write instructions an agent can follow?

Write every step as a goal, an action and a transition, give each step one job, and say plainly when the agent must stop and hand the work back to a person.

Microsoft’s guidance for declarative agents asks for that structure: the step’s goal, the action and tool to use, and the transition that says when to move on or finish. For Copilot Studio agents, Microsoft adds that instructions work much like code: if a long set stops working, add them back one at a time, testing after each.

Three rounds of patching the prompt each fixed one problem and caused another. I rewrote it in Microsoft’s structure. Merging two actions in one message was one of three defects: when a step both showed the overview and opened the matrix, the next dictation landed in the wrong place. One step per message fixed that one. The rewrite then brought three smaller formatting problems, which we fixed afterwards.

The stopping rule matters most. The people team wanted a person in charge of every judgement, so when the writer asks the agent to complete the remaining sections itself, it must refuse and ask a question instead. An early version filled in four sections on its own. We now rerun that test at every publication and platform change. Microsoft’s declarative-agent page also warns that automatic model updates can change how an agent follows step-by-step instructions. I assume the same applies in Copilot Studio, so I rerun the tests even when I’ve touched nothing.

How do you test an agent with the people who will use it?

Give it to two or three of them, on the channel they will really use, with their own work, and write down what they say before you change anything.

After the first round of feedback, four of six comments went into the prompt, one was reinforced and one was postponed. The sponsor signed off only after a trial run of their own; a preview from me did not count. We chose the Microsoft Copilot channel over Teams because it gave users the dictation microphone they needed. The agent is now in use.

How is a Copilot Studio agent billed today?

In Copilot Credits. Microsoft replaced messages with credits on 1 September 2025, and each answer, action or grounding step now costs a fixed number of them. You pay from a monthly prepaid pack, through pay-as-you-go on an Azure subscription, or both.

For agents on the standard harness, Microsoft’s billing page (updated 15 October 2026) lists 1 credit for a classic answer, 2 for a generative answer, 5 for an agent action and 10 for tenant graph grounding, which is optional for each agent. One grounded answer can cost 12 credits.

A pack holds 25,000 credits a month, and unused credits expire at month end. Pay-as-you-go has no upfront commitment and is billed in arrears. Microsoft’s October 2026 licensing guide lists $200 per pack per month, billed annually, and $0.01 per credit on pay-as-you-go, in US dollars and subject to change. Resellers set their own prices, so confirm the price and billing term with yours. Credits normally reach the pay-as-you-go meter only after the packs are used up, and an environment can be set to bill its overage there once its own allocation runs out.

A Microsoft 365 Copilot licence (Microsoft’s pages now call it Microsoft Copilot) includes employee-facing use of Copilot Studio agents: no credits are charged when the agent runs under the licensed person’s identity, within fair usage limits. Some features, such as computer-using agents, aren’t included. A colleague without the licence draws credits from the same agent. Microsoft’s page on assigning user licences says people who use a published agent need no special licence.

What should you check before you publish?

Estimate the monthly credits, find out who holds a licence, set a limit on the agent and confirm which kind of agent you are building.

Illustrative numbers of my own: forty people each ask five questions on 22 working days, which makes 4,400 answers. At 2 credits each that is 8,800 credits, about a third of one pack. Switch on tenant graph grounding and it becomes 52,800 credits, just over two packs. If all forty hold a licence and use the agent inside Microsoft’s channels, that usage is included.

Set a limit. Packs are shared across the tenant, and Microsoft’s billing page says that at 125% of prepaid capacity custom agents are disabled and new conversations are rejected until capacity is increased or reset, so on packs alone a runaway second agent could switch off the first. Its own example shows an environment billed to pay-as-you-go is unaffected, so where pay-as-you-go covers the overage no hard stop arrives and the per-agent limit is the brake. An admin sets that monthly limit in the Power Platform admin center, with notifications and a hard stop, whether the environment uses packs or pay-as-you-go. Limits aren’t applied instantly, so leave some margin.

Agents in Copilot Chat and agents in Copilot Studio can draw on different capacity, so find out which route yours takes before you buy.

Microsoft now documents billing per agent runtime, which it calls a harness; Microsoft Learn lists four. The rates above are for the standard harness, where billing starts after you publish. The GitHub Copilot harness, a Copilot Studio framework and not the GitHub Copilot service, charges credits from the moment you start building.

Send me your shortlist and your credit estimate, and I will tell you where I would start.

Sources

If you are working on Copilot adoption, a training or a talk, or you are starting out in this world, write to me.