The work worth automating is repetitive, knowledge-heavy and rule-based. Answering the same question from a document set. Comparing a new policy against the old one. Troubleshooting a known system. I've built Copilot Studio agents for exactly that, turning a one-person spreadsheet into a shared knowledge base and cutting a monthly policy review down to minutes.
The trick is a narrow agent plus a couple of flows aimed at one painful task, built well. Two real examples, kept anonymous.
Five jobs, and what each costs to run
Four of these five jobs cost under NZ$120 a month to run, so the build is the real cost. The table models standard-harness agents and uses the credit rates on this site as at 28 October 2026 (a scripted answer about 1 credit, a generative one about 2, grounding in your own data about 10 more, a connector or agent action about 5, so an autonomous run of four actions about 20) at USD 0.01 a credit pay-as-you-go, converted at 1.72, not a Microsoft NZ price. The volumes are my assumptions, so replace them with yours.
| Job | Where it fits in an NZ firm | Assumed volume | Credits per answer | Credits a month | Per month, NZD |
|---|---|---|---|---|---|
| Monthly policy or rule comparison | Accounting firm reading an IRD or industry change | 20 runs | 12 (generative plus grounding) | 240 | NZ$4.13 |
| Shared knowledge agent replacing a spreadsheet | Office of 25 asking about one internal system | 240 questions | 2.6 on average | 624 | NZ$10.73 |
| Staff policy and HR questions | Builder or clinic with a handbook and leave rules | 300 questions | 12 | 3,600 | NZ$61.92 |
| Support triage that looks up a record | Firm with a line-of-business Power App | 400 questions | 17 (2 + 10 + 5) | 6,800 | NZ$116.96 |
| Autonomous inbox triage | Property manager or trades office sorting shared mail | 500 triggers | 20 (4 × 5) | 10,000 | NZ$172.00 |
These rates are for agents on the standard harness. Agents on the GitHub Copilot harness, which Microsoft made generally available in August 2026, are billed on usage from the moment you start building, including preview tests and evaluations, and a Microsoft 365 Copilot licence does not cover that usage. These weights do not predict GitHub Copilot harness costs.
The USD 200 prepaid pack of 25,000 credits only beats pay-as-you-go once you use more than 20,000 credits a month, and none of the five jobs gets there alone. How the meter works is in what counts as a Copilot Credit, and the full price model is in Copilot Studio pricing in NZ. If the job is a fixed sequence with no questions in it, a flow does it for less, which Copilot Studio or Power Automate works through. The largest of these builds, the support agent, has its own write-up in the Copilot Studio support agent.
Example one: the answer that lives in one person's head
Some questions only come up every few months. Someone dealt with it last time, wrote down what they sent, and moved on. Then it comes up again half a year later and the only way to answer it is to go fishing through old emails to find what was said the last time.
In one team, this knowledge was being kept in an Excel spreadsheet that one person filled in by hand and referred back to each time. Two problems with that. It's slow, and worse, it's a one-person knowledge base sitting in one person's OneDrive. The day that person leaves, the knowledge walks out with them.
The fix was to move the knowledge out of the spreadsheet and into a shared SharePoint source, then point Copilot agents at it to answer questions directly, retrieval-style. Now anyone can ask and get the answer in seconds, and the knowledge belongs to the team instead of one inbox.
Example two: the monthly policy check nobody enjoys
Some teams have to stay on top of rules that change without warning. A regulatory or government policy gets updated, and they have to work out what changed and what they now need to action against their own contracts and obligations. Done by hand, that's a monthly slog of reading, comparing, and writing up.
So I built an agent for it. When the new version comes in, you drop it straight into the chat. The agent already holds the previous policy, so it runs the comparison, does an online check in case there's anything it missed, and produces a change-and-action document in the exact format the team uses. A monthly manual review turned into something that happens in seconds, with a clean document at the end that someone just has to action.
A related build sits at bigger scale: the Copilot Studio support agent I built to troubleshoot a complex internal system cut bug-fixing time by about 80%. Same pattern, larger surface.
Where it works, and where it doesn't
The limits matter, because an agent in the wrong place just adds noise.
It works when the task is repetitive, knowledge-heavy, and rule-based, and the knowledge can be pulled into one place. Retrieval, comparison, triage, answering from a known source. That's the sweet spot.
It doesn't work when the task needs real human judgement, when the inputs are all over the place, or when no one will own keeping the agent's knowledge current. An agent fed stale or messy information is worse than the manual process it replaced, because people stop trusting it.
If the phrase Model Context Protocol keeps turning up while you research this, it is worth ten minutes: MCP explained without the protocol talk covers what it is and whether it applies to you yet.
What good looks like
- Scope it narrow. One painful job that works every time beats a do-everything assistant that does nothing well.
- Centralise the knowledge, then keep it live. Get it out of the spreadsheet and the personal inbox, and make sure it updates as things change.
- Pin down the output. If the agent produces a document, decide the format up front so the result is ready to use, not ready to reformat.
If your team has a job like this
This is the kind of thing I build for teams. If there's a repetitive, knowledge-heavy task eating hours in your week, a focused Copilot Studio agent can probably take it off you. Work out the running cost as well as the build cost before you commit: how Copilot Credits are consumed covers the meter. Book a free 30-minute call to find the best one to start with, or to hear that it's not worth it yet.