Once a week, a few people meet for a short time. They look at how much the helpers cost, who used them, and whether anything odd happened. Then they decide one or two small things.
A short meeting every week, always with the same agenda, keeps AI spending under control. You look at spend against budget, who's run out, who isn't using their allowance, which agents cost more than expected, and you write down the decisions.
Run a timeboxed weekly cost review with a fixed agenda: spend versus forecast, limits reached, underused seats and allowances, agent anomalies and route mix, then decisions with owners. Prepare the pack automatically from admin centre exports so the meeting spends its time on decisions.
Why weekly, and why fixed
Monthly reviews are too slow for metered spend. By the time the monthly report arrives, a misbehaving agent has had four weeks to run. A weekly review catches drift while it's still small, and a fixed agenda keeps it short, because nobody has to work out what the meeting is for.
I've kept it to 20 minutes because a longer slot invites discussion that belongs elsewhere. If an item needs more than a few minutes, it gets an owner and leaves the meeting.
Who comes
The budget holder chairs it. The people who own production agents attend for their item. Someone from IT who can see the admin centres brings the figures. Finance can read the pack without attending, which Microsoft's reader roles for cost dashboards are designed to allow.
The agenda
| Minutes | Item | The question |
|---|---|---|
| 0 to 4 | Spend against forecast | Are we inside the likely case, and if not, why? |
| 4 to 8 | Limits reached | Who hit a per-user or per-agent limit, and what were they doing? |
| 8 to 11 | Underused allowance and seats | Who has a licence or allowance they barely use? |
| 11 to 16 | Agents | Any agent whose cost, volume or route mix moved? |
| 16 to 20 | Decisions | What are we changing, who owns it, and by when? |
The agents item looks at more than cost. For an agent that makes or supports decisions, I'd look at the split between cases it handled, cases it sent to a person and cases it stopped on, as in Act, ask a person, or stop. A cheaper week caused by the agent quietly approving more cases deserves a question.
The pack
The meeting only works if the figures are ready before it starts. I'd prepare a one-page pack from exports:
- Copilot Credit consumption by policy, user and service, from Cost Management in the Microsoft 365 admin center.
- Copilot Studio agent consumption from the Power Platform admin center.
- Azure costs for Foundry resources from Microsoft Cost Management.
- Active Copilot users from the Microsoft Copilot usage report.
Microsoft says the Copilot usage report becomes available within 72 hours of the end of the day it covers, and its Cost Management overview refreshes every 4 hours. So the Friday pack really describes the week to about Tuesday or Wednesday. I'd note that on the pack rather than pretend it's live.
A worked example of the decisions item
These are illustrative.
| Decision | Owner | By |
|---|---|---|
| Move 8,000 unused credits from the policy agent to the supplier agent's cap | Budget holder | Monday |
| Ask two people with no Copilot activity in 28 days whether they still need a seat | Team lead | Next review |
| Check why the supplier agent sent 18% of cases to people, up from 11% | Agent owner | Wednesday |
Where I got stuck
I haven't settled how far to automate the pack. A Power Automate flow could pull some figures, but not every admin centre view has an export I've confirmed, and I'd rather have a slightly manual pack that's right than an automated one that quietly drops a source. That's on my list to test.
Sources
- Microsoft Learn, Monitor Copilot Credit spending (Overview tab refreshes every 4 hours), accessed 11 October 2026.
- Microsoft Learn, Understand Microsoft Copilot reports (72-hour availability), accessed 11 October 2026.
- Microsoft Learn, Understand usage-based billing and cost management for Copilot Credits (reader roles), accessed 11 October 2026.