Forecasting Copilot Credits from decision volumes

Start from the work

When people forecast AI spend, they usually take last month's bill and add a percentage. That's fine for licences, which only change when someone buys or cancels a seat. It's weak for metered use, because the bill follows activity, and activity follows the work.

The work is easier to forecast than the bill. A service desk knows roughly how many tickets arrive in January. A finance team knows month-end is busy. Those volumes already sit in ticket logs, case systems and approval histories, which is the same data behind the decision inventory.

The sum

For each decision type an agent supports:

Forecast credits = decisions expected × share that go through the agent × credits per decision

Then add up the decision types and multiply by the credit price. Microsoft's pay-as-you-go rate is $0.01 per Copilot Credit.

Credits per decision should come from measurement once the agent is live. Before that, estimate it from the agent's design, as in Budgeting for an agent before it goes live.

A worked example

These numbers are illustrative.

A service desk's ticket system shows these monthly volumes for password and access requests last year: about 2,000 most months, rising to 3,200 in September when new starters arrive. An agent handles 70% of these requests, and the first two months of logs show an average of 14 credits per request.

If the finance team had used last month's bill plus 5%, they'd have budgeted around $206 for September and been caught out by the new-starter peak. Volume-led forecasting sees it coming because the ticket system already did.

Watch the share as well as the volume

The share that goes through the agent tends to grow as people trust it, and it can jump when someone adds a new trigger or puts the agent in a new channel. I'd forecast it as its own line and agree changes to it in advance, so a jump in spend has a named cause.

Check the forecast

Each month, compare forecast with actual and note the gap and its cause: volume, share, or credits per decision. Over a few months that tells you which input is least reliable. Microsoft Cost Management can also send forecast alerts on Azure budgets, which notify you when projected spend is likely to cross a threshold. Microsoft's documentation says budgets are evaluated against costs every 24 hours and cost data typically takes 8 to 24 hours to arrive, so an alert can arrive a day or two after the spending that caused it.

What I'm still checking

For Copilot experiences billed through the Microsoft 365 admin center, such as Cowork, I haven't yet confirmed what forecasting the new Cost Management dashboard offers beyond trends and limits. Until I've seen it, I'd build the forecast in a spreadsheet from volumes and compare it with the dashboard's actuals.

I'm also unsure how stable credits per decision will be as Microsoft changes models. Copilot's Auto mode picks a model per request, and Microsoft says spending policies can set which models are available to groups. A change of model could move the rate without any change in the work.

Sources

• Microsoft, Copilot Studio licensing guidance, accessed 11 October 2026.
• Microsoft Learn, Tutorial: Create and manage budgets (forecast alerts, evaluation every 24 hours), accessed 11 October 2026.
• Official Microsoft Blog, Introducing the new Copilot with Home, Code and Autopilot, 25 September 2026, accessed 11 October 2026.