Copilot Credits are a currency. Who's keeping the books?

On August 3, 2026, Microsoft quietly turned every Copilot Studio agent into a spender.

That's the date Copilot Studio switched to usage-based billing. Every agent interaction now consumes Copilot Credits — Microsoft's billing unit across its agent platform, renamed from "messages" last September. Credits cost $0.01 each on the pay-as-you-go meter, or about $0.008 if you prepay through capacity packs: $200 a month for 25,000 credits, pooled across every agent in the tenant. Unused pack credits don't roll over. Consumption beyond the pack spills onto the Azure invoice at the full pay-as-you-go rate.

If you run finance at a Microsoft shop, read that paragraph again slowly, because it describes something you already know how to handle — you just haven't been asked to handle it here yet.

Copilot Credits are a virtual currency. A prepaid asset, purchased in blocks, consumed by activity, with two exchange rates (prepaid and spot), a monthly expiration, and an overage mechanism. Finance has dealt with instruments like this forever — postage meters, gift card float, prepaid cellular minutes. And the discipline has always been the same: you keep books on them. You know what was purchased, what was consumed, what it was consumed on, what expired unused, and what the effective rate turned out to be.

Nobody is keeping books on Copilot Credits.

The questions the admin center can't answer

Microsoft's admin tooling will show you consumption — total credits used, prepaid versus pay-as-you-go, by product. Genuinely useful. But watch what happens when the CFO asks the next layer of questions:

• Which agent spent these credits, and who owns it? Packs are pooled tenant-wide. The customer-service agent, the HR bot someone built in an afternoon, and the pilot nobody remembers are all drinking from the same pool.
• What's our burn-down? 25,000 credits reset monthly with no rollover. Are you exhausting the pack on the 22nd and paying a 25% premium for the rest of the month — or leaving 40% on the table every cycle? Both are real money; both are invisible without tracking.
• What did a credit actually cost us? Blend prepaid draws, expired credits, and overage spillover, and your effective rate is neither $0.008 nor $0.01. It's a number someone should be able to state.
• What's the month-end treatment? Prepaid packs want amortization against actual consumption. Overage wants an accrual before the Azure invoice lands. Expired credits want to be a visible line, not a silent write-off.
• Whose budget does any of this hit? "The tenant's" is not a cost center.

Then it gets worse: the two-meter problem

Custom Studio agents that call Azure OpenAI models incur per-token charges on top of their credit consumption. Same agent, two meters, two different units, landing on different invoice lines — and most organizations discover the second meter around month three, when the Azure bill arrives carrying charges nobody mapped to the agent that generated them.

So the true cost of one agent is: its credit draw (valued at your actual blended rate) plus its token spend plus its share of the pack it drinks from. No calculator on the internet computes that. It's not a calculation problem. It's an accounting problem.

What keeping the books looks like

The answer isn't another dashboard. It's the same discipline finance applies to every other currency it touches:

Attribution. Every credit consumed maps to a named agent, an owner, and a department — and consumption that maps to nothing posts to a suspense account where it stays visible until someone claims it. An agent nobody owns spending a currency everybody shares is precisely the situation suspense accounting was invented for. That is agent-level attribution in practice — and it is how Meridian treats every other source on Integrations.

Valuation, honestly labeled. Prepaid draws valued at the pack rate, overage at spot, with the basis stated on every row — because a pack-draw valuation is an allocation, and your auditors will care about the difference.

Burn-down and alerts. Pack exhaustion projected before it happens, expiring credits surfaced before they expire.

Reconciliation. The attribution detail ties out to the Azure invoice — subledger to bill — so the number finance reports is the number finance paid.

And the books proper: chargeback entries to the departments whose agents spent the credits, accruals for the overage, amortization for the packs — posted to the GL like every other cost, because that's what it is.

This is what Meridian is built to do — and Copilot Studio credit metering is landing in it now: per-agent attribution, valued at your actual rates, reconciled to your Azure invoice, in the same ledger as your OpenAI, Anthropic, and Bedrock spend — one agent, two meters, one set of books.

Your agents started spending a currency in August. The two-week, read-only Agent Spend Assessment will tell you where it's going — including the spend hiding on the second meter.

Microsoft, Copilot, Copilot Studio, and Azure are trademarks of the Microsoft group of companies. Onaro is not affiliated with, endorsed by, or certified by Microsoft. Pricing figures reflect Microsoft's published rates as of September 2026; consult Microsoft's official licensing documentation for current terms.