Copilot Credits, easy to approve and easy to lose track of
Microsoft named the meter: Copilot Credits are the single currency for agentic AI across Cowork, Copilot Studio, Dynamics 365 agents, and the Work IQ APIs, all on one pooled tenant balance at one cent per credit, replacing the prior message based meter in September 2025. They decrement your Azure commitment like any other Azure spend, which makes them easy to approve and easy to lose track of, and the license and the meter are two separate lines that budget separately or surprise together.
Prepared by Redress Compliance · August 8, 2026 · Microsoft advisory. Based on 30 to 40 Microsoft EA renewals benchmarked 2024 to 2025.
Executive summary
The credit hides the bundle, and the task complexity sets the bill.
A credit looks like one cent, and what it packages is the model, the runtime orchestrating the agent, the context retrieved from mail and files, and the tool actions taken, which is why it bills by how hard the task was: a light task runs under two dollars and a heavy one over fifteen.
And you cannot reverse a credit into a clean token rate, which is exactly what makes the meter hard to benchmark against a direct model price.
The subscription still covers everyday Copilot Chat and the in app experiences; the credits meter the autonomous work on top.
All three purchase routes burn the MACC, and the prepay floor is 0.8 cents.
Pay as you go bills monthly in arrears, Capacity Packs run $200 per tenant per month for 25,000 credits resetting monthly, and the Pre-Purchase Plan tiers from 5 percent at a $3,000 commitment to 20 percent at $3 million.
An effective floor of 0.8 cents that even the deepest commitment never beats: unused prepay credits expire at term end, and pools sized on vendor forecasts expired with 10 to 30 percent unused in the deals we reviewed.
The choice is not whether credits touch the commitment, it is whether you stay flexible or prepay for a capped discount.
The MACC makes the meter feel free, and governance slips first. Credits drawn against an underused MACC felt free, so spend governance was the first control to slip: the burn helps the commitment along, nobody owns the meter, and the invoice arrives as Azure spend everyone already approved.
The controls install before provisioning, not after the first invoice, spend caps at the tenant, per persona thresholds, and the task mix modeled in dollars first, because the buyers who did kept the meter predictable and the ones who prepaid on a forecast wrote off credits.
The buy or build comparison is the negotiation's honest anchor. Heavy agentic workloads priced per unit of inference ran several times the cost of the same work built on direct model rates, and the same task on Claude direct costs a fraction per inference unit.
The gap being the managed agent rather than the model: the deepest prepay still pays four fifths of list while batch processing and prompt caching on a direct contract cut effective rates far below that.
The managed runtime is worth something, and pricing it explicitly, against the direct alternative, is what keeps the premium honest.
The three purchase routes, and what to watch
| Route | How it works | What to watch |
|---|---|---|
| Pay as you go | One cent per credit, billed monthly in arrears | Fully variable, visible only after the fact |
| Capacity Packs | $200 per tenant per month for 25,000 credits | Unused credits reset monthly, steady volume only |
| Pre-Purchase Plan | Annual pool, 5 to 20 percent tiered discount, paid up front | The discount caps at 20 and unused credits expire |
| All three | Decrement the MACC like any Azure spend | Easy to approve, easy to lose track of |
The tier ladder's fine print is the floor.
The Pre-Purchase Plan runs from 300,000 credits at $3,000 with 5 percent off through 30 million at $300,000 with 10, to 300 million at $3 million with the headline 20.
An effective per credit price falling only from 0.95 to 0.8 cents: the deepest commitment in the program still pays four fifths of list, unused credits expire at term end.
And the forecast that overshoots is money burned, which is why the pools size from measured task mix and never from the adoption pitch.
The single pool, and what it covers
- Cowork and Copilot agents: the autonomous task work reading mail and files and writing back, the heaviest draw in the pool.
- Copilot Studio and Dynamics 365 agents: the custom and first party agents teams build and deploy, on the same balance.
- The Work IQ APIs: generally available from June 16, 2026 and billed from the same pool, extending the meter to the API surface.
- Copilot Chat for unlicensed users: consumption metered to the pool rather than a seat, the quiet expansion path.
- The double edge: one meter simplifies the invoice and concentrates every AI workload onto a single currency whose pricing the vendor controls.
The Copilot Credits cost brief
The task mix modeling, the tier arithmetic, the MACC interaction, and the governance thresholds worked end to end.
Get the white paper →Buy or build, the comparison that anchors the price
The credit cannot be reversed into a token rate because it bundles the model with the runtime, the retrieval.
And the tool actions, which is precisely why the buy or build comparison runs on outcomes: the same task costed on Copilot Credits against a direct model contract with batch processing and prompt caching, where the direct route's effective rate falls far below the 0.8 cent prepay floor and the difference prices the managed agent.
The managed runtime is worth real money for estates without platform engineering; it is worth pricing explicitly for estates with it.
The seat side subscription economics run in the Copilot licensing guide, the EA structure the credits negotiate inside in the Enterprise Agreement guide, the MACC sizing discipline in the Microsoft overspending report.
And the direct model economics on the other side of the comparison in the token economics report and the Anthropic pricing history.
- Percentile standing for your exact deal size and industry, from real closed transactions
- Scenario simulation before the call: test alternative terms and see the financial impact of each
- A negotiation playbook, talking points, and a two page executive brief on day one
What we saw across EA renewals, 2024 to 2025
Across roughly 30 to 40 Microsoft EA renewals Fredrik Filipsson benchmarked in 2024 and 2025, the Copilot conversation moved from a flat per user add on to a variable meter inside a single year:
Of pools sized on vendor forecasts, written off unused at term end.
Heavy agentic workloads against the same work built on direct model rates.
The pattern matches every consumption meter in this practice, with one Microsoft specific twist: the MACC interaction makes the credits the easiest AI spend in the enterprise to approve, because the burn counts toward a commitment someone already signed.
And the easiest to lose track of for the same reason.
The buyers who kept it predictable modeled the task mix in dollars first, installed the caps and per persona thresholds before provisioning, held the license and the meter as two separately budgeted lines, and treated the prepay ladder as what it is, a capped discount against an expiring pool.
Sized from measured burn or not at all.
Your first five moves
- Model the task mix in dollars before provisioning, light tasks under two dollars and heavy ones over fifteen.
- Install spend caps and per persona thresholds first, because governance slips when the MACC makes credits feel free.
- Budget the license and the meter as two lines, the subscription and the agentic work it does not cover.
- Size any prepay from measured burn, against the 10 to 30 percent that expired unused, and remember the 0.8 cent floor.
- Run the buy or build comparison on outcomes, pricing the managed agent against direct model rates. The Microsoft practice runs the model with you.
Frequently asked questions
What are Microsoft Copilot Credits?
The single usage meter for agentic AI across Cowork, Copilot Studio, Dynamics 365 agents, and the Work IQ APIs, pooled at the tenant at one cent per credit and charged on top of the per user Copilot subscription, which still covers everyday Chat and the in app experiences.
The currency replaced the prior message based agent meter in September 2025, and all consumption decrements the Azure MACC.
How much do Copilot Credit tasks cost?
By complexity, not by message: a light task runs under two dollars and a heavy one over fifteen, because each credit bundles the model, the orchestrating runtime, the context retrieval, and the tool actions, billing by how hard the task was.
The bundle is also why a credit cannot be reversed into a clean token rate, making the meter structurally hard to benchmark against direct model prices.
How do you buy Copilot Credits?
Three routes, all decrementing the MACC: pay as you go at one cent billed monthly in arrears, Capacity Packs at $200 per tenant per month for 25,000 credits that reset monthly, and the Pre-Purchase Plan with tiered discounts from 5 percent at $3,000 to 20 percent at $3 million.
The prepay floor is 0.8 cents even at the deepest tier, and unused prepaid credits expire at term end.
Should you prepay for Copilot Credits?
Only against measured burn: pools sized on vendor forecasts expired with 10 to 30 percent unused in the deals we reviewed, and the discount caps at 20 percent, so the forecast that overshoots burns money the discount never recovers.
Pay as you go through a measurement period, model the task mix in dollars, and let the demonstrated volume justify whichever tier it actually reaches.
Why does the MACC connection matter for Copilot Credits?
Because it makes the meter frictionless in both directions: credits burn down the Azure commitment like any other spend, so approval is easy and the consumption even helps an underused MACC along, which is why spend governance was the first control to slip in our renewals.
The caps and per persona thresholds install before provisioning, since the invoice arrives inside Azure spend everyone already approved.
Are Copilot Credits cheaper than direct model contracts?
No, and the gap is the managed agent rather than the model: heavy agentic workloads ran several times the cost of the same work built on direct rates, and batch processing plus prompt caching on a direct contract cut effective rates far below the 0.8 cent prepay floor.
The managed runtime is worth real money for estates without platform engineering, and the buy or build comparison on outcomes is what prices that premium honestly.
