Executive summary

Microsoft changed the unit of Copilot from a seat to a meter, so the cost of an agent is now set by how it is built, not by how many people use it. For three years the assistant was sold per seat at a price known on day one. The agentic surface that drives Copilot growth today, Copilot Studio agents, autonomous flows, and Copilot Cowork, bills instead on a metered currency called the Copilot Credit, which replaced per-message billing in September 2025. This briefing works through how the meter operates, the published rate card, the three ways to buy credits, how that spend touches the Azure commitment most enterprises already hold, and the terms to secure before routing material AI spend through Microsoft.

Our advisors sit on the buyer side only. Across more than 500 enterprise engagements, the buyers we advise have negotiated over $2.4 billion in software and cloud contracts at an average saving near 38 percent, and our audit defence work averages a 72 percent reduction against the initial claim. The figures below are Microsoft's published prices and the public mechanics that frame a Copilot Credit commitment, not a quote.

1. The consumption shift, from messages to credits

On September 1, 2025, Microsoft retired the per-message billing that had governed Copilot Studio and replaced it with a single metered currency, the Copilot Credit. Under the old model a message was a roughly comparable unit whatever the agent did. Under credits, the exchange rate floats with architecture. A scripted answer costs a fraction of a cent, a reasoning-heavy response costs more than a dollar, and both run on the same platform.

Microsoft is extending the same meter beyond Copilot Studio. Copilot Cowork bills on usage through credits, which signals that everything agentic is moving to consumption while seats remain only for the in-app assistant. The Microsoft 365 Copilot seat still lists at $30 per user per month, but that seat now sits beside a meter that scales with machine behaviour rather than headcount.

For a buyer, the consequence is structural. The cost of an agent is decided by the people who design it, and the builder gives no warning as expensive features are added. The friction of buying is low, because agents consume under an Azure agreement you already hold, with identity and billing you already trust. Low friction is exactly where commercial detail gets skipped.

This is not a pricing tweak. It is a change in who controls the cost line. When Copilot was a seat, procurement owned the number, because the number was a seat count multiplied by a list price. When Copilot is a meter, the people who build agents own the number, because every grounding call, tool call, and reasoning step they add changes it. The job of this briefing is to put that control back where it can be governed.

2. How the money works, the published credit rate card

A Copilot Credit is debited whenever an agent does work. The price of a credit is fixed. The number consumed per action is not, and that rate card is the most important table in Copilot economics. Microsoft publishes the consumption values, and the extremes are where the budget is decided.

The figures that matter are the extremes. A scripted answer is one credit. A reasoning call is a hundred credits per ten responses, a hundred-fold spread inside one product. The operations also stack within a single turn. Microsoft's own example is a grounded generative response at twelve credits, ten for grounding plus two for the answer. Add tools and reasoning and one turn passes a hundred credits.

Two agents that look identical to a user can differ many times over on the invoice, purely because of how each was built. That is the single fact a cost review must absorb. Credits are pooled at the tenant level, not charged per seat, so the meter rewards designs that ground, reason, and call tools only when the task truly needs them.

A worked example of how a turn adds up

Take a service desk agent that answers a policy question. If it returns a scripted answer, the turn costs one credit. If it generates the answer, two. If it grounds that generative answer on tenant data through Microsoft Graph, the grounding adds ten, so the turn is twelve. If it then calls a connected tool to open a ticket, add the action cost, and if it routes the hardest cases through a premium reasoning model, the turn climbs again. The same business outcome, a answered question and an opened ticket, can land anywhere from one credit to well over a hundred depending only on the path the designer chose.

Multiply that single turn by the volume of a production agent and the annual figure is decided long before finance sees it. At a million turns a year, a design that averages two credits a turn and a design that averages forty credits a turn are not a rounding difference. They are two different budgets wearing the same product name.

3. The three ways to buy, and who should use which

Beyond the zero-rated inclusion path, where internal agents for users who already hold a Microsoft 365 Copilot licence consume no credits, Microsoft sells the same credit three ways. Each carries a different unit price and a different failure mode.

Each fits a different stage of maturity. Pay as you go is the most expensive credit and the most flexible, the right place to start while you learn what an agent burns. Prepaid packs cut the rate by roughly a quarter for a predictable base load, but enforce at 125 percent of a shared, non-rolling pool. Exceed it and agents are disabled until you add credits or enable pay as you go. The pre-purchase plan takes the rate lower for an annual commitment and adds automatic overflow, but belongs only to mature estates with the consumption history to commit without over-buying.

The structures are not mutually exclusive, and the mistake is to treat the decision as a single, permanent pick. The pattern that works is layered and sequenced. Start on pay as you go and inclusion, gather two to three months of real consumption, then cover the demonstrable base load with packs or a pre-purchase commitment at the lower rate, keeping pay as you go enabled underneath as the overflow valve.

One nuance on the inclusion path deserves attention, because it is where real savings hide. An agent that serves your own licensed staff for internal work can run without consuming credits, while a near-identical agent pointed at customers or unlicensed users runs on the meter. Audit your existing agents for ones that are billed but could be free, and reserve metered credits for genuinely external-facing work. That single review often returns more than any rate negotiation.

4. How credits touch your Azure commitment

For enterprises carrying a Microsoft Azure Consumption Commitment, credits are not only a cost. Pay as you go and pre-purchase credits are billed through Azure and decrement the MACC the way native Azure spend does. That cuts two ways, and the buyer should hold the number before the seller introduces it.

If you are underconsuming a MACC, and many enterprises are after committing to an ambitious figure, routing agent spend through credits converts a promise you have already made into capability the business wants, at no incremental commitment. If, instead, Microsoft is sizing a new or renewed commitment, an AI roadmap is the perfect reason to inflate it, hardening a speculative forecast into a fixed obligation.

The practical move is to separate the two questions Microsoft will try to merge. Question one is whether buying AI through Azure helps you retire commitment you have already made. Often it does, and that is a real benefit worth taking. Question two is whether your AI roadmap should be used to raise the size of your next commitment. That is a separate decision, and it should be made on its own evidence, not folded into an AI enthusiasm that the account team is happy to encourage.

5. Where the cost leaks, agent design is the budget

Because the meter follows architecture, the cost of an agent is set in engineering and discovered in finance. The chart below shows the relative credit burn of four agent designs answering the same business question. It is an illustrative index with the scripted design set to 100, not a market benchmark and not a quote.

The lesson is direct. The same answer can cost a few credits or a few thousand, depending on whether it grounds on tenant data and whether it calls a reasoning model. A design that grounds on every turn when most turns need no grounding will run a large bill for no added value. Costing the design before it ships is the only control that works upstream of the invoice.

Make the design specification a cost document

Require every agent to ship with a costed design specification, credits per turn, expected volume, purchase path, and inclusion status. Treat that document as a release gate. An agent without a credit forecast is an agent without a budget, and it will find its budget on the next Azure invoice.

6. Governing runaway consumption

A meter with no governance is a blank cheque written in tokens. The control framework is operational before it is contractual, and it rests on measurement, gating, and ownership.

Turn on the Copilot Credits report and read it weekly. You cannot size a commitment, or govern spend, against consumption you have not measured. Set alert thresholds on the shared pool well below the 125 percent enforcement point, so a runaway agent triggers a human review rather than a service interruption or a surprise overage.

Assign an owner for agent spend the way you assign an owner for cloud spend. Per-seat assistants belong with whoever runs the Microsoft agreement. Metered agents belong with the cloud commercial owner because they land on the Azure bill. An organisation that has not named those owners is measuring only part of its Copilot exposure.

The three reports to run from week one

Run three views and you will catch most surprises early. The first is credit consumption by agent, which shows where the spend concentrates and which agents deserve a design review. The second is consumption by action type, grounding against reasoning against tool calls, which tells you whether the expensive verbs are being used where they earn their cost. The third is the pool position against the enforcement threshold, which protects production agents from going dark on a shared-pool exhaustion at the worst moment.

None of these reports is exotic. They are the same discipline a mature cloud cost practice already applies to compute and storage, pointed at a new meter. The organisations that struggle are the ones that treat Copilot as a feature upgrade rather than a consumption service, and so never stand up the reporting until the first large invoice forces them to.

7. What to demand in writing

The friction is low, which is exactly why the terms get skipped. A meter is repriced more easily than a seat, because a seat price change is visible and a consumption value change is buried in a rate card most buyers never read. Before routing material AI spend through Microsoft, confirm each of these in the agreement, not on a call.

8. The commitment calendar and our recommendation

Copilot Credit commitments reward a calendar, not a meeting. Microsoft's fiscal year ends June 30, and its quarters close at the end of September, December, and March. Quarter ends sharpen the flexibility an account team can reach, so a buyer with a consumption baseline already in hand can time the conversation rather than react to it.

Our recommendation: instrument before you commit, build internal agents on the zero-rated inclusion path first, cost every agent design as a release gate, layer and sequence the three purchase paths with pay as you go held underneath as overflow, and know your MACC position before any AI conversation. Treat the published rate as fixed and the agent design as the variable, because design is where the Copilot bill is actually decided.

The discipline is the same one that protects every software category. Treat the first proposal as an opening position, insist on measurement before you scale, and never let a pilot convert to a production commitment without a contract negotiated as deliberately as the agent was designed. The buyers who hold their consumption data, their MACC position, and their design costs before the conversation starts are the ones who set the number rather than receive it.

Sources: Microsoft published Copilot Studio credit rate card, Microsoft 365 Copilot list pricing, and Azure consumption commitment mechanics, as available at the time of review. Modelled indices are Atonement Licensing advisory illustrations, not a quote.

Frequently asked questions

What is a Copilot Credit?

Microsoft's metered currency for agent work, debited whenever an agent answers, grounds on data, calls a tool, or runs a flow step. It replaced per-message billing in September 2025. A credit lists at $0.01 pay as you go or $0.008 in a pack, and one interaction can consume from 1 to over 100 credits depending on design.

How do you buy Copilot Credits?

Three published ways beyond the zero-rated inclusion path. Pay as you go bills $0.01 per credit via Azure. Prepaid capacity packs list at $200 for 25,000 credits, $0.008 each, no roll over, with 125 percent enforcement. An annual pre-purchase plan lists at up to 20 percent off with automatic pay as you go overflow.

Do Copilot Credits count toward an Azure MACC?

Pay as you go and pre-purchase credits are Azure billed and count toward a Microsoft Azure Consumption Commitment. Capacity packs are the least aligned. For an underused MACC, agent spend can retire commitment you have already made, so confirm your chosen path decrements the MACC in writing.

Why do Copilot Credit costs vary so much?

Because consumption follows agent design, not user count. Scripted answers are 1 credit, generative 2, grounding 10, reasoning 100 per ten responses, and they stack within a turn. Two agents that look identical to a user can differ many times over on the invoice purely because of how each was built.

Related research

Related reading: Microsoft Licensing hub, Microsoft Licensing Experts, Azure MACC Negotiation Guide, and Microsoft EA Negotiation Playbook.