What an AI agent actually costs: seats, credits and consumption units

01 / 15

What does an AI agent actually cost?

Nothing in Microsoft's stack is priced per agent. That is the first thing to establish, because most budget conversations start from the assumption that agents are the billable unit, and every published Microsoft price contradicts it. Microsoft's own licensing FAQ says it flatly: blueprints, agents and agent instances are not licensed separately and do not count as individual licensing units.

What you actually pay for is one of three things: a human being with a licence, a unit of consumption, or an identity that needs a mailbox. The third is the one people miss.

This matters more than a pricing curiosity because the three units are not interchangeable, they are administered in different portals, they draw down different commitments, and they are enforced at different thresholds. Which one you land on is decided by architecture choices — how the agent is triggered, whose identity it runs under, who it serves — made long before anyone looks at an invoice.

02 / 15

The seat that makes consumption free, through four gates

Here is the single most valuable and most widely mis-stated fact in this subject. A Microsoft 365 Copilot licence makes most Copilot Studio activity No charge — classic answers, generative answers, agent actions, tenant graph grounding, the AI tools, content processing. Not discounted. Zero.

That is genuinely the best commercial lever available, and it is why a seat that looks expensive next to a cent-per-credit meter often is not. But the inclusion is defined in a footnote with four cumulative conditions, and the footnote is where the money is.

03 / 15

The four conditions, stated plainly

Microsoft's wording is that employee-facing usage is included in the Microsoft 365 Copilot user subscription licence when the user of the agent is licensed and the agent operates using that authenticated user's identity. Unpack it and there are four separate ways to fall off.

• Business-to-Employee only. The inclusion is scoped to employee-facing scenarios. An agent built for customers is outside it, no matter who is licensed inside your tenant.
• The agent runs as the licensed user. Not merely that a licensed user exists somewhere in the tenant. If the agent authenticates as a service identity, this condition is not met.
• For agent flows, only one trigger qualifies. The inclusion applies to runs triggered by the When an agent calls the flow trigger. Agent flows using any other trigger consume credits at the standard rate of 13 per 100 actions.
• Fair usage limits apply, and Microsoft may change them. The footnote reserves the right to update the limits as the product evolves. No numeric limit is published anywhere. You cannot model against it, and you cannot plan on it staying where it is.
• Two named exclusions sit outside all four. Computer-Using Agent usage is explicitly not included in the Copilot subscription licence. And generative answers still incur charges unless the agent was created in Agent Builder and the response does not use tenant graph grounding.
• One documented silence. The voice rates — 10, 35 and 75 credits per minute — sit in a table with no Copilot-licence column at all. Whether voice is zero-rated is not documented. Do not assume it is.

04 / 15

The three units, and what each one meters

The unit changed under people's feet once already. On 1 September 2025 the currency for agents changed from messages to Copilot Credits, and as of 29 August 2026 Microsoft's own pages have not finished catching up — more on that below.

05 / 15

So what does an agent cost? Three pages, three answers

This is where a careful reader gets caught, because Microsoft states the per-agent question three incompatible ways and never reconciles them in one place.

06 / 15

The rule that actually decides whether an agent costs anything

Strip away the licensing prose and there is one clean, well-documented rule underneath, and it is about data rather than agents.

An agent that reasons over its own instructions or public web content costs nothing extra. An agent that reaches into shared tenant data starts a meter. Microsoft's extensibility cost guidance states it directly: declarative agents relying on instructions or public website grounding do not incur extra costs, while agents accessing SharePoint or Copilot connectors generate usage-based charges metered in Copilot Credits. Agents using metered consumption are off by default for Copilot Chat users.

That single sentence will do more for a budget than any licensing table. The expensive agents are the useful ones, and they are expensive precisely because they touch your data.

07 / 15

What the prices actually are, as of 29 August 2026

Every figure below was read from an official Microsoft or Anthropic pricing page on 29 August 2026. Prices move, and the Microsoft 365 suites moved on 1 July 2026, so treat the date as part of the number.

08 / 15

A note on bundle arithmetic

Do not compose bundle prices from their parts. Business Standard with Copilot is 23.50 dollars per user per month, while Business Standard at 14.00 plus Copilot Business at 18.00 comes to 32.00. The bundles are priced independently and the difference is not a discount you can quote — it is a different SKU.

One more correction worth making because the stale figure circulates widely. There is no seat minimum for Microsoft 365 Copilot on any current official page. The 300-seat purchase minimum was removed in January 2024. The live 300 figure is a ceiling on Copilot Business, not a floor, and inverting the two is the most expensive misreading available in this subject.

09 / 15

Which AI spend draws down your Azure commitment?

If you hold a Microsoft Azure Consumption Commitment, this is the question that changes the shape of a deal — and the documentation is uneven in a way worth knowing before you rely on it.

10 / 15

Two traps inside the commitment story

The first is a setup condition that silently costs you the benefit. Microsoft is explicit that commitment benefits apply only when the selected subscription links to a billing account that includes the commitment — and that if you use a different billing account or subscription, you still pay for consumption but it might not count. Note the hedge. Verify the billing account before you enable anything, because the drawdown is not retroactive.

The second is unreconciled between two Microsoft pages. The Marketplace commitment rules require that a purchase be completed through the Azure portal, while Copilot Credit billing is configured in the Microsoft 365 admin center. No page bridges the two. If a Copilot Credits commitment drawdown is load-bearing in your business case, get it confirmed in writing rather than inferred from a PDF.

And keep Azure prepayment separate in your head. Consumption covered by prepayment does not contribute to a commitment — but the prepayment purchase itself decrements it. That is the opposite of most people's mental model.

11 / 15

Three enforcement thresholds, one currency

When credits run out, three different things happen depending on which surface you are on. Microsoft documents each one clearly and never shows them together, which is how a team gets surprised on a Tuesday.

12 / 15

The pre-purchase plan is a one-way door

Prepaying credits buys a discount and gives up every form of flexibility, which is a reasonable trade only if the forecast is good.

The pre-purchase plan runs a one-year term and renews automatically by default. Microsoft states that cancel and exchange operations are not supported and that all purchases are final. You cannot split or merge a plan. Pre-purchase discounts do not combine with other discounts you hold, so an existing agreement can quietly make the listed tier percentage unachievable.

One thing to note precisely, because it is easy to assert and not supported: what happens to unconsumed units at the end of the term is not documented. The page says the plan runs for its term or until units run out, and stops there. If someone tells you they expire, or that they roll over, ask them for the page.

13 / 15

Where the documentation disagrees with itself

Naming the contradictions is more useful than picking a winner. All four were checked on 29 August 2026.

14 / 15

How I would model this

Author-proposed, not vendor guidance, and written for the version of this question that shows up in a board pack rather than an architecture review.

• Count humans first, actions second, agents never. Build the model on licensed users and metered actions. An agent count is a useful engineering metric and a misleading financial one.
• Treat the Copilot seat as a consumption hedge, not just a feature. At 30 dollars a month it removes metered charges for a large slice of employee-facing agent activity. Model the seat against the credits it displaces, not against its feature list.
• But size the credits as though the zero-rating does not hold. Three of its four conditions depend on how each agent is built, and the fourth is an unpublished fair-use limit. A model that assumes zero credit spend for licensed users is one architecture decision away from being wrong.
• Separate the two agent populations. Agents that reason over instructions and public content are close to free. Agents that touch tenant data are the ones with a meter. Forecast them separately, because they scale differently and only the second one grows with your data.
• Check the billing account before enabling anything. Commitment drawdown depends on the linked subscription sitting in the billing account that holds the commitment, and it is not applied retroactively.
• Do not prepay a curve you have not measured. The pre-purchase plan is final, non-cancellable, non-exchangeable, auto-renewing, and its treatment of unused units is undocumented. Run pay-as-you-go for a quarter and buy the discount against real numbers.

15 / 15

When is this worth bringing in outside help?

Not for a first agent. Build one, watch the meter, and you will learn more in a fortnight than any model will tell you.

It is worth help at three points. The first is a business case that has to survive finance — where the difference between a seat, a credit and a consumption unit decides whether the number is right, and where a commitment drawdown assumption may be resting on a PDF rather than a product page. The second is an estate already running agents across Copilot Studio, Microsoft 365 and Foundry, where nobody owns the whole picture and three enforcement regimes are quietly in play. The third is a prepay decision, because that one is irreversible by Microsoft's own wording.

Disclosure: Avalon Web Services sells Microsoft and AI integration work, including the Copilot pilot package linked below, so read this section as an interested opinion. Every price and rule above is cited to a vendor page you can open yourself, which is the part that does not depend on trusting us.