The situation
A client governance question landed on my desk recently. The organisation is rolling out Microsoft 365 E7, and their IT leadership wanted a straight answer before enabling broader agent functionality: can Copilot generate unexpected costs, and if so, where?
Reasonable question. The answer turned out to be less obvious than expected, because the licensing conversation around Copilot has produced a specific misconception that I have now seen repeated by more than one person in the same email thread.
The misconception goes like this: Researcher and Analyst are included with the licence for normal use, but when you ask them to do something heavy, deep planning, data visualisation, analysis across multiple documents, you cross a line into “agent operations” and start consuming Copilot Credits.
It sounds plausible. It maps neatly onto how metered cloud services usually work. It is also wrong, and getting it wrong sends you off governing the wrong thing entirely.
What the documentation actually says
There are two separate systems here, and most of the confusion comes from treating them as one.
The first is a query allowance. Any user with a Microsoft 365 Copilot licence can run a fixed number of combined Researcher and Analyst queries per calendar month. At the time of writing that figure is 25, and it resets on the 1st rather than on a rolling 30-day basis. Microsoft’s own documentation carries a caveat that the number may change, so verify it before you build a policy around it.
The second is Copilot Credits. These are a metered currency with a real unit price, purchased as prepaid capacity or billed through a pay-as-you-go meter linked to an Azure subscription. They are consumed by custom agents built or published through Copilot Studio, and by Cowork.
These do not overlap. The query allowance is not denominated in credits, and credits cannot be spent to extend it. Researcher and Analyst are core parts of the Copilot chat experience, and their consumption of compute is already accounted for in the licence. There is no threshold at which a Researcher query starts drawing on a credit balance, regardless of how many documents it ingests or how deep the reasoning goes.
The practical consequence is important: the query allowance is a capacity constraint, not a cost one. It cannot produce a surprise on an invoice. It can produce a user complaining that Researcher has stopped working until the 1st of next month, which is a different problem with a different fix.
There are important details to note:
- Researcher, Analyst and Facilitator are the built-in reasoning agents included with the Copilot licence.
- There is no native admin report showing remaining query usage per user. Users have to self-track, which makes the cap harder to govern than it first appears.
- Admins can block Researcher and Analyst from the Microsoft 365 admin center, but individual users cannot disable or unpin them.
- The allowance is per user. It is not a tenant-wide pool that heavy users can drain.
Where the credits actually apply
Having established that the built-in agents do not bill, the next question is what does. This is where Copilot Studio comes in, and where the answer is more nuanced than “custom agents cost money.”
For employee-facing scenarios, the inclusion is broad. When a user licensed with Microsoft 365 Copilot interacts with a Copilot Studio agent, and that agent operates under the authenticated identity of that licensed user, the core agent activity is included at no charge. That covers classic answers, generative answers, agent actions, tenant graph grounding, agent flow actions, the text and generative AI tool tiers, and content processing. Voice tiers include core agent activity as well.
So a custom agent built for internal staff, used by licensed staff, largely does not consume credits.
The exposure sits in the exceptions, and these are worth knowing precisely because they are easy to walk into:
- Computer-Using Agents are not included in the Microsoft 365 Copilot licence. They bill at the agent action rate regardless of who invokes them.
- Agent flows only qualify for the free inclusion on one trigger type. Runs triggered by the agent calling the flow, with an authenticated licensed user, are included. Agent flows using other triggers consume credits at the standard rate. The same flow can be free or metered depending on how it was started.
- Generative answers have a narrower carve-out than the table suggests. Agents incur charges for generative answer responses unless the agent was created in Agent Builder in Microsoft 365 and the response does not use tenant graph grounding.
- Unlicensed and external users are not covered. An agent exposed to customers or to staff without a Copilot licence bills for every interaction.
- Bring-your-own-model configurations are out of scope entirely and are billed separately.
- The inclusion is subject to fair usage limits, which Microsoft reserves the right to update. “No charge” does not mean unlimited.
E7 does not cover every variant of Copilot
This is the point I would most want an E7 customer to take away, because the SKU name works against it. E7 reads as the top tier, the one where everything is included, and that is how it tends to get discussed in commercial conversations. It is not how the Copilot portfolio is licensed.
The licence covers Copilot Chat, Copilot in the Office applications and Teams, and the built-in reasoning agents described above. It does not cover every experience that carries the Copilot name.
Cowork is the clearest example. It is not part of the licence, and there is no per-user Cowork entitlement to assign. It runs on Copilot Credits against usage-based billing, which means it bills through an Azure subscription rather than through the Microsoft 365 subscription the organisation already pays for. Cost scales with how much work is asked of it, not with how many people are licensed.
That distinction matters more than it first appears. With a per-user licence, the seat count is a natural ceiling: you know the maximum before you start. With a consumption model, there is no equivalent boundary. And as covered in the next section, usage-based billing is specifically the model without a hard stop.
The Frontier preview period muddied this further. During the preview, Cowork was available at no cost, which created a reasonable impression that it came with the licence. That period ended on 1 July 2026, and tenants that had not configured usage-based billing lost access at that point. If Cowork stopped working in your environment around then, that is the reason, and the underlying position is that it was never part of the entitlement to begin with.
The general lesson is worth applying beyond Cowork. When a new Copilot capability is announced, the useful question is not whether the organisation has a Copilot licence, but which billing model that specific capability uses: included in the licence, drawn from a query allowance, metered against Copilot Credits, or billed separately through Azure. Those are four different answers with four different governance consequences, and the licence tier alone does not tell you which one applies.
The part that surprised me
I went into this assuming that enabling pay-as-you-go was the cautious option, because it comes with budgets and alerts attached. That assumption is backwards.
Under the prepaid capacity model, Copilot Studio has genuine enforcement. When a tenant reaches 125% of its prepaid capacity, custom agents are disabled. Ongoing conversations finish, but subsequent attempts to invoke the agent are rejected until capacity is increased or resets. Users see a message telling them the agent has reached its usage limit. An administrator gets an email and a notification in the Power Platform admin center. Agent flow enforcement works separately and more surgically: new flow runs are blocked while the parent agent continues to answer normally for non-flow interactions.
Pay-as-you-go removes that. With a billing plan linked to an Azure subscription, overage enforcement does not apply, because the overage simply bills to Azure instead.
Meanwhile, the budgets you attach to a pay-as-you-go billing policy trigger email notifications when spending crosses a threshold, but do not enforce the limit or prevent your organisation from exceeding it. Usage continues uninterrupted. I have seen this described elsewhere as a hard cap. It is not, and Microsoft’s own documentation is explicit on the point.
So the position is this: prepaid capacity has a stop. Pay-as-you-go has a notification. Turning on pay-as-you-go to gain cost visibility also removes the mechanism that would have halted consumption.
There is a per-agent control that partially compensates. In the Power Platform admin center, under Licensing, Copilot Studio, Manage Agents, you can set monthly consumption limits for individual agents to cap usage before enforcement is triggered. That is worth configuring on anything autonomous.
Summary of scenarios
| Scenario | Billing behaviour | What controls it |
|---|---|---|
| Researcher or Analyst query, licensed user | No credits. Draws on the monthly query allowance | Per-user allowance, resets monthly. Admin can block the agents entirely |
| Facilitator in Teams meetings, licensed user | Included with the licence | Admin agent settings |
| Custom Copilot Studio agent, licensed user, internal use | Included at no charge for core activity | Restrict who can create and publish agents |
| Custom agent used by unlicensed or external users | Consumes credits per interaction | Publishing restrictions, DLP, per-agent limits |
| Computer-Using Agent, any user | Consumes credits. Not covered by the licence | Per-agent consumption limits |
| Agent flow called by the agent, licensed user | Included at no charge | Trigger design |
| Agent flow started by any other trigger | Consumes credits at standard rate | Trigger design, per-agent limits |
| Generative answer using tenant graph grounding | Consumes credits unless built in Agent Builder without grounding | Grounding toggle per agent |
| Bring-your-own-model agent | Billed separately, outside Copilot Credits | Azure subscription controls |
| Cowork | Outside the licence. Consumes credits billed through an Azure subscription | Requires usage-based billing to be enabled. Off by default since 1 July 2026 |
How to verify the position on a tenant
The thing that makes this genuinely difficult to govern is that Copilot billing is managed in two different administrative surfaces. Checking one and concluding you are safe is the obvious failure mode, and I suspect it is a common one.
Microsoft 365 admin center, Copilot, Cost management. This governs Cowork and the Work IQ API. If usage-based billing has never been activated, the page shows an “Unlock AI experiences enabled by usage-based billing” splash with a Get started button rather than the Overview and Consumption dashboards. That splash state is your confirmation that no default spending policy exists.
Power Platform admin center, Licensing, Copilot Studio. This governs custom agents. Look at the Capacity summary. Zero billing plans and zero total Copilot Credits under pay-as-you-go means no Azure subscription is linked for consumption. Check the prepaid capacity table as well, and then check the Environments tab, because pay-as-you-go can be scoped per environment. A single environment bound to a billing policy is enough to open a consumption path.
On the tenant I was reviewing, both surfaces came back clean: no spending policy, no billing plan, no linked subscription, no prepaid capacity, and zero consumption across every environment including the developer and sandbox ones. Makers can build and test custom agents there, but nothing can bill until someone deliberately creates a billing plan.
One caveat worth stating plainly to anyone you report this to: that is a snapshot of the current configuration, not a permanent lock. Any Global Administrator can create a billing plan later. The correct phrasing is that no consumption path is currently active, not that consumption is impossible.
If your organisation’s billing is managed through a partner, there is an additional consideration. The Cost management page prompts for this directly, and setting up credits through the partner keeps charges on the existing agreement rather than creating a separate direct-billing relationship.
Conclusion
Two things are worth separating here. The first is that E7 does not cover every variant of Copilot, and the tier name encourages the opposite assumption. Cowork is the obvious case, sitting outside the licence entirely and billing through Azure, but the wider point is that each Copilot capability has its own billing model and the licence tier does not tell you which one.
The second is that the cost exposure in an E7 deployment is not where most people look for it. Researcher and Analyst, the two features that feel expensive because they visibly do a lot of work, are capped by a query allowance and cannot generate spend. The exposure sits in custom agents reaching unlicensed users, Computer-Using Agents, agent flows with the wrong trigger type, and Cowork, and none of those can bill anything until a billing plan exists.
That makes the governance answer simpler than the licensing documentation implies. Leave billing not configured until there is a reason not to. Restrict agent publishing to a named group. Set per-agent consumption limits on anything autonomous. Verify both administrative surfaces rather than one, and re-verify periodically, because the configuration is not immutable.
And if anyone tells you that a budget in the Microsoft 365 admin center will stop the spend, check the documentation before you rely on it. This is something to keep in mind.
