Copilot Studio Autonomous Agents: Control Copilot Credit Costs With Hard Limits

New York consultancy eSoftware Associates (ESW) has published a guide on Copilot Studio agent pricing and governance. It was announced through a GlobeNewswire release that The Manila Times carried on October 19, 2026. ESW is a Microsoft partner that sells Copilot consulting, so the guide is partly marketing. Its central warning still holds up against Microsoft's own documentation. Once agents run on schedules and event triggers instead of waiting for a prompt, Copilot Credits become a real budget line, and some of the controls that look like spending limits only send alerts.

Here is what the guide says, what Microsoft's documentation confirms, and where admins need to read more carefully.

Why this is coming up now​

ESW's guide, dated October 16, follows Microsoft's October 15, 2026 announcement of three Copilot Studio additions for autonomous work:

• Agent accounts. An agent gets its own Microsoft Entra account, a mailbox, Teams presence and a directory entry.
• Business process memory. An agent with an account keeps one shared memory for everyone who works with it.
• New triggers for agents built on the GitHub Copilot harness.

All three are rolling out through Microsoft's Frontier preview program. Microsoft's FAQ on agent accounts labels the feature prerelease documentation and says its availability and capabilities may change. Whether a tenant has these features depends on enrollment and licensing. ESW adds that full agent accounts need an early-release environment, Microsoft Agent 365 licensing for makers and a Frontier for AI Teammates license for each agent. That licensing detail comes from ESW and should be checked against your own agreement.

Section summary: These are preview features, not general availability. The cost and governance questions apply as soon as a tenant turns them on.

The pricing math, checked​

Copilot Studio bills agent work in Copilot Credits. Microsoft's own pricing page confirms the prepaid rate: a customer pays up front ($200.00/pack/month for 25,000 Copilot Credits). Microsoft also offers a pay-as-you-go option that requires no up-front license commitment and allows customers to pay only for the Copilot Credit capacity they use at the end of the billing period. Several independent licensing trackers put that meter at $0.01 per credit via Azure. ESW also says unused pack credits don't roll over, and third-party pricing guides agree, though I couldn't confirm that in Microsoft's own pages.

Microsoft's Copilot Studio billing documentation lists these per-feature rates for standard-harness agents:

ESW's example agent does four things per run: one trigger, two tool calls and one generative answer. That is three agent actions at 5 credits plus one answer at 2 credits, or 17 credits per run. Here is what that comes to over a 30-day month:

The arithmetic matches ESW's "about 73,000" and "about 510" figures. In a 31-day month the 10-minute schedule reaches about 75,900 credits. The point is the ratio: the same agent costs roughly 144 times more depending on one setting. ESW's CEO Russell Kommer put it this way: "The schedule tends to be the biggest cost decision, and someone makes it in five seconds."

Keep in mind that 17 credits per run applies only to ESW's simple example. Add tenant graph grounding (10 credits), switch to a reasoning model (premium token charges) or call agent flows, and the cost per run goes up. Microsoft recommends forecasting with its Copilot Studio agent usage estimator before building, and that is worth doing.

Section summary: The prepaid pack price is confirmed by Microsoft. How often the agent runs matters far more than the per-action rates.

"But we already pay for Microsoft 365 Copilot"​

The press release says agent runs are billed in Copilot Credits even when an organization has Microsoft 365 Copilot licenses. As a planning rule for autonomous agents, that is fair. As a general statement about all Copilot Studio usage, it goes too far.

Microsoft's billing documentation says employee-facing use of Copilot Studio agents is included when the user has a Microsoft 365 Copilot license and the agent operates using that user's authenticated identity. The exceptions explain why autonomous agents get billed:

• Agent flows are free only when started through the "When an agent calls the flow" trigger by an authenticated, licensed user. Microsoft's admin guidance says flows an agent starts while running autonomously, and flows with other triggers, are billed.
• Agents built on the GitHub Copilot harness are billed for everything. Microsoft says building, testing, evaluating and running them all consume credits, and none of it is included with a Microsoft 365 Copilot license.
• Computer-Using Agents are not included in the license either.

So the question to ask is whose identity the agent runs under and which harness it uses, not simply whether users are licensed. ESW also points out that event triggers currently run on the maker's credentials. That is a data-access concern as well as a billing one: if the builder can see the whole finance share, so can the agent.

Controls that stop spending, and controls that only alert​

This is the most useful part of the story for admins. ESW recommends six measures before the first agent goes live: a per-agent credit cap with a hard stop, an environment capacity cap, a schedule that matches real need, a named owner, Microsoft Purview auditing, and a review before publishing anything that runs on a schedule, sends email or writes to a system of record. Microsoft's documentation shows these controls do different jobs.

1. Set a per-agent hard stop​

Microsoft's Power Platform admin guidance describes the steps:

1. In the Power Platform admin center, go to Licensing > Copilot Studio.
2. Select Manage Agents.
3. Select the agent, then Set limit.
4. Enter the monthly Copilot Credit limit.
5. Set the notification threshold.
6. Turn on Stop usage if the agent should switch off when it hits the limit.
7. Select Save.

Microsoft says an agent limit contains one workload. It does not cap total consumption across an environment.

2. Make the environment allocation an actual limit​

On its own, an environment allocation doesn't stop anything. Microsoft says it isn't a boundary if the environment can keep drawing from tenant capacity or from pay-as-you-go billing. To contain prepaid use, allocate credits to the environment and, under Capacity overages, clear Draw from the available capacity in my tenant. That option is selected by default for eligible environments that have no allocation set. Clearing it still doesn't block usage billed through a linked Azure pay-as-you-go plan.

3. Don't treat Azure budgets as a limit​

For pay-as-you-go environments, Microsoft says Azure budgets and alerts send notifications but don't stop Copilot Studio usage. Pair them with per-agent limits that have Stop usage turned on.

4. Know about the 125% cutoff​

Under the prepaid capacity model, Microsoft disables custom agents once consumption reaches 125% of capacity. Users then see a "There is a billing issue" message. How far that reaches depends on how credits are allocated. In Microsoft's own example, an environment with its own unused allocation, or one billed pay-as-you-go, isn't affected when the shared tenant pool goes into overage. Agent flows follow a separate rule: when prepaid capacity is fully used, new flow runs are blocked, but the parent agent keeps working for everything else.

ESW's practical point stands. One runaway agent drawing from a shared pool can take down the agents your teams rely on. Isolating allocations per environment is what keeps one agent's overspend from spreading.

5. Owners and audit logs give visibility, not limits​

Microsoft's own pre-deployment checklist asks admins to decide who reviews usage and responds to alerts, so a named owner is good practice. But an owner and a Purview audit log don't stop spending. ESW says Purview logs Copilot Studio creation, publishing and interaction events. Turn auditing on, but count it as forensics, not as a spending control.

Section summary: Only per-agent limits with Stop usage, and environment allocations with tenant draw turned off, actually stop spending. Budgets, alerts, owners and audit logs only tell you about it.

Shared memory is a data-governance issue​

Cost isn't the only risk. Microsoft's preview FAQ is unusually frank about shared memory for agents with accounts:

• Summaries of shared files and chats can appear in answers to any user, even users without permission to the originals and regardless of sensitivity label.
• Microsoft says shared memory "has no privacy boundary between users."
• End users can't opt out during the preview, and admins can't turn shared memory off without turning off agent accounts entirely.
• Shared memories are deleted automatically after 28 days without any interaction.

ESW suggests treating the agent's audience like a distribution list: limit it to a team that already shares the same access. That is a sensible precaution, and it is the strongest argument for ESW's review before publishing.

Reading the source with some skepticism​

This is a vendor's guide promoting its own consulting services, and readers should weigh it that way. Recommendations such as naming an owner and starting with "one recurring process" are ESW's advice, not Microsoft requirements. Some figures in the guide, such as the non-rollover claim, are confirmed by third parties rather than Microsoft. ESW's list of controls also gives owners and logging the same billing as hard stops, even though Microsoft's documentation distinguishes them.

On the main points, though, ESW's guide agrees with Microsoft's documentation: the run schedule drives cost, per-agent limits with Stop usage are the real cap, and pay-as-you-go budgets only alert. It also aligns with a deadline Microsoft has already applied: developer and trial environments moved to usage-based billing on September 1, 2026.

Quick checklist before your first autonomous agent​

• Model the cost per run and the schedule in Microsoft's usage estimator.
• Confirm the harness and identity scenario, including whether the agent runs under the maker's credentials.
• Set a monthly per-agent limit with Stop usage on.
• Allocate environment credits and clear Draw from the available capacity in my tenant.
• If pay-as-you-go is linked, add Azure budgets, but don't treat them as a limit.
• Turn on Purview auditing and check consumption weekly at first.
• Limit the audience of any agent with an account to a team that already shares the same access.
• Consider whether a Power Automate cloud flow would do the job instead. Microsoft says those use Power Automate licensing, not Copilot Credits.

An agent that waits for a prompt is easy to budget for. One that runs every ten minutes needs a hard limit before it goes live.