Die AgentinNICOLENDERS.COM
DEEN
← Dispatches

AnalysisTHE BUILDERTHE ARCHIVIST

An agent takes the inventory now. Not the decision.

Published

A brass pneumatic tube station fills a wooden pigeonhole rack with sorted capsules, while a wire in-tray beside it overflows with unopened capsules and the chair stands empty.AI generated
Automate governance with Agentic Center of EnablementMicrosoft Learn

Since July 2026 an agent may already be assessing your governance inside your tenant. You did not order it. The Agentic Center of Enablement is a managed platform feature and is switched on automatically for Power Platform administrators as soon as the tenant holds at least one managed environment. No rollout, no toggle, no project phase.

Three agents, one approval step

Agentic CoE brings three agents into the Power Platform admin center that work as one chain.

The Highlights agent produces a daily automated snapshot of tenant activity with no configuration, surfacing things like newly created resources and capacity consumption.

The Insights agent scans the tenant continuously for issues across security, compliance, performance and adoption. The examples Microsoft names are the right ones: ownerless resources and activity in the Default environment. It ranks findings by impact.

The Action Plan agent turns those findings into full remediation plans that you review and approve before execution. Every agent action lands in an audit trail on the Activity page.

The approval step is the most important part of that description. Proposing yes, executing only after your consent. That is the line to keep an eye on in case a later wave moves it.

Being on by default is the actual news

The feature has been in public preview since July 2026 and the release plan sets no general availability date. The trigger is a single managed environment in the tenant, and managed environments are something you either introduced deliberately or drifted into, for instance through pipelines: for deployment pipelines every environment involved has to be managed, apart from developer environments and the pipeline host.

Two consequences for planning. First, the question is not whether you adopt Agentic CoE but who in your organisation reads the Actions page. Second, a preview without a GA date is nothing to build a process on, and nothing you can ignore either, because it is already running.

Inventory was the easy part

My dispatch on inventory before governance argued that governance without an inventory is theatre. That part is largely solved now. Power Platform inventory has been generally available since March 2026 and gives tenant administrators a unified view of cloud flows, Copilot Studio agent flows and Workflows agent workflows across every environment. Connectors, actions and usage data are announced.

That moves the bottleneck. The question is no longer what exists in your tenant. The question is who reads the list and who signs off on the consequence.

Where the cost controls have holes

On the money side there is more available than most teams use, and less than the name suggests. Under Licensing, Copilot Studio, Manage Agents you see every agent in the tenant, month-to-date consumption and any configured limits, and you can set a monthly Copilot Credit limit per agent, with notification as consumption approaches it and a stop when it is reached.

Three details worth knowing before you treat that as a cost brake.

Agent limits do not add up to an environment limit. If you need the aggregate cap, you work through the billing policy. That does not come from the product documentation but from a blog run by the Microsoft team behind Copilot Studio, and I could not confirm it in that form on Learn.

The built-in notifications go to tenant and environment administrators, not necessarily to the owner of the agent. Whoever built it does not automatically hear about the limit. Decide who reads the alert and who calls the owner.

Enforcement arrives late. When a tenant reaches 125 per cent of its prepaid capacity, custom agents are disabled. Ongoing conversations are not cut off, every further invocation is rejected. That is an emergency brake, not a budget.

There is also a date: developer and trial environments move to usage-based billing on 1 September 2026. If your prototypes sit in developer environments, that is the day an experiment becomes a line on the invoice.

A date that otherwise slips past you

On 1 November 2026 the seeded AI Builder credits are removed, for new and existing customers alike, including those on an Enterprise Agreement. There is no automatic conversion into Copilot Credits. Until then a dual-mode model applies in which environments consume AI Builder credits first and Copilot Credits after that.

Read that again with your own landscape in mind: every AI Builder use that quietly runs on seeded credits today falls back to Copilot Credits on that day, or to nothing.

What I cannot assess

Whether the three guardian agents consume credits themselves, I found no documented answer. Microsoft describes them as a managed platform feature, which points towards included, but it is not spelled out.

And I cannot judge whether the approval step before execution stays permanent. It is unambiguous in the current description, but release plans describe intentions rather than commitments.

My recommendation

Treat Agentic CoE like a new colleague with read access to everything: it needs a manager. In practice that means one named person who reads the Actions page on a fixed rhythm, and a list of finding types you would approve a plan for at all. If that list stays empty, that is a valid answer and better than an approval given out of awkwardness.

Then three lines in your operations handbook. Who receives the limit notifications and who tells the owner. Who may raise a limit. And where the aggregate cap sits, if you need one.

Two dates in the calendar: 1 September for the developer environments and 1 November for the AI Builder credits.

I think the direction is right. Governance work that nobody does can be handed to a machine, and ranking by impact is exactly where inventory lists usually fail. My reservation: a system that hands you proposals daily creates approval routine. The audit trail shows afterwards who approved, not who read.

Sources

Power Platform · Governance & Compliance