
Last quarter, I sat with an IT director at a 1,800-person manufacturer who had spent eighteen months on a Copilot governance plan. Sensitivity labels. Purview rolled out. SharePoint oversharing scrubbed. The whole program buttoned up before broad rollout. He was proud of it. He should be. Then he opened his Copilot Studio admin center and we counted together: 94 agents built by business users in the last 120 days. He had no owners mapped for 61 of them. Three were hitting Dynamics. One was pulling from a finance SharePoint site his labels never got to in time.
That's the governance story in 2026, and it's the one most mid-size orgs are missing. The Copilot governance plan you wrote a year ago was about Copilot Chat touching too many files. The one you need this quarter is about the agents your users are already building — quietly, in the background, each one a tiny application with its own data path, its own permissions, and its own blast radius if something goes wrong.
Twelve months ago, the governance question was: can we trust Copilot not to surface the HR folder nobody labeled? That was a real problem, and the Purview work most of you did was real work. But it was fundamentally a read problem — Copilot reading data it shouldn't see.
Agents flip the model. A Copilot Studio agent is not a reader. It's a maker. It can call connectors, hit external APIs, trigger Power Automate flows, write back to systems, and operate on behalf of a user who may or may not understand what they just published. Microsoft's own recent guidance makes this plain: with Entra Agent IDs now mandatory, authorization happens at the identity layer — the agent respects enterprise boundaries whether or not its prompt tells it to. That's a big step forward. It's also an admission of what the prior year of agent building looked like: prompt-based "please be careful" instructions standing in for actual controls.
If your governance plan still ends at Purview labels and tenant hygiene, you're one quarter behind. And this is the part that stings for mid-size orgs: you don't have a dedicated agent governance team. You have the same two or three people who did the Copilot rollout, now also responsible for a sprawling, user-generated application layer they didn't build and don't fully see.
Here's the pattern I'm seeing across the mid-size orgs we work with. Marketing builds a "campaign brief generator" agent. Finance builds something to pull AP aging. Someone in Operations wires an agent to a vendor's web API because they got tired of copying data. HR builds a policy Q&A bot and quietly points it at the full policy library — including the drafts.
None of these people are malicious. Most of them are your highest-performing employees. They're doing exactly what the Copilot pitch promised: automating their own work. The problem is that each agent is now:
Multiply that by 80, 150, 300 agents, and you have the same sprawl problem SharePoint created ten years ago — except SharePoint didn't take actions on your behalf. Agents do.
You don't need an enterprise AI governance platform to get a grip on this. You need four things, in this order.
Identity. Every agent in your tenant should have an Entra Agent ID, Conditional Access applied, and connector permissions surfaced as API permissions in Entra. If you're on E3 or E5, this is table stakes and the tooling is there — turn it on.
Access control. Stop sharing agents with individual users. Share them through Entra security groups, every time. The reason is operational, not philosophical: when the maker leaves or changes roles, group membership is how you cleanly transfer or retire the agent. Direct-share agents become orphans within a quarter.
Data protection. Your Purview labels and DLP policies should extend to agent connectors — not just to Copilot Chat. Most mid-size orgs discover their DLP rules weren't written with agents in mind. Audit them.
Observability and lifecycle. Build an agent registry. Every agent gets a named business owner on record and a quarterly review date. The agent doesn't exist until it's in the registry. If a maker won't put their name on it, that itself is the signal.
None of this requires a six-figure platform. The hard part isn't tooling. It's the operating cadence — who reviews the registry, how often, with what authority to retire agents that don't have an owner. That's a program decision, not a product one.
Open your Copilot Studio admin center on Monday morning and get the count. Just the count. How many agents exist in your tenant, built by whom, in the last ninety days. Most IT leaders I show this to don't have the number memorized — and the gap between what they thought and what the screen says is the entire argument for why agent governance belongs on the executive agenda this quarter.
Then do one thing: pick the ten agents with the most users and assign named business owners to them this week. Not IT owners — business owners. The person in the department who will defend the agent's existence when it comes up for review. That single act establishes the operating pattern. Everything else — the registry, the lifecycle policy, the audit — follows from the principle that no agent runs in your tenant without a human on the hook for it.
Governance isn't the opposite of adoption. It's what makes adoption durable — the agents that stick around are the ones somebody owns, defends, and keeps alive. The ones nobody owns were going to break something eventually. Better to find them now than after the audit.
This is piece two of the eight-pillar executive plan for Copilot adoption. If your enablement program is being designed around these agents — not just the license — that's where the real adoption curve bends.
Stop Pissing Off Your Software Users! There's a Better Way...
VisualSP makes in-app guidance simple.