MCP Lands in Copilot Studio — and Purview Catches Up

MCP Support in Copilot Studio Is the Week's Lead Story

The Microsoft 365 Roadmap entry for Copilot Studio MCP tools dropped on May 15 with a May preview date and a June GA target. Starting this month, makers can wire MCP-compliant servers directly into Copilot Studio workflows — proprietary APIs, dynamic knowledge sources, custom actions — without hand-rolling a custom connector for each one.

Why this matters now: every mid-market client we've talked to in the last quarter has the same problem. They have a working agent prototype in Copilot Studio, and they have a library of internal APIs that aren't in the Power Platform connector catalog. Closing that gap previously meant either a custom connector build (slow, brittle, connector-team bottleneck) or abandoning Copilot Studio for something heavier. MCP removes that choice. If your backend team can expose an MCP-compliant endpoint, the Copilot Studio maker can consume it without a connector deployment cycle.

What to do: Map your planned Copilot Studio agents against internal systems that lack Power Platform connectors. Prioritize the two or three where connector-build latency has been blocking your rollout timeline. Stand up an MCP server for those systems in the preview window so your governance team can evaluate it before June GA rather than after.

Dynamic Tool Discovery Removes the Redeploy Tax

One entry below on the roadmap, Dynamic Tool Discovery for Declarative Agents (published May 13, GA June 2026) solves a different but related frustration. Today, adding a capability to a declarative agent means republishing it — which in an enterprise tenant means going back through your change management and deployment pipeline.

With dynamic tool discovery, developers update the MCP server and users see the new capability immediately. The agent itself doesn't need a new version. Microsoft is explicit that enterprise RAI controls and admin deployment gates remain in place; what's removed is the redeploy-for-every-tool-update cycle, not the governance layer.

For clients running agents against fast-moving internal systems — a ticketing tool that adds fields quarterly, a data platform that rotates schemas — this is meaningful. The ops burden of keeping agents current drops considerably.

What to do: Before June GA, review which of your deployed or planned declarative agents pull tools from sources that change frequently. Those are the candidates to migrate to MCP-backed dynamic discovery first. Confirm with your security team that the existing agent deployment gates (user/group assignment) are configured to your policy before you enable dynamic tool updates.

Purview DLP for External Emails: A Gap You Probably Have Open

Microsoft Purview DLP for external email processing in Copilot and Copilot Chat hits preview in June and GA in July. The control is straightforward: Copilot and Chat will not process emails from external senders when the policy is enabled. Agents built in Copilot Studio and published to M365 Copilot are also in scope.

Most organizations that have deployed M365 Copilot to date have leaned on sensitivity label policies for document-level control and assumed email risk was lower. It's not. Copilot's ability to summarize and synthesize across the mailbox means a well-crafted inbound email from outside your tenant is now a potential prompt injection vector or a source of external data that shouldn't ground internal responses. This policy closes that surface.

What to do: Flag this for your Purview admin and your Copilot governance lead today, even though GA is July. The policy design decision — do you block all external email grounding, or do you scope it by department or data classification — needs to be made before the control lands, not after. Treat the preview window in June as your testing and stakeholder-alignment period.

GCC-High Orgs Finally Get Prompt-Level DLP

The Purview DLP for M365 Copilot in government clouds entry (published May 11, GA June 2026) is the most important signal this week for any client operating in GCC-High or GCC. Prompt-level DLP — blocking Copilot from responding when the prompt itself contains sensitive data, and preventing that data from being used as grounding — has been available in commercial cloud for some time. Government tenants have been waiting.

This matters beyond compliance checkbox territory. Agencies and contractors in GCC-High environments have staff who are accustomed to treating Copilot like a search box. Without prompt-level controls, a user can inadvertently paste CUI or export-controlled data into a prompt and have it used for web grounding. That's a data handling incident in most program office frameworks.

What to do: If you're managing a GCC-High Copilot deployment, get a sensitive information type inventory done before June. Prompt-level DLP is only as good as the SITs you've defined. If your SIT library reflects your commercial tenant's defaults and hasn't been tuned to your program's data taxonomy, you have roughly six weeks to fix that before the control goes live.

Copilot Studio Computer Use: Watch the October Timeline

The Computer Use capability for Copilot Studio has been in preview since April and is targeting October GA. It enables goal-oriented UI automation directly in workflows — agents that can operate web applications or desktop software without a formal API, adapting when the UI changes.

October is far enough out that it's not an immediate action item, but it's worth flagging to your RPA team now. If you have a process automation backlog where no API exists and Power Automate desktop flows have been too brittle to maintain, this is the path to watch. Admin governance controls (access controls, hosted browser management, environment-level permissions) are part of the GA spec, which suggests Microsoft has thought about the blast radius of an agent that can click through arbitrary UIs.

---

Next week, watch for any further GCC-High Copilot feature parity announcements — the government cloud catch-up cadence has accelerated and the PowerPoint Copilot entry for government clouds suggests more commercial-to-GCC parity is queued.