Customer story · Energy

Dozens of agents in production from 120+ citizen developers, none without a security review

We secured and governed Copilot Studio for an energy company, so 120+ makers could build agents without exposing its data.
Let’s talk
Sector
Energy
Makers
120+
Licensing
Microsoft 365 E5
Duration
4 months, 2025
Data residency
Canada

The situation

78%
of AI users bring their own AI tools to work
Microsoft Work Trend Index, 2024
97%
of organisations with an AI-related breach lacked proper AI access controls
IBM Cost of a Data Breach, 2025
US$670K
more per breach where shadow AI is high
IBM Cost of a Data Breach, 2025
  • Copilot Studio came with the Power Platform licences, and 120+ business users wanted to build agents.
  • The Power Platform team could run the platform. Nobody owned the security questions.
  • Which data can an agent reach, which systems can it call, who can talk to it, and who checks before it goes live?
  • Blocking it would push agents out of sight. Leaving it open would expose the data.

The architecture

Makers build only in the Personal Productivity environments. Connectors go through the governance page, and nothing reaches Prod without the review committee.

What we put in place

1. A dedicated place to build

Makers never build in the default environment.

  • Default environment locked
  • Personal Productivity Dev, Test and Prod
  • Promotion through Power Platform pipelines, never by hand

2. A governance page for connectors

Connectors carry most of the risk. Every request answers six questions before it is allowed:

  • Why the business needs it
  • What data it touches, and its sensitivity label
  • Whether data leaves the tenant or Canada
  • How it authenticates: user or service account
  • Who owns it
  • Which environment it is allowed in

3. A review before production

Every maker can build. Nothing reaches Prod until the Copilot admin and the security admin check:

  • Authentication
  • Knowledge sources
  • Sharing scope
  • Connectors used
  • Data sensitivity

4. Identity and data rules

  • Entra ID on every agent
  • SharePoint first, other sources on need
  • Purview sensitivity labels
  • Teams and Microsoft 365 Copilot only
  • Data stays in Canada
  • Power Automate, custom connectors and MCP case by case

5. Visibility after go-live

  • Defender for Cloud Apps watches agents and prompts
  • Purview audit logs flow into Sentinel, the SIEM
  • An inventory of every agent and its owner

6. Makers who know the rules

  • A maker guide
  • Training during reviews, on the maker’s own agent

Checked against Microsoft’s own list

In February 2026, Microsoft’s Defender research team listed the ten most common Copilot Studio agent misconfigurations. This design catches eight.

MisconfigurationWhere this design catches it
Agent shared with the whole organisationSharing scope is set at the review
No authentication requiredEntra ID required, checked at the review
Risky HTTP Request actionsEvery connector goes through the governance page
Email-based data exfiltration“Does data leave the tenant?” is a required answer
Agent runs with the maker’s credentialsUser or service account is declared per connector
MCP tools configuredMCP allowed case by case, through the governance page
Dormant connections, actions or agentsThe agent inventory shows what is still in use
Orphaned agent with no ownerEvery agent and connector has a named owner

Not covered: hard-coded credentials and generative orchestration without instructions. Both are next on the review checklist.

What changed

  • Dozens of agents in production, all through the review
  • Each one with a known owner, known connectors and a known audience
  • Published only to Teams and Microsoft 365 Copilot
  • Shorter reviews: the hard questions were answered before the committee asked
What we learned

The connector template, built as a gate, became a design tool: makers thought about security while building.

If you are in the same situation

If Copilot Studio came with your licences and your teams want to build, this is what we bring, adapted to your size and your data:

  • An environment strategy, with the default environment locked and a pipeline to production
  • The connector governance page, built on your own data classification
  • A pre-production review checklist, and who sits on the review
  • Identity, knowledge-source and labelling rules: Entra ID, SharePoint, Purview
  • Monitoring connected to what you already run: Defender for Cloud Apps, Purview audit, Sentinel
  • A maker guide and training delivered during reviews
  • An agent inventory with an owner for every agent

Four months for 120+ makers. A smaller organisation needs the same controls with one reviewer and a checklist.

Opening Copilot Studio to your teams?

Tell us where your teams want to build. We will tell you what to lock down first, and what it would take.

Let’s talkAI Agents

The stack

Build and deploy

  • Microsoft Copilot Studio
  • Power Platform environments
  • Power Platform pipelines

Actions and integrations

  • Power Automate
  • Custom connectors
  • Model Context Protocol (MCP)

Identity and data

  • Microsoft Entra ID
  • SharePoint
  • Microsoft Purview sensitivity labels

Monitoring

  • Microsoft Defender for Cloud Apps
  • Microsoft Purview Audit
  • Microsoft Sentinel

Channels

  • Microsoft Teams
  • Microsoft 365 Copilot

Sources

  1. Microsoft and LinkedIn, 2024 Work Trend Index: AI at Work Is Here. Now Comes the Hard Part (May 2024)
  2. IBM, Cost of a Data Breach Report 2025 (July 2025) Figures as reported in IBM’s announcement of the report and consistent across independent coverage.
  3. Microsoft Defender Security Research Team, Detecting and mitigating common agent misconfigurations (February 2026)