Discover how to create practical technical guardrails for AI agents built by end users in Copilot Studio and other low-code agent platforms.
Shadow IT and shadow automation are not new problems, but AI agents change the scale and risk. Business users can now create assistants that read email, access business data, use connectors, create tickets, send messages, update records, and trigger workflows. These agents may start as personal productivity tools, but they can quickly become unmanaged automation with real permissions, real data, and real business impact.
In this session, we will explore how organisations can move from reactive case-by-case approval to a policy-driven guardrail model. Instead of relying on every user to make the right technical decision, IT and security teams need safe defaults, permission boundaries, connector policies, environment separation, approval requirements, monitoring, and clear ownership.
The focus is on practical technical policy: what should be allowed by default, what should require review, and what should be blocked entirely. We will cover common risk areas such as mailbox access, prompt injection through email content, over-permissioned connectors, maker-owned credentials, unmanaged actions, data leakage, weak ownership, and missing audit trails.
Using Microsoft Copilot Studio and Microsoft 365 as concrete examples, we will discuss guardrails that also apply to other agent platforms: Data Loss Prevention policies, connector classification, environment strategy, identity and credential models, action risk levels, human approval points, logging, lifecycle management, and exception handling.
The goal is not to stop users from creating agents. The goal is to make safe agent creation the default path. Participants will learn how to design a governance model where users can innovate inside clear technical boundaries, while IT and security teams retain visibility and control over data access, tools, actions, and risk.

After this session, participants will be able to:
Identify where user-created AI agents become a governance risk, especially when they can access email, use connectors, read business data, and take actions across systems.
Design practical technical guardrails for agents, including environment strategy, connector policies, permission boundaries, credential choices, approval requirements, monitoring, and lifecycle controls.
Apply a simple allow / restrict / review / block model to decide which agent capabilities should be enabled by default, which require approval, and which should be prevented entirely.

Level 200-300
45 minutes + questions