DE

Microsoft Best Practices - AI Agents

AI apps and agents secure by design: Microsoft Purview as the control layer

Practical article on Microsoft Purview, DLP, audit, DSPM and protection mechanisms for AI apps, agents and developer workflows.

Start Quick CheckBack to best-practice library

Microsoft Purview enables developers with strong data security across AI apps and agents

Context

AI agents do not only access knowledge. They read files, use tools, execute actions and connect systems. Security therefore shifts from simple app approval to a harder question: which data may the agent see, use and pass on?

Typical scenario

A team builds an internal agent for support and process questions. At first, it only answers from SharePoint. Later it should create tickets, process form data and work with local developer tools. Without runtime controls, DLP and audit, an assistant becomes a process actor that is difficult to monitor.

Technical implementation steps

  1. Create an agent profile for every agent: purpose, owner, data sources, actions, audience and approval status.
  2. Separate environments in the Power Platform admin center: do not mix development, test and production.
  3. Define DLP policies for connector groups and review risky combinations such as company data plus public consumer services.
  4. Limit knowledge sources: add SharePoint sites, files and systems only when ownership and data classification are clear.
  5. Create test cases: wrong question, sensitive data, external user, prompt injection and unauthorized action.
  6. Enable audit, monitoring and escalation paths before the agent performs production actions.
  7. Define a lifecycle: owner review, source review, answer quality, DLP matches and retirement of unused agents.

Microsoft best practices in implementation

  • Document every agent with owner, purpose, data sources and permitted actions.
  • Plan DLP rules for prompts, responses, tool calls and sensitive content early.
  • Ensure audit and investigation capability before agents trigger production processes.
  • Assess local and cloud-based agents separately because local agents use different data paths.
  • Bring security and development teams into one approval flow.

Common mistakes

  • Treating agents like normal chatbots.
  • Checking only knowledge sources while ignoring actions and connectors.
  • Enabling audit only after an incident.

Azuric perspective

Azuric would deliberately start small: one clear use case, controlled data sources, no critical actions without approval and a measurable security model.

Key takeaway

AI agents need governance before scale. Purview can bridge development, security and compliance.

Sources

This article is an original Azuric perspective. The following sources are used as technical references; content is not copied.