Microsoft Dataverse plugin: unleashing coding agents on the enterprise | OD849

Kent Weare and Suyash Kshirsagar show how a Dataverse “skills” plugin can make agent-driven Dataverse work more reliable by grounding agents in real Dataverse metadata (tables, columns, relationships, views, security, and solutions) and executing tasks autonomously.

Overview

Coding agents can be effective, but in enterprise systems they often fail without domain-specific tooling: they hallucinate schema, misunderstand security constraints, and generate workflows that don’t match the platform’s real capabilities. This session introduces a Dataverse plugin designed to reduce those failures by giving agents a constrained, metadata-driven interface to Dataverse.

The presenters describe and demo:

• A plugin that exposes Dataverse concepts to an AI agent with guardrails:
  • Tables, columns, and relationships
  • Views
  • Security model elements (business units, roles, profiles, sharing)
  • Solutions
• How a natural language request can trigger multi-step, autonomous execution, including:
  • Provisioning/setup steps
  • Data imports
  • Validation steps
• The plugin architecture and how it integrates with an MCP server.

Demo scenarios highlighted in the session

Natural-language operations against Dataverse

A user issues simple prompts to connect to CRM and perform Dataverse operations using the plugin, with the plugin providing the agent the necessary platform context (schema and constraints) to avoid guessing.

Task and data hygiene workflows

The session includes examples of using natural language to query work items (for example, open follow-up tasks for the week) and then perform cleanup actions to improve data hygiene.

Security model creation and validation

A key enterprise focus is using natural language to create and validate more complex Dataverse security models. The session calls out completing setup tasks such as:

• Business units
• Roles
• Profiles
• Validation
• Sharing

What the plugin is trying to solve

• Reduce hallucinations by grounding the agent in Dataverse metadata.
• Make agent-driven changes safer by constraining actions to what Dataverse actually supports.
• Enable reliable, repeatable multi-step execution for enterprise Dataverse development and administration tasks.