Power Apps MCP Server Adds Closed-Loop Learning

Introduction

Microsoft is adding a practical new capability to enterprise agents in Power Apps: closed-loop learning. For organizations using AI agents for structured tasks like invoice processing, this matters because agents can now learn directly from user corrections in production instead of relying only on static prompts, documents, or custom retraining workflows.

The feature is available for agents connected to the Power Apps MCP server, beginning with the data entry tool.

What’s new

Microsoft’s update introduces two learning mechanisms that work together:

• Memory-based optimization stores user corrections from the Agent feed as structured memory.
• On future runs, the agent retrieves those memories and applies them to similar tasks.
• Genetic-Pareto optimization then generalizes repeated corrections into broader rules that become part of the agent’s default behavior.
• The process runs automatically in production with no extra configuration, data pipeline, or manual ML workflow required.

A simple example: if a user repeatedly changes UK to United Kingdom during invoice processing, the agent learns that normalization pattern and starts applying it automatically. Over time, it can extend that logic to similar cases such as USA or DE.

Why it matters for enterprises

This is designed for organizations running agents at scale. Instead of each correction being a one-off fix, the system turns feedback into organization-wide improvements scoped to the tenant and business process.

That makes closed-loop learning different from consumer-style AI memory features, which usually personalize responses for one user. In Power Apps MCP server, the goal is to improve task accuracy for everyone using the agent.

Microsoft says the approach helped reduce manual edits in offline testing with the UK Electoral Commission:

• Fields requiring manual correction dropped from 64% to 48%
• 1,045 fewer fields required edits across 4,277 field instances
• F1 accuracy improved from 66.4% to 74.6% in 10 independent runs

The biggest gains came from business-specific formatting and extraction rules, such as:

• Expanding abbreviated country names
• Formatting UK addresses correctly
• Using legal entity names instead of brand labels
• Selecting gross invoice totals instead of partial or ex-VAT values

Impact on IT admins and makers

For Power Platform admins, this could lower the overhead of maintaining enterprise agents. Teams may be able to improve accuracy through normal user feedback rather than building separate evaluation and retraining processes.

For makers and business users, it means data entry agents should become more aligned with internal conventions over time, reducing repetitive corrections and improving consistency.

Next steps

• Review where your organization uses Power Apps agents for structured extraction tasks.
• Identify business processes where users frequently make repeat corrections.
• Monitor data entry scenarios to see whether closed-loop learning can reduce manual cleanup.
• Validate governance and tenant scoping requirements before broader rollout.

This update signals a broader shift in enterprise AI: agents that do not just respond, but continuously improve from real operational feedback.