Automate·Advanced·50 min·Updated Sep 30, 2026

Hand off a Power Automate step to a Copilot Studio agent

Use the Execute Agent action to route a reasoning-heavy step in a flow to a Copilot Studio agent instead of hard-coded logic.

Download PDF

Good fit for

Microsoft 365

Power Automate

Works With

Copilot Studio

Prerequisites

An existing Power Automate flow with a step that currently needs manual judgment

Business Outcome

Flows that can handle judgment calls (not just fixed rules) without a human in the loop for every case.

Workflow Overview

Power Automate

Copilot Studio

Step 1: Identify the judgment-call stepPower Automate

Open the flow and find the exact action where it currently pauses for a person to read something and decide — classifying an expense report's risk level before routing it, say.

If that decision could already be written as a Condition or Switch action using fields already in the data, it doesn't actually need an agent; the real candidate is a step where the input is free text or otherwise too unstructured for a fixed rule.

Step 2: Build a narrow Copilot Studio agent for that decisionCopilot Studio

In Copilot Studio, build an agent scoped to exactly that one decision, with clear instructions and a structured output format (a category, a confidence score, a short reason) rather than free text.

Step 3: Wire it into the flow with Execute AgentPower Automate

Add the "Execute Agent" action at that step in Power Automate, pass in the relevant record data, and map the agent's structured output fields to the flow's next conditions.

Prompt idea:

Classify this expense report as Low, Medium, or High risk based on amount, category, and whether it matches a pre-approved vendor list. Return only the risk level, a one-sentence reason, and a confidence score from 0-100.

Step 4: Add a low-confidence fallbackPower Automate

Route anything below a confidence threshold (for example, under 70) to a human reviewer instead of letting the flow act on a guess.

Check the work

Run the flow against known cases first and compare the agent's classification to what a human would have decided.
Confirm the low-confidence branch actually routes to a person, not a dead end.
Spot-check a sample of live runs weekly to catch drift in the agent's decisions.

Source: Rishona Power Platform, "Copilot Studio/Power Automate: Call an agent to run during a flow" (January 2026)

Expected Outcome

A Power Automate flow that calls a Copilot Studio agent mid-run and branches on its structured response.

Is AI actually needed here?

The example given — classifying risk by amount, category, and vendor-list match — is itself expressible as a Power Automate Switch/Condition action; those three factors are exactly the kind of fixed rule Power Automate was built to branch on, not a reasoning call that needs an agent.

Non-AI alternative: A Power Automate Condition or Switch action already branches on amount thresholds, category, and vendor-list membership directly — reserve Execute Agent for a step that genuinely can't be expressed as a fixed condition, like reading free-text context.