Workflows

Build an autonomous agent using Workflows — the next generation of autonomous agents in Copilot Studio — with an event-driven trigger and a non-deterministic inline agent that books calendar time and updates a to-do.

Level 300 60 min advanced Report an issue

Workflows

Build an autonomous agent in Microsoft Copilot Studio using Workflows — an agent that runs on a trigger instead of waiting for a chat, and hands real work to a non-deterministic inline agent.

Lab Details

Level	Persona	Duration	Purpose
300	Maker	60 minutes	After completing this lab, participants will be able to create a Workflow in Copilot Studio, fire it from an event-driven trigger, and embed a non-deterministic inline agent that reasons over the trigger’s data and acts through tools (Outlook calendar, Microsoft To Do, and web search) — going beyond the classic, deterministic autonomous-agent model.

Why This Matters
Introduction
Core Concepts Overview
Documentation and Additional Training Links
Prerequisites
Summary of Targets
Use Cases Covered
Instructions by Use Case
Summary of Learnings
Conclusions & Recommendations

Why This Matters

Makers and Developers — want your agents to do work on their own, in the background, the moment something happens — instead of only answering when a user types a message?

Think of Workflows as the autonomous side of Copilot Studio:

Classic agents are conversational — they wait for a user turn, then respond.
Workflows are event-driven — they start from a trigger (a schedule, or a change in an external system) and run to completion with no one watching.
The new agent node lets a Workflow hand a step to a reasoning agent rather than a fixed, deterministic action — so the automation can decide how to accomplish a goal, not just replay hard-coded steps.

Common challenges solved by this lab:

“I want automation that kicks off when a record changes, not when someone opens a chat.”
“My flows are rigid — I need a step that can reason about messy, real-world inputs and choose what to do.”
“I want an agent that can read my calendar, book time, and update a task — on its own.”

In 60 minutes, you’ll build autonomous Workflows that act on their own when something happens — an inline agent that blocks focus time and enriches a task, a multi-branch order-management workflow with human-in-the-loop approval, an MCP-powered inventory agent that checks stock and creates Dataverse tasks, and a workflow that calls a published price-quote specialist agent.

Introduction

Workflows are the next generation of autonomous agents in Microsoft Copilot Studio. Where a classic agent waits to be asked, a Workflow listens for a trigger and then executes — making it ideal for “when X happens, do Y” automation across Microsoft 365 and hundreds of connectors. What makes the new experience powerful is the agent node: a Workflow can now delegate a step to an inline, non-deterministic agent that reasons over the trigger’s data and calls tools to get work done, rather than following a fixed script.

Real-world example: You keep a running list of things you need to do, but you never actually block time to do them. In this lab you’ll build a Workflow that watches a Microsoft To Do list. The moment you add a task, the Workflow’s inline agent reads the task, checks your Outlook calendar for an appropriate slot within your working hours, books a focus block, and then updates the to-do with a due date, a reminder, helpful notes, and a category — all autonomously. You add one task; the agent does the planning.

This example is covered in Use Case 1 of this lab. It establishes the foundation — trigger plus inline agent plus tools — that the later use cases build on.

Core Concepts Overview

Concept	Why it matters
Workflow	The autonomous unit in Copilot Studio. It runs on a trigger instead of a chat turn, so work happens in the background without a user present.
Trigger	What starts a Workflow. Triggers can be manual, scheduled (recurrence), or connector-based (an event in an external service, like a new Microsoft To Do task).
Inline agent (agent node)	A non-deterministic agent embedded directly in a Workflow step. Instead of a fixed action, it reasons over the input and decides how to achieve the goal.
Dynamic content / `/` token	A reference to data produced earlier in the Workflow (e.g., the trigger’s to-do Title), inserted into the agent’s instructions so it acts on the actual item at run time.
Tools	The capabilities you grant the agent so it can take action — here the Work IQ Calendar MCP (Outlook calendar), the Update to-do action, and Web search.

Documentation and Additional Training Links

Prerequisites

Access to Microsoft Copilot Studio with appropriate licensing, in an environment where Workflows are enabled.
A work or school Microsoft 365 account with Outlook (calendar) and Microsoft To Do provisioned — the inline agent reads and writes this account’s calendar and tasks.
A Power Platform environment where you can edit Dataverse table data and toggle environment settings (System Administrator or System Customizer).
Sample data pre-loaded into the Dataverse tables used by the order-management use cases (“Tasks” table).
Permission to create connections for the services used across the lab: Microsoft To-Do (Business), Work IQ Calendar / Work IQ, Office 365 Outlook, Microsoft Dataverse / Dataverse MCP, Human Review, and Microsoft 365 Copilot.
Basic familiarity with the Copilot Studio interface.

[!IMPORTANT] Use the same work account for Copilot Studio and every connected service in this lab (Microsoft To Do, Outlook, Dataverse, M365 Copilot, and approvals). The triggers, inline agents, and tools all act on one identity — if they don’t match, a workflow can’t see your data or write back to your calendar, tasks, or Dataverse records.

Summary of Targets

In this lab, you’ll build several autonomous Workflows that act on event-driven triggers and hand real work to inline and published agents. By the end, you will be able to:

Create a Workflow and configure manual, scheduled, and connector-based triggers.
Embed a non-deterministic inline agent in a workflow, give it goal-oriented instructions, and reference earlier data with the / dynamic-content token.
Equip agents with tools across Microsoft 365 and Dataverse (Work IQ Calendar, Update to-do, Dataverse MCP, web search) and use M365 Copilot inside a workflow.
Add a human-in-the-loop approval step so a person validates an AI-drafted action before the workflow continues.
Publish, run, and monitor workflows end-to-end — including calling a separate published agent from a workflow branch.

Use Cases Covered

Step	Use Case	Value added	Effort
1	Automate Task Time-Blocking with a Workflow and an Inline Agent	Build an autonomous, trigger-driven Workflow whose inline agent reasons over a task and acts across your calendar and to-do list	15 min
2	Setting Up the Order Management Workflow	Configure connections, ownership, and publish a pre-built multi-branch classification workflow	15 min
3	Use M365 Copilot and Add a Human-in-the-Loop	Validate the Customer Inquiry path: M365 Copilot drafts a response, human approves, workflow replies	15 min
4	Build an Inline Agent for Inventory Management	Add an MCP-powered inline agent that checks warehouse stock and creates Dataverse tasks	15 min
5	Call a Price Quote Specialist Agent from a Workflow (Bonus)	Wire a published agent into a workflow branch to generate and send price quotes	10 min (extra)

Instructions by Use Case

Use Case #1: Automate Task Time-Blocking with a Workflow and an Inline Agent

Build a Workflow that watches a Microsoft To Do list and, whenever a task is added, lets an inline agent autonomously block focus time on your calendar and enrich the task.

Summary of tasks

In this section, you’ll create a To Do list to watch, build a Workflow with a To Do trigger, add a non-deterministic inline agent with goal-oriented instructions and tools, publish it, and then trigger it with a real task and verify the results.

Scenario: You capture work in Microsoft To Do but never actually reserve time to do it. You want an assistant that, the moment you jot down a task, finds an appropriate slot in your workday, books a calendar block, and updates the task with a due date, reminder, notes, and category — without you lifting a finger beyond adding the task.

Objective

Create and publish the To Do Time Block Workflow: a To Do trigger feeds a new task to an inline agent that schedules a calendar block within your working hours and updates the to-do — then run it end-to-end and confirm the calendar event and the enriched task.

Step-by-step instructions

Create the To-Do list that drives the workflow

In Microsoft To Do (to-do.office.com), select New list in the left sidebar, name it Time Block, and press Enter to create it. This is the list the workflow will watch — adding a task here is what triggers the automation.

[!TIP] Sign in to To Do with the same account you use for Copilot Studio. The workflow’s trigger connects to your To Do, so the list and the workflow must live under one identity.

Create the workflow and configure its trigger

In Copilot Studio, make sure you’re in the right environment, then on the Home page select Workflow (Automate business processes with agents and actions). Rename the workflow from Untitled Workflow to To Do Time Block by selecting the title at the top of the designer.

[!TIP] Workflows are the autonomous side of Copilot Studio — they run on a trigger instead of waiting for a user to chat. You’ll find them in the left navigation next to Agents.
Select the Start trigger node. In the configuration panel, change Trigger type from Manual to Connector, then choose Select trigger… and search for To Do. Under Microsoft To-Do (Business), choose When a new to-do in a specific folder is created. This makes the workflow fire automatically whenever a task lands in the list you specify.

[!IMPORTANT] Pick the (Business) connector, not (Consumer) — the lab uses your work/school To Do account, which is where Copilot Studio and the calendar tools also run.
When prompted, select Create new connection and, in the account picker, sign in with the same work account you’re using for this lab. Once the connection shows your account, set To-do List to Time Block from the dropdown.

[!WARNING] In the connection account picker, double-check you pick your lab account — not a personal or other work account. The workflow will read and write that account’s To Do and calendar.

Add and configure the inline agent

Select the + after the trigger and choose Agent to add an agent node. (Copilot Studio automatically wraps it in an Apply to each loop so the agent runs once per new to-do.) When prompted, Create new connection with your lab account. Leave the Agent dropdown on New agent for this workflow — this creates a brand-new, non-deterministic agent dedicated to this workflow rather than reusing an existing one.

[!TIP] This is the heart of the lab: instead of hard-coded actions, the workflow hands the task to an agent that reasons about how to schedule it. The default model is Claude Sonnet 4.6.
In the Instructions box, give the agent its goal. Enter it in three parts so the agent receives the actual to-do title at run time:
1. Type the first line of text:
```
Find a time that is appropriate for this task 
```
2. Type / to open the dynamic-content picker, then select Title (under When a new to-do in a specific folder is created). This inserts the to-do’s title as a token.
3. Continue typing the rest of the instruction right after the token:
```
on my calendar in the next couple of business days and block that time on my calendar to work on this task. Then update the to-do with a due date a week out from the time block we created. Add appropriate notes and suggestions in the notes on key things I might want to consider and categorize it as blue. Also set a reminder 2 days before the due date. Make sure that the format of the Calendar notes and To-Do are formatted properly for the user to be able to read them well. Only schedule the time block within my normal working hours of 8:00 AM to 5:00 PM.
```
[!IMPORTANT] The /Title token is what makes this dynamic — the agent receives each new task’s real title, so its scheduling decision is grounded in the actual to-do rather than a fixed string. The final sentence constrains when the agent books focus time; without it, the agent may pick the first technically-free slot (even very early morning).
Give the agent the tools it needs to act. In the agent panel:
- Under Tools, select + (Add tool), search for Work IQ Calendar, choose Work IQ Calendar (Preview) — the MCP server for Outlook calendar operations — and Create new connection with your lab account, then Add.
- Add a second tool: + → search Update to-do → under Microsoft To-Do (Business) choose Update to-do → Add (it reuses your To Do connection).
- Turn on the Web search toggle so the agent can look things up (e.g., venue or location details referenced in a task).
[!TIP] The Work IQ Calendar tool lets the agent read your free/busy and create calendar events; Update to-do lets it write the due date, notes, category, and reminder back onto the task.

Publish, then trigger the workflow

Select Save, then Publish.

[!WARNING] Publishing is required. A workflow only listens for its trigger once it has been published — so it will not run from a new to-do until you publish it. If you add the to-do before publishing, nothing happens. After publishing, the Activity and Monitor tabs become available.
Open Microsoft To Do, go to the Time Block list, and add a new task: Book Skull’s Rainbow Room in Nashville for Next Friday. Creating this task in the watched list is what fires the workflow’s trigger.

[!TIP] Use a task whose title implies a real-world commitment with a time and place — it gives the agent something concrete to reason about (find a slot, block the calendar, look up the venue).

Watch the workflow run

Back in the workflow, open the Activity tab and select Refresh to look for the run. When the new run appears (status Running), select it to open the run details and watch the workflow execute node-by-node (trigger → Apply to each → Agent).

[!IMPORTANT] Be patient — the To Do trigger polls on a schedule, so the run can take a few minutes to appear. Press Refresh in the Activity list a few times until it shows up.
In the run details, select the Agent node to expand its details. You’ll see the agent’s own narration of what it did — checking your free/busy, choosing a time block inside your working hours, creating the calendar event, finding the to-do, and updating it. This is the non-deterministic part: the agent decided the specifics, they aren’t hard-coded in the workflow.

[!TIP] The Agent’s Outputs / Response is a great teaching moment — it shows the model reasoning step-by-step and calling the Work IQ Calendar and Update to-do tools on its own.

Verify the results

Go back to Microsoft To Do and refresh the Time Block list. Open your task — it has been updated by the agent: a due date (a week out), a reminder set 2 days before the due date, a blue category, and richly formatted notes (target date, the scheduled work block, and an action checklist).

[!TIP] Switch the list to List view and open the task to see the full notes the agent wrote — it even researched the venue with web search.
Open Outlook → Calendar. The agent has added a time block event to reserve time to work on the task — on the next business day it found free and within your 8:00 AM–5:00 PM working hours — with a body summarizing the target date and venue. The workflow ran end-to-end with no further input from you.

Congratulations! You’ve built an autonomous Workflow with a reasoning inline agent.

Use Case #2: Setting Up the Order Management Workflow

Configure and publish a pre-built Order Management Workflow that classifies incoming emails and routes them through different processing paths — then verify the setup end-to-end by testing the “Other” classification path.

Summary of tasks

In this section, you’ll configure connection references in the Power Apps solution, transfer workflow ownership, open the workflow to verify canvas-level connections, explore the email classification categories, publish, and verify everything works by sending a test email that triggers the “Other” classification path.

Scenario: Your environment contains a pre-built Order Management Workflow that automatically classifies incoming emails with “Order Management” in the subject. It routes emails into one of four categories — Quote Request, Supplier Delay, Customer Inquiry, or Other — and takes different automated actions for each. Before the workflow can run, you need to configure all its connections and link them in the solution.

Objective

Complete the setup of the Order Management Workflow: configure all solution connection references, transfer ownership, verify canvas-level connections, explore the classification categories, publish, and validate end-to-end by sending a test email through the “Other” path.

[!IMPORTANT] This use case is the foundation for all subsequent use cases in this module. The connections, ownership, and publishing you complete here are required for the Customer Inquiry, Quote Request, and Supplier Delay paths covered later. Take extra care during setup — if a connection reference is missed or ownership isn’t transferred correctly, later use cases will fail silently or with hard-to-diagnose errors.

Step-by-step instructions

Take ownership and configure the solution

[!IMPORTANT] Complete the solution-level setup first. The steps below configure the node connections, link the connection references in the Power Apps solution, and transfer ownership — all of which must be done before the workflow can be published or tested. Skipping or rushing through these steps is the most common cause of “publish failed” errors later. Read each note carefully.

In Copilot Studio, navigate to Workflows and open the Order Management Workflow.

[!NOTE] The workflow has been pre-built for your lab environment. You’ll see a trigger node (“When a new email arrives”), a Classify node, and several branch paths — one for each email category.
Open each node on the canvas by clicking on it so its connection can resolve. Start with the trigger: click When a new email arrives and, under Connections, wait for the connection to load — it should update automatically to your lab user. Most other nodes resolve the same way. Two nodes need a connection created manually:
- M365 Copilot node — open the node and, under Connections, select Create new connection > Create, then sign in with your lab account.
- Human review node — open the node and, under Connections, select Create new connection > Create, then sign in with your lab account.
Open Power Apps (make.powerapps.com), ensure you are in the correct environment, navigate to Solutions, and open the LAB: Order Management solution. In the left pane, select Objects, then open Connection References. For each connection reference, select Edit and choose the connection you just created from the dropdown (it should now appear).
- You should see five connection references. If a sixth one is present whose connection id starts with crc3b_draft_bRURqJ.cr.shared_a365outlookmailmcp, it is not needed for the workflow to run and can be safely removed from the solution.
- For all others, select the connection and click Save, then confirm with Save changes.
[!WARNING] Always sign in with your lab account when creating connections — not a personal or different work account. All connections in this workflow must use the same identity, or the workflow will fail at runtime with permissions errors.
Navigate to the Cloud Flow object (the Order Management Workflow) inside the solution and select the flow. Choose Set primary owner, remove the existing owner, set yourself (your lab user) as the new owner, and select Save. A green bar should appear confirming the owner was successfully changed.

[!IMPORTANT] Without the ownership change, publishing is disabled by default, so the workflow cannot be triggered end-to-end — you can still test individual nodes inside it, but no real incoming email will start a run. Once you own the workflow, the Publish button becomes available. Ownership must be transferred after the connection references are configured.
Go back to the workflow, select Save, and Publish. Verify the workflow was successfully published after changing the ownership — publishing only succeeds once you own the flow.

Open and explore the Order Management Workflow

Select the trigger node (“When a new email arrives”). Next to Subject filter, select the code symbol </> (Switch to expression mode) to reveal the expression @{string('Order Management')}. This means the workflow only fires for emails whose subject contains “Order Management” — all other emails are ignored.

[!IMPORTANT] Keep this filter in mind when testing: every test email you send must include “Order Management” in the subject line, or the workflow will not trigger. This is a common source of confusion when tests “don’t seem to work.”
Select the Classify node. Review the four classification categories configured in the node. Each category has a name and example text that the AI model uses to determine where an incoming email should be routed:
- Quote Request — emails requesting pricing or quotes for products/services
- Supplier Delay — notifications about delays from suppliers
- Customer Inquiry — general questions from customers about orders, shipping, etc.
- Other — anything that doesn’t fit the above categories (spam, promotions, irrelevant)
[!TIP] Improving classification accuracy: You can add multiple examples to each category to handle different phrasings and edge cases. For instance, a “Quote Request” might come as “Can you send me pricing for 500 units?” or “We’d like a formal quotation for the attached spec.” Adding diverse examples helps the model generalize better. A good testing exercise is to try common cases and edge cases using the Test tab in the Classify node, observe how categories get assigned, and refine the category descriptions accordingly.

Test the classification

Test the classification before sending a real email. In the When a new email arrives section, find Body (The body of the message), click into the box, and paste:
```
Hi, I would like to request a quote for 500 units of Product A.
```
Select Run test and observe the predicted category — it should classify as Quote Request.

Test the “Other” classification path end-to-end

Open Outlook (outlook.office.com) and compose a new email to your lab user account with the following:
- Subject: Order Management - Congratulations! Your order desk has been selected
- Body:
```
Dear Order Manager,

Congratulations! Your order management team has been selected in our monthly
business draw and is entitled to claim an exclusive cash reward.

To release your reward, please reply with your full name and contact details,
and a small processing fee may apply.

Act soon, as unclaimed rewards expire.

Regards,
Promotions Team
MegaDraw Rewards
```
Send the email. This spam/scam email should be classified as Other by the workflow.

[!NOTE] The subject must contain “Order Management” to match the trigger’s subject filter. The body content is intentionally irrelevant to quotes, suppliers, or customer inquiries — it should fall through to the “Other” category.
Wait a few seconds for the workflow to trigger and process the email. Then navigate to your Archive folder in Outlook. You should see the test email has been automatically moved there — confirming the “Other” classification path is working correctly.

[!TIP] If the email doesn’t appear in the Archive folder, check the Activity tab in the workflow in Copilot Studio to see if the run triggered. If no run appears, verify the workflow is Published (not Draft) and that the email subject includes “Order Management.”

The test email moved to the Archive folder by the workflow

Congratulations! The Order Management Workflow is now fully configured and operational.

You’ve completed the foundation that all subsequent use cases build on — the workflow is published, connections are linked, and you’ve verified the “Other” classification path works end-to-end.

Use Case #3: Use M365 Copilot and Add a Human-in-the-Loop in Order Management

Configure and validate the Customer Inquiry branch so the workflow drafts a response with M365 Copilot, pauses for a Human review, and only replies to the customer after approval.

Summary of tasks

In this section, you’ll inspect the Customer Inquiry routing branch, review how the M365 Copilot and Human review nodes are configured, publish any updates if needed, and run an end-to-end test with an approval email.

Scenario: A customer sends a product question to the order desk. Instead of crafting the answer manually every time, the workflow uses M365 Copilot to prepare a grounded draft response, sends that draft to a human reviewer for approval, and then either replies to the customer or flags the message for follow-up depending on the approver’s decision.

Objective

Validate the Customer Inquiry path of the Order Management Workflow: confirm the M365 Copilot draft prompt, inspect the Human review approval experience, test the path with a sample customer email, approve the proposed reply, and verify the customer response lands in your inbox.

Step-by-step instructions

Inspect the Customer Inquiry branch

In Copilot Studio, open the published Order Management Workflow and select the Classify node. Observe the Customer Inquiry branch that leaves the classification node and note how it routes customer questions into a draft-and-review sequence.
Select the M365 Copilot node in that branch. Review the prompt and confirm it uses the Customer question input based on the incoming email body. This is the grounded draft-generation step that prepares the proposed response for the reviewer. The prompt asks M365 Copilot to read the customer’s question, search the existing email threads of the user that owns the node connection (you) for any prior context or earlier reply on the same topic, and then draft a complete, ready-to-send reply — answering clearly when it finds an answer, or writing a brief holding response when it can’t (without inventing product facts), always addressed “Dear Customer,” and signed off as “Contoso Electronics Order Management”.

[!NOTE] M365 Copilot node scope: The M365 Copilot node excels at addressing questions that leverage Microsoft Graph (emails, Teams chats, etc.) — think of it as prompts you would send in M365 Copilot chat. Note that the operations this node performs are read-only — it can search and retrieve information from emails, chats, and files, but it cannot send emails or Teams messages. If you need those write operations, use an Agent node with access to the relevant Work IQ tools instead.

[!NOTE] Fresh lab accounts have no history to ground on: Because this lab uses a freshly provisioned account with no prior emails or Teams messages, the M365 Copilot node won’t find real substance to build its answer from — so expect a generic holding reply. This step is illustrative: in a real mailbox with genuine history, M365 Copilot would ground the draft in actual prior threads and context.
Select the Human review node and inspect its configuration carefully:
- In the message, confirm the proposed reply drafted by M365 Copilot is inserted for the reviewer to read.
- Notice that the review request is delivered over Outlook — observe how we use an Outlook email to ask a person to approve before the workflow continues.
- The reviewer responds through a single Yes/No input (Send proposed reply from M365?). Select + to expand and see the other input types available — text, email, number, date, and more. You don’t need to add any here, but it’s worth understanding the variety of formatted inputs you can request and then reuse cleanly in downstream nodes.
Select the If/Else node after Human review and observe the decision logic. When the reviewer answers Yes, the workflow sends the proposed reply to the customer; otherwise, the original email is flagged in the inbox for manual follow-up.

Run an end-to-end Customer Inquiry test

Open Outlook (outlook.office.com) and send the following email to your lab user account:

Subject: Order Management - Question about iPad Air warranty and MDM

Body:

Hi team,

Quick question before we expand our last order. What is the standard warranty
period on the iPad Air (M2), and can the devices be enrolled in our mobile device
management (MDM)? I think we asked something similar a few months back but I can't
find the reply.

Thanks,
Jordan Kim
IT, Alpine Ski House

The customer inquiry test email being composed in Outlook

Return to the workflow and open the Activity panel. Refresh until the new run appears, then open it and observe the path: Classify → Customer Inquiry.

[!IMPORTANT] This path pauses at the approval step. If the run seems to stop, that is expected — the workflow is waiting for the Human review action to be completed from Outlook.
Wait until the run reaches the Human Request node. Once M365 Copilot finishes exploring your emails and drafting an appropriate response, a green check appears on the node and you should be notified of an incoming review email within a matter of seconds.
Go to Outlook and open the incoming approval email. Review the proposed M365 Copilot reply, then select Yes to approve sending it.

[!TIP] If you don’t see the approval email, try refreshing Outlook. If it still doesn’t land, go back to the workflow, open the Build tab, re-establish the connection on the Human review node by creating a new one, then select Save > Publish and run the test again.
Return to your inbox and verify that the approved reply email lands there. This confirms the workflow resumed after human approval and sent the response automatically.

Congratulations! The Customer Inquiry path is now drafting, reviewing, and replying successfully.

You’ve confirmed that the workflow can generate a grounded response with M365 Copilot, pause for human approval, and then continue automatically based on the reviewer’s decision.

Use Case #4: Build an Inline Agent for Inventory Management

Extend the Supplier Delay branch with an inline Agent that uses MCP servers to reason over the delay notice, check current stock, and create a follow-up Dataverse Task.

Summary of tasks

In this section, you’ll add a new agent node to the Supplier Delay branch, connect the Warehouse and Dataverse MCP servers, define the agent’s instructions and structured output, publish the workflow, and test the scenario with a supplier delay email.

Scenario: A supplier warns that an incoming shipment is delayed. The workflow shouldn’t just forward the email — it should reason about the impact by checking current inventory, determine urgency, and create a structured Dataverse task so the replenishment team can act immediately.

Objective

Build and validate the Supplier Delay path of the Order Management Workflow by adding the Inventory Task Agent, equipping it with MCP tools, generating structured output, and verifying that it creates the expected Dataverse task after a test email is received.

Step-by-step instructions

Add and configure the Supplier Delay agent

In the Classify node, locate the Supplier Delay category. Select the + next to that branch and choose Agent to add a new agent step for this path.
In the new node, leave Agent set to New agent in this workflow. This creates an inline agent dedicated to this workflow path. Keep the default AI model.

[!TIP] A strength of AI-native nodes in Workflows is that you can choose a different model per node. This inline agent runs on Claude Sonnet 4.6 for its reasoning, while the Classify node uses the lighter GPT-4.1 mini — a cost-optimized choice for that lightweight classification step. Matching the model to the work each node does keeps the workflow both capable and economical.
Select Expand (the two arrows next to …) so you can edit the full agent configuration, then rename the node title from Agent to Inventory Task Agent.
Under Tools, navigate to Model Context Protocol (MCP servers), then add and connect both of these tools:
- Microsoft Dataverse MCP Server
- Warehouse MCP
Sign in or connect as prompted so both tools show as available in the agent.

In the Instructions box, paste the following agent instructions:

Purpose
You triage supplier shipment-delay notices into restock tasks for Contoso Electronics. A delay notice names a delayed product but does not include stock levels, so you use the Warehouse tool to find the current stock and judge how urgent the delay is, then record a task in Dataverse.

Inputs
Delay notice: [email body]
Review by (Due Date): [due date]

Guidelines
Take the current stock and all restock details from the Warehouse tool; the email does not contain them.
Judge urgency from the current on-hand quantity: zero on-hand means the delay causes a stockout and is urgent; low but above zero means the buffer is at risk; healthy stock means the delay is tolerable.

Steps
Read the product or SKU named in the delay notice.
Call the Warehouse restock-date lookup once, passing the exact SKU from the notice (no extra spaces). It returns the current quantity, next shipment date, expected quantity, and supplier. Do not call the stock-check tool.
Decide the urgency of the delay from the current on-hand quantity.
Use Dataverse to create one row in the Task table in the current environment, setting only these columns:
Subject: Restock delay - [product] ([SKU]).
Description: the current on-hand quantity and warehouse; the delayed next shipment date, the expected quantity, and the supplier; the urgency judgement; and a one-line action for the replenishment team.
Priority: High if the on-hand quantity is zero, otherwise Normal.

Rules
Create exactly one Task row and set only the columns listed above.
If the product is not found in the Warehouse, record it as unrecognised for a person to check.

The Supplier Delay agent instructions pasted into the configuration panel

Replace the placeholders in the instructions with dynamic content:
1. Select [email body], choose the lightning icon, search for body, and insert Body from When a new email arrives.
2. Select [due date], choose the lightning icon, then look above the search bar for the diamonds icon — Ask Copilot to generate an expression.
3. Enter: Add 3 days to the date and time of arrival of the trigger email
4. Confirm Copilot returns:
```
addDays(triggerOutputs()?['body/receivedDateTime'], 3)
```
  Select Insert.
Scroll to the Output section and change the dropdown to Structured output. Add these three properties:
- sku (string) — The SKU named in the delay notice
- stock (number) — Current stock quantity from the warehouse for the SKU named
- risk (boolean) — True if current stock is zero
[!TIP] “Request human assistance when unsure”: Notice the button in the agent configuration panel. This feature sends a structured message to the connection owner when the agent encounters ambiguity or conflicting information — it dynamically generates questions based on the problem at hand, allowing human judgment for edge cases while keeping the automation running.
Minimize the view, then select Save and Publish so the Supplier Delay branch is live with the new agent configuration.

Test the Supplier Delay path

In Outlook, send the following email to your lab user account:

Subject: Order Management - Shipment delay - LumiRead E-Reader

Body:

Hello Contoso,

We're writing to let you know that the upcoming shipment of the LumiRead E-Reader
(16GB), SKU-LUMI, has been delayed and will not arrive on the originally planned
date.

We apologise for the inconvenience and will keep you updated.

Regards,
LumiRead Supply Team

The supplier delay test email being composed in Outlook

Return to the workflow’s Activity panel, refresh until the run appears, and open it. Observe the path: Classify → Supplier Delay → Inventory Task Agent.
Select the Inventory Task Agent node in the run details and scroll through the tool runs. You should see the warehouse lookup and the Dataverse describe/create operations executed by the agent:
- Using the Warehouse MCP, the agent looks up the current stock for the delayed product (the SKU named in the email).
- Using Dataverse, it then inspects the Task table and creates a new row there, filling in the subject, description, priority, and due date exactly as the instructions specified.
Together, these tool calls show how the agent turned an unstructured delay email into a concrete, trackable restock task.
In the same run details, review the structured output and confirm it contains:
- sku = SKU-LUMI
- stock = 0
- risk = true
Open Power Apps (make.powerapps.com), navigate to Tables → Task, and find the row Restock delay - LumiRead E-Reader (16GB) (SKU-LUMI). Verify the task has a Due Date three days from now and Priority = High.

Congratulations! The Supplier Delay path is now using an MCP-powered inline agent.

You’ve added a reasoning agent that enriches a supplier delay notice with live warehouse data, creates a Dataverse task, and emits structured output for downstream visibility.

Use Case #5 (Bonus): Call a Price Quote Specialist Agent from a Workflow

Connect the Quote Request branch to a reusable, published Price Quote Agent that combines SharePoint knowledge, Dataverse product data, and WorkIQ Mail to produce and send a professional quote.

Summary of tasks

In this section, you’ll inspect the existing Price Quote Agent, wire it into the Quote Request branch, pass the sender and email body as dynamic inputs, publish the workflow, and verify the generated quote email end-to-end.

Scenario: A customer asks for pricing on a bundle of products. Rather than rebuilding the quote logic inside the workflow, you can reuse an already-published agent that knows how to look up pricing guidance, combine it with product catalog data, and send a polished quote response.

Objective

Configure the Quote Request path to use the published Price Quote Agent, provide it with the email sender and request body at runtime, and verify that it completes the quote process and sends the resulting email.

Step-by-step instructions

Inspect the reusable Price Quote Agent

In Copilot Studio, select Agents in the left navigation. Ensure the New experience toggle (top-right of the Agents list) is ON — this gives you the modern agent builder.
Open Price Quote Agent. In the new experience’s Build tab, inspect the agent and confirm these components are present on the right-hand panel:
- Tools: Microsoft Dataverse MCP Server and Work IQ Mail (Preview)
- Knowledge: Order Management (SharePoint knowledge source with pricing guidance, customer tiers, and related policy)
- Skills: price-quote — the structured skill for composing and sending the quote email
- On the left, the Instructions pane holds the agent’s broad, conversational guidelines about handling price questions. Notice they stay deliberately high-level: the precise, procedural steps live in the price-quote skill instead — a best practice that keeps the instructions readable while the repeatable logic is encapsulated in a skill.
Open the price-quote skill and read through what it does. The skill spells out exactly how to build a quote: read the customer name and the requested items from the request, use Dataverse to look up the customer’s account and employee count and derive their pricing tier (Small / Mid / Large), take each item’s SKU and list price from the Sales & Pricing Guide, then calculate the line totals, subtotal, tier discount, and final total. It finishes by reading the payment terms and delivery lead time and using Work IQ to email a fully formatted quote to the customer — applying only the prices and discounts found in the knowledge, never estimates.
Publish the Price Quote Agent by selecting Publish in the top-right corner.

Configure the Quote Request branch

Return to the Order Management Workflow and locate the Quote Request category in the Classify node.
Select the + next to Quote Request and choose Agent.
In the Agent dropdown, select Price Quote Agent — the existing published agent.

[!IMPORTANT] Only published agents appear in the dropdown: When using the Agent node with an existing agent, only agents that have been published (not just saved) will appear in the agent selection dropdown. If your agent is missing from the list, go to the agent page and publish it first.

In the Message field, enter:

Prepare a price quote for this customer request:
Customer email: [email sender]
Request: [email body]

Replace the placeholders with dynamic values:
- Replace the [email sender] placeholder (select the arrow >), choose the lightning icon, search for from, and insert From.
- Replace the [email body] placeholder (select the arrow >), choose the lightning icon, search for body, and insert Body.
Select Save, then Publish.

Test the Quote Request path

In Outlook, send the following email to your lab user account:

Subject: Order Management - Quote request - Adventure Works

Body:

Hello Contoso team,

We're fitting out a new office and would like a quote for:
- 10 Sony WH-1000XM5 headphones
- 5 iPad Air (M2, 256GB)

Delivery to our Santa Cruz, California office. Could you confirm your best price,
the payment terms, and the delivery estimate?

Thanks,
Nancy Anderson
Procurement, Adventure Works

The quote request test email being composed in Outlook

Return to the workflow’s Activity panel, refresh until the run appears, and confirm the route Classify → Quote Request → Agent completed.
Select the Agent node in the run details and read the completion message describing what the Price Quote Agent did to prepare the customer response.
Open Outlook and verify that the generated price quote email has landed in your inbox.
Bonus: Return to Price Quote Agent and select the Preview tab (top of the new experience). As a bonus, observe how the agent reasons over a question by combining the different tools at hand. Ask: What are the iPad Air prices applicable to Fabrikam? Watch the extended reasoning — the agent issues a Dataverse read_query that looks up the number of employees where the account name is Fabrikam, then connects that result with the SharePoint knowledge about customer tiers: it works out that Fabrikam falls into the Mid tier and shows the discounts the SharePoint handbook says should apply.

You can also try asking We have received this customer email: [] and paste the email you used to test the workflow. You’ll see how the agent reasons over the answer — and how, this time, it loads the price-quote skill to build a full quote following the skill’s specified steps.

Congratulations! The Quote Request path is now powered by a reusable published agent.

You’ve connected a workflow branch to a published agent that blends knowledge, data, and mail actions — a strong pattern for reusing sophisticated business logic across multiple automations.

Summary of Learnings

True learning comes from doing, questioning, and reflecting — so let’s put your skills to the test.

To get the most out of Workflows in Copilot Studio:

Trigger first, chat never — a Workflow runs on an event (a schedule or a connector trigger), so it works in the background without a user present. Pick the trigger that matches the “when” of your automation.
Use an inline agent for the messy middle — when a step needs judgment (which slot? how to phrase the notes?), the agent node lets the model reason and act, instead of forcing every decision into deterministic logic.
Ground the agent with dynamic content — the / token feeds the agent the actual trigger data (the to-do Title), so each run acts on the real item, not a fixed example.
Tools are the agent’s hands — without the Work IQ Calendar and Update to-do tools (and web search), the agent could reason but not act. Grant exactly the tools the goal requires.
Constrain behavior in the instructions — small additions like “only within my 8:00 AM–5:00 PM working hours” meaningfully change the agent’s choices. Iterate on the instructions, then re-publish.
Publish before you test — the trigger is only live after a publish, and changes don’t take effect until you re-publish.

Conclusions & Recommendations

[!IMPORTANT] Workflows golden rules:

Match the trigger to the event that should start the work — manual for testing, recurrence for schedules, connector for “when something changes.”

Reach for an inline agent when a step needs reasoning; keep deterministic actions for steps that must always behave identically.

Reference trigger data with the / dynamic-content token so the agent acts on the real input every run.

Give the agent only the tools it needs, and state constraints (working hours, formatting, categories) explicitly in the instructions.

Always Save → Publish before testing, and re-publish after any change. Use the Activity and Monitor views — and the Agent node’s run details — to see exactly what the agent decided and did.

With these foundations, you can move from agents that talk to agents that act on their own — reacting to events, reasoning over real data, and getting work done across Microsoft 365.

Workflows

Workflows

Lab Details

Table of Contents

Why This Matters

Introduction

Core Concepts Overview

Documentation and Additional Training Links

Prerequisites

Summary of Targets

Use Cases Covered

Instructions by Use Case

Use Case #1: Automate Task Time-Blocking with a Workflow and an Inline Agent

Objective

Step-by-step instructions

Create the To-Do list that drives the workflow

Create the workflow and configure its trigger

Add and configure the inline agent

Publish, then trigger the workflow

Watch the workflow run

Verify the results

Congratulations! You’ve built an autonomous Workflow with a reasoning inline agent.

Use Case #2: Setting Up the Order Management Workflow

Objective

Step-by-step instructions

Take ownership and configure the solution

Open and explore the Order Management Workflow

Test the classification

Test the “Other” classification path end-to-end

Congratulations! The Order Management Workflow is now fully configured and operational.

Use Case #3: Use M365 Copilot and Add a Human-in-the-Loop in Order Management

Objective

Step-by-step instructions

Inspect the Customer Inquiry branch

Run an end-to-end Customer Inquiry test

Congratulations! The Customer Inquiry path is now drafting, reviewing, and replying successfully.

Use Case #4: Build an Inline Agent for Inventory Management

Objective

Step-by-step instructions

Add and configure the Supplier Delay agent

Test the Supplier Delay path

Congratulations! The Supplier Delay path is now using an MCP-powered inline agent.

Use Case #5 (Bonus): Call a Price Quote Specialist Agent from a Workflow

Objective

Step-by-step instructions

Inspect the reusable Price Quote Agent

Configure the Quote Request branch

Test the Quote Request path

Congratulations! The Quote Request path is now powered by a reusable published agent.

Summary of Learnings

Conclusions & Recommendations