Overview
Voiceflow integrates with Microsoft Power Automate so your AI agent can submit an employee request to a protected cloud flow and return its pending approval reference. Microsoft Power Automate runs the external workflow; Voiceflow calls its approved endpoint and uses the response in the conversation.
Examples of when to use Microsoft Power Automate
- Connect the business service — prepare its authorization, schema and test resources.
- Request an employee approval — invoke the configured service and interpret its current result.
- Recover a failed request — inspect execution and downstream state before retrying.
Studio
Use an editable test project in Agent studio. Your integration developer or administrator prepares the service below, supplies its endpoint and credentials, and gives you a verified test request plus its expected result. Keep provider-side deployment with that owner.
Code
Use a Voiceflow personal or service key with project access, a current Voiceflow CLI and the selected test Environment. You prepare the provider service and create, execute and attach the Voiceflow request separately. The examples use a test project; they can read or change connected systems.
Connecting Microsoft Power Automate
Prepare a deployed endpoint and the credentials for its approved operation.
Prepare the provider service
Studio
Ask your Microsoft Power Automate administrator or integration developer to complete the service setup linked below. Obtain all four handoff items before creating the Voiceflow tool:
- The deployed HTTPS endpoint and its exact authentication headers.
- The credential value for the Voiceflow Secret.
- A request schema, response schema and verified test employee/item.
- A successful test response and a rejected unauthorized request, confirmed in the provider’s execution history.
The sample request fields on this page are a recipe contract for your existing workflow. Have the owner map them to that workflow; a provider account alone does not create this endpoint.
Code
Create and restrict the HTTP-triggered flow
A Power Automate administrator creates a cloud flow with When an HTTP request is received. In Who can trigger the flow?, choose Specific users in my tenant and enter the calling service principal’s object ID. Do not leave that list empty: an empty list permits users across the tenant. Define required string fields request_id, employee_id, item and justification in the trigger’s JSON schema. Save the flow and copy its current HTTP POST URL.
Create an approval and return a receipt
Validate that the employee may request the item, choose the approver from your approved business rules and deduplicate request_id in a durable request store before creating work. Use Create an approval, retain the returned Approval ID and save its mapping to request_id. Add Response with status 202 and JSON request_id, approval_id and status: "pending". For a repeated request, return the saved receipt instead of creating another approval.
Use Wait for an approval with the saved Approval ID for later processing, then update the request record and notify the verified employee through your existing notification route. Do not put Start and wait for an approval before the receipt: it blocks until someone responds. The receipt confirms a request is waiting, not that it has been approved.
Deploy the token relay
Register a confidential application in Microsoft Entra ID and obtain its tenant ID, application/client ID and credential. The flow’s allowlist must use this application’s service principal object ID. For this public-cloud example, acquire a client-credentials token for audience https://service.flow.microsoft.com/; the v2 scope is https://service.flow.microsoft.com//.default, preserving the trailing slash.
Download power-automate-relay.mjs and serve-power-automate.mjs. Use Node.js 22 for this example and keep both files in one directory. Configure TENANT_ID, CLIENT_ID, CLIENT_SECRET, FLOW_URL and a generated RELAY_TOKEN in your hosting service’s secret store. Start node serve-power-automate.mjs behind your HTTPS reverse proxy, forwarding /approval to its configured PORT (8080 by default). The relay obtains a fresh Entra token for each request, calls one fixed flow and returns only the validated receipt. It uses a 15-second total upstream deadline, with up to 4 seconds for the token and 10 seconds for the flow call, to leave room inside Voiceflow’s standalone endpoint wait. It never automatically retries the flow POST. A timeout still requires reconciliation because the flow may have created the approval.
The relay authenticates Voiceflow to the flow; your application and flow still own employee authorization, durable deduplication and final notification. Deploy those controls before exposing the action. Keep the Microsoft credential on your backend; Voiceflow stores only the separate relay token.
Verify the deployed handoff
Test a known employee/item through the relay. Require a matching 202 receipt, an approval visible to the intended approver and a request-store record linking the request ID to that approval. Repeat the same request ID and confirm there is still one approval. Verify an invalid relay token is rejected before a flow call, and an unauthorized employee is rejected before approval creation. Complete the approval once and verify your final notification reaches the intended employee.
Provider setup: HTTP trigger authentication, approval action behavior and client-credentials tokens.
Create the Voiceflow request
Studio
Store the service credential
Open your test project in Agent studio, then More → Secrets. Create power_automate_relay_token with the credential supplied by the service owner. Creating its default also writes Main’s Published value, so keep initial setup in a test project. If the selected Environment has a different Draft credential, set its Draft override too.
Create a request variable
Open More → Variables, create power_automate_request with an empty default and ask your application developer to populate it with the prepared request. For a Studio Run test, temporarily set its default to the developer’s prepared test string in this isolated test project; backend runtime variables do not automatically populate Studio Run. Clear the default after that test.
Configure the API request
Open Agent → Global tools → + → API and name the tool Request equipment approval. Choose POST, set the owner-supplied endpoint and add the headers below. Set the raw JSON body to {payload}.
Bind inputs before running
Open the attached tool’s inputs. Open authorization and set its Default value to the Secret power_automate_relay_token, and set payload’s Default value to power_automate_request. Keep Async execution off so the agent receives this call’s response. The provider workflow may still return a pending receipt.
Code
Set VF_TOKEN, PROJECT_ID and ENVIRONMENT for the test project. Set PROVIDER_CREDENTIAL to the separate relay token. The service owner supplies PROVIDER_URL and the verified test request.
Save this definition as power-automate-tool.json, replacing the illustrative URL with PROVIDER_URL. Braced names declare inputs on this API tool; they are bound to Secrets/variables on its attachment later.
Save apiTool.id as API_TOOL_ID, then discover input IDs:
Save apiTool.id as API_TOOL_ID, then discover input IDs:
Creating the Secret default also writes Main’s Published value. If your selected Environment uses a different credential, set its Draft override before testing. Match each variables[].name to its id for the attachment; resolve unresolvedReferences before continuing.
Request equipment approval
Provide a verified request and use the result from the current call.
Invoke Request equipment approval when the employee asks for equipment approval and confirms the item and justification. The authenticated application supplies the verified identity and prepared request variable before the call; keep those inputs bound to trusted defaults. A matching approval reference with pending status means the request is waiting for approval. It does not mean the equipment was approved. For a pending result, explain that work is waiting. For a current failure or missing required fields, say the result could not be confirmed; never reuse an earlier result or automatically repeat an uncertain write.
Prepare the request
Code
The application developer builds the request from an authenticated employee/item and the confirmed task. Save this illustrative body as power-automate-request.json, replacing test identities with values the application has verified:
Serialize it with the literal-safe request encoder: node encode-request.mjs power-automate-request.json > power-automate-payload.txt. It preserves quotes, newlines and literal text such as {employee_name} without treating that text as a Voiceflow variable or Secret. Do not concatenate employee text into quoted JSON placeholders.
Studio
Have your developer deliver the encoded string and populate power_automate_request in the authenticated client’s next action before this tool runs. They send variables on the runtime action; creating the variable or session alone does not fill it. For Studio Run, use the temporary test default described above, then clear it. The provider service must enforce employee access and validate the request; tool descriptions do not enforce either.
Code
Read the confirmed body, encode it, and build standalone inputs. This tests the saved request against the selected Environment’s Draft; it does not attach the tool or publish the agent.
Require success: true, inspect output.status and output.body, and check the business fields against the service owner’s response schema. truncated: true means the returned body is incomplete. On timeout or error, inspect provider execution before retrying a write.
Let the agent use the request
Studio
Describe the trigger and result use
Open the attached Request equipment approval tool and set Trigger to the condition below. Its input defaults must already point to the Secret and prepared request variable; the model must not collect credentials or choose another employee’s identifier.
Test the conversation
In Run, use the verified test request. Ask “Please request a replacement headset.” and confirm the intended tool runs with the fixed identity and prepared payload. Inspect the current tool result in Run and compare it with the pending approval and request-store record. Also test the rejected-resource case: the agent must not claim success or expose another employee’s data.
Code
Save this as power-automate-attachment.json. Replace YOUR_API_TOOL_ID and each input ID with the values discovered above. The payload and credential are defaults, not model-collected inputs.
Your authenticated backend supplies the encoded payload as variables.power_automate_request on the next runtime action. Use a test client bound to this project’s Draft, ask for equipment approval and inspect the emitted tool call, current response and agent answer. Verify both the successful and rejected-resource cases. The backend/provider enforces access; the description is model guidance, not an authorization gate.
The attached API tool returns the provider body to the agent. For a Workflow’s Capture response, paths start at that body: for example status rather than standalone output.body.status. A missing capture can leave an earlier variable value unchanged; clear result variables before a Workflow call and branch on the current HTTP/business outcome. These Global tool examples use the current result directly.
Maintaining the connection
Inspect provider execution, update credentials and test the intended deployment.
Replace a credential
Inspect the original provider execution
Use the request ID and execution time to find the provider run and its downstream record. A timeout can follow a completed side effect; reconcile that state before repeating a write.
Update the intended credential version
Studio
Open the same project and More → Secrets, select the provider Secret and update the target Environment’s Draft value for testing. After validation, explicitly set the Published value for the intended Environment; this changes the credential serving live traffic.
Code
Retain secret.id from creation as SECRET_ID, set PROVIDER_CREDENTIAL to the replacement and use the value operation below. After validation, explicitly set the Published value for the intended Environment; this changes the credential serving live traffic.
Code
Save the Draft update body as credential-value.json:
For the selected Environment’s Published credential, change versionVariant to published after the Draft test. To replace the project default, omit environmentAlias and versionVariant; this also writes Main’s Published value. Test the intended connection before changing a live default.
Before launch, select the target Environment and supply the provider Secret’s Default value for its published agent; a Draft override alone does not supply Published. Publish the tested configuration, then repeat the conversation test against that Environment’s published client and verify the provider account/resource it reaches.
Rotate the credential with the provider owner, update the relevant Secret default/Environment overrides, and retest the existing tool. Removing a tool from Voiceflow does not remove the provider deployment or revoke its credential. Disable/revoke access in Microsoft Power Automate when retiring the integration, then remove the unused attachment and API resource.