Power Automate
Neil Haddley • February 17, 2026
Connect Power Automate to Azure AI Foundry
I added an Azure AI Foundry model call to a Power Automate Flow.
I created a Foundry
I created a new resource group
Review + create
Create
Go to resource
Go to Foundry portal
I copied KEY1 and the API Endpoint (I would need to adjust the API Endpoint later)
I created a new solution
I added an environment variable to the solution
I created the lead qualification foundry endpoint environment variable
I created the lead qualification foundry api key environment variable
+New|Table|Tables
Start with Copilot
Lead table is related to (standard) Dataverse Account and Contact tables
I clicked Save and exit
The solution had one table and two environment variables
+New|App|Model-driven app
Create
+ Add page
Dataverse table
Add (Lead)
View new page
The app was added to the solution
+ Deploy model
gpt-4o-mini
Standard
I copied model Endpoint
I updated lead qualification foundry endpoint environment variable
+New|Automation|Cloud flow|Automated
The flow is triggered when a Dataverse row is added, updated, or deleted
I updated the trigger to run only when a new Lead record was added or when the Description column of an existing row was updated
I used an Enabled variable to know if AI Enable AI processing was True (notice that Dataverse maps Yes to 0 and No to 1)
EXPRESSION
1not(triggerOutputs()?['body/hadd_enableaiprocessing'])
I created a Condition based on the Enabled variable value
I called the AI model. I used the environment variables to set the URL and API-Key values
PROMPT
1{ 2 "messages": [ 3 { 4 "role": "system", 5 "content": "You are an expert sales manager. Your task is to classify a lead based solely on the provided description and output a structured assessment.\n\nUse the following categories for the lead's potential:\n- **Very Promising**: High urgency, clear budget, decision-maker involved, strong fit with our offering.\n- **High Potential**: Good fit, expressed interest, but may lack immediate budget or timeline clarity.\n- **Good Chance**: Moderate interest, some qualification criteria met, but needs nurturing.\n- **Moderate Interest**: Vague interest, initial inquiry, still exploring options.\n- **Low Interest**: Unlikely to convert soon; may be research-only, no budget, or mismatched needs.\n- **Other**: If none of the above apply, provide a brief, accurate label.\n\nIn addition, provide:\n- **score**: A numeric lead score from 0 to 100, where higher values indicate stronger potential and fit.\n- **confidence**: Your certainty in this assessment, expressed as one of: \"Low\", \"Medium\", or \"High\".\n- **explanation**: A brief 1–2 sentence justification for the category, score, and confidence level.\n\nOutput your response **only as a JSON object** with the following keys:\n- \"category\": string\n- \"score\": integer (0–100)\n- \"confidence\": string (\"Low\", \"Medium\", or \"High\")\n- \"explanation\": string\n\nDo not include any other text, markdown, or formatting—just the raw JSON." 6 }, 7 { 8 "role": "user", 9 "content": "@{triggerOutputs()?['body/hadd_description']}" 10 } 11 ], 12 "max_tokens": 400, 13 "temperature": 0 14}
I used Compose to fetch the result (content) from the body of the HTTP response
EXPRESSION
1body('Call_LLM_(HTTP)')?['choices'][0]?['message']?['content']
I used Parse JSON to extract the category, score, confidence and explaination values
EXPRESSION
1outputs('Get_content_(Compose)')
SCHEMA
1{ 2 "type": "object", 3 "properties": { 4 "category": { 5 "type": "string" 6 }, 7 "score": { 8 "type": "integer" 9 }, 10 "confidence": { 11 "type": "string" 12 }, 13 "explanation": { 14 "type": "string" 15 } 16 } 17}
I used a Dataverse Update a row action to update the record (row)
EXPRESSION
1body('Parse_Result_(Parse_JSON)')?['explanation']
EXPRESSION
1body('Parse_Result_(Parse_JSON)')?['score']
I used the model driven app to manually update the Description column value
I checked the Update a row action
I viewed the updated row in the model driven app