Automate Document Processing with AI Builder and Power Automate

1. AI Builder in Power Automate: The Automation Engine

While Power Apps brings AI to the user's fingertips, Power Automate brings AI to your business processes. By integrating AI Builder into a cloud flow, you can create automations that trigger on an event (like a new email or a file being created), use an AI model to understand the content, and then take action in other systems—all without human intervention.

To get a general sense of the types of models available and the overall process, the following article from Avantiico provides a good high-level summary.

As the article explained, the core of this integration is the AI Builder connector in Power Automate. It contains a rich set of actions that correspond to the different AI models. Your main task as a developer is to:

1. Choose the correct trigger for your process.
2. Add the appropriate AI Builder action and provide it with the necessary input (e.g., file content, text).
3. Parse the output from the AI model and use that data in subsequent actions (e.g., writing to a Dataverse table, sending an email, or starting an approval).

2. Practical Example: Automating Document Processing

Let's dive into a detailed, hands-on scenario. Imagine your organization receives construction funding request forms as PDF attachments in a shared mailbox. Your goal is to create a flow that automatically:

• Triggers when a new email with an attachment arrives.
• Uses a custom AI Builder document processing model to extract data like Loan Number, Draw Amount, and line items from the PDF.
• Creates a new "Loan Draw" record in Dataverse with the extracted information.

This is a classic automation pattern and a perfect demonstration of AI Builder's power. We will follow a guided lab that walks through building this exact solution.

While the lab also covers training the model, you can treat that part as a review, since we've covered it in previous lessons. Our primary focus will be on Exercise #2, which details the Power Automate flow integration.

Deconstructing the Flow

Let's break down the key steps and concepts from that lab exercise, which are fundamental to any document processing flow.

1. The AI Builder Action:
The core of the flow is the Extract information from documents action. You select your published custom model, specify the document type, and pass the file content from the email attachment as input.

2. Handling Model Outputs:
After the action runs, its outputs are available as dynamic content. This is where your developer skills come into play. The outputs are structured and predictable:

• Simple Fields: For each field you defined in your model (e.g., Loan Number), the action provides [Field Name] value (the extracted text) and [Field Name] confidence score.
• Tables: For tables (e.g., Budget Items), the action provides an array of objects. Each object in the array represents a row, and its properties correspond to the columns you defined (e.g., Budget Items Description value, Budget Items Amount value).

This GIF illustrates how you map these outputs to the inputs of another action, such as creating a Dataverse record.

3. Processing Tabular Data:
Because the table data is an array, you must use an Apply to each control to iterate over it. In the lab, the goal was to create a single string of all work items. The pattern is:

1. Initialize a string variable before the loop.
2. In the Apply to each loop (which iterates over the Budget Items output), use the Append to string variable action to add the details from each row.

4. Data Transformation with Expressions:
Raw data from a document often needs cleaning. The lab provides an excellent example: the Draw Amount is extracted as text (e.g., "$10,000.00"). To store this as a currency value in Dataverse, you must first remove the non-numeric characters. The expression used is:
string(replace(outputs(...),'€',''))
(Note: The lab might use '€' or '$', the principle is the same).
Given your background in C# and data manipulation, you can see this is analogous to string manipulation and type casting. Power Automate expressions are a powerful tool for these transformations.

3. Integrating a Text Processing Model

Not all AI automation involves complex documents. You can also automate processes based on unstructured text using prebuilt models.

Scenario: Let's say you want to analyze customer feedback submitted via a web form and route it based on sentiment.

The flow would be:

1. Trigger: When a new response is submitted (from the Microsoft Forms connector).
2. Get Details: Get response details.
3. Analyze Sentiment: Add the AI Builder action Analyze positive or negative sentiment in text. Pass the feedback text from the form response to the Text input field.
4. Condition: Add a Condition control. Check the output of the sentiment action. The condition could be Overall text sentiment is equal to Negative.
5. Branching Logic:
   • If yes (Negative): Create a task in Planner for a manager to follow up and send an empathetic email to the customer.
   • If no (Positive/Neutral): Add the feedback to a "Positive Feedback" list in SharePoint.

This pattern is incredibly useful for triage, routing, and understanding customer voice at scale.

Accessing Nested Outputs with Expressions

Sometimes, the dynamic content pane might not show the exact value you need, or you may want to access a specific piece of data from the complex JSON output of a model. As a developer, you can write expressions to navigate this JSON directly.

The image below shows a Predict action's output being parsed by an expression in a Compose action.

This demonstrates that you have full control to access any part of the model's response, just as you would when consuming a REST API in C#. The ? is the safe navigation operator, equivalent to the null-conditional operator (?.) in C#, preventing the flow from failing if an intermediate property is null.

Conclusion

In this lesson, you have learned how to embed AI intelligence directly into your automated backend processes using Power Automate. This moves AI from a tool for user assistance to a driver of core business automation.

Key Takeaways:

• The AI Builder connector in Power Automate is your gateway to using models in flows.
• The implementation pattern is: Trigger -> AI Action -> Parse Output -> Subsequent Actions.
• Custom document processing models return structured fields and tables. Table outputs are arrays that must be processed using an Apply to each loop.
• Prebuilt text models (like sentiment analysis) provide a simple way to add intelligence to text-based triggers.
• Expressions are crucial for transforming and cleaning extracted data and for accessing deeply nested values in a model's output.

What's Next?

We've now worked with a custom document model and discussed a prebuilt text model. In the next lesson, we will focus on another common scenario: using the prebuilt text recognition AI model in a flow to extract text from an image. This will cover how to handle unstructured text within images, such as text from a photograph of a sign or a screenshot, further expanding your automation toolkit.