Lesson: AI Builder Model Types

Introduction: Why Model Selection Matters

In the modern landscape of business automation, the ability to turn unstructured data—like emails, invoices, or customer feedback—into actionable insights is a significant differentiator. AI Builder, a feature within the Microsoft Power Platform, provides a low-code interface that enables teams to build and deploy artificial intelligence models without needing to be professional data scientists. However, the true power of AI Builder is not found in the platform itself, but in the intelligent selection and configuration of the specific model type that fits your unique business problem.

Choosing the wrong model type is a common point of failure for many organizations. If you attempt to use a document processing model to analyze sentiment in social media posts, you will find yourself frustrated by poor accuracy and high maintenance costs. Conversely, understanding the taxonomy of models—knowing when to choose a prebuilt model versus a custom one, or a structured data model versus an unstructured data model—allows you to solve complex automation challenges with minimal technical overhead. This lesson will walk you through every major model type available in AI Builder, helping you map specific business requirements to the right AI capability.

Understanding the Two Categories: Prebuilt vs. Custom

Before diving into specific models, it is vital to distinguish between prebuilt models and custom models. This distinction dictates how much data you need to provide, how much effort is required to set up the solution, and how the model behaves over time.

Prebuilt Models

Prebuilt models are "off-the-shelf" solutions provided by Microsoft. They are trained on massive datasets covering common business scenarios, such as reading business cards or extracting information from tax forms. They require zero training on your part; you simply connect them to your data source and start using the output.

Custom Models

Custom models are tailored to your specific organization’s data. If you have a unique type of invoice that no standard model recognizes, or if you need to predict which customers are likely to churn based on your specific historical CRM records, you will need to build a custom model. These require you to upload your own training data, label the information, and manage the training lifecycle.

Callout: The "Build vs. Buy" Decision for AI When deciding whether to use a prebuilt or custom model, ask yourself two questions: Does a generic version of this task exist (e.g., extracting a date from a receipt), and do I have enough historical data to teach a machine to do this task specifically for my company? If the task is universal, start with a prebuilt model. If the task is highly specific to your internal processes, invest the time to build a custom model.

Deep Dive: Document Automation Models

Document automation is perhaps the most popular use case for AI Builder. These models are designed to ingest physical or digital files and convert them into structured data that your applications can process automatically.

1. Document Processing (Formerly Form Processing)

The Document Processing model is the workhorse of automated data entry. It is designed to extract key-value pairs, tables, and checkboxes from documents. You use this when you have consistent document layouts, such as purchase orders, utility bills, or insurance claims.

How to configure it:

1. Define Fields: In the AI Builder portal, you specify the fields you want to extract (e.g., "Invoice Number," "Total Amount").
2. Upload Samples: You must provide at least five examples of each document type you want the model to handle.
3. Labeling: You manually draw boxes around the fields on your sample documents so the AI understands where to look.
4. Train and Publish: Once the model learns the patterns, you publish it to your environment.

Tip: Always aim for at least 10-15 high-quality samples for each document layout. While the platform says five is the minimum, providing more variety in your training set significantly improves the model's ability to handle document skew, different fonts, and varying print qualities.

2. Prebuilt Receipt Processing

If your company manages employee expenses, the Receipt Processing model is an essential tool. It is a prebuilt model that automatically recognizes receipt data, including the merchant name, transaction date, tax amount, and total. You do not need to train this; it is ready to go as soon as you connect it to a Power Automate flow.

3. Prebuilt Invoice Processing

Similar to receipt processing, the Invoice Processing model is specialized for the B2B world. It understands the nuances of invoices, such as line items, sub-totals, and vendor addresses. Because it is prebuilt, it is resilient to the thousands of different invoice formats used by vendors globally.

Deep Dive: Text and Language Models

Text models allow your applications to "read" and "understand" natural language. These are essential for automating customer support, sentiment analysis, and content classification.

1. Category Classification

This is a custom model used to tag text based on predefined categories. For example, you might want to automatically route incoming customer support emails to the "Billing," "Technical Support," or "Sales" department.

• Input: A block of text (email body, chat transcript).
• Output: A category tag (e.g., "Billing").

Step-by-Step Implementation:

1. Prepare your dataset: You need a CSV or Excel file containing two columns: "Text" and "Category."
2. Choose the Model: Select "Category Classification" in the AI Builder portal.
3. Train: Upload your dataset. The AI will analyze the word associations for each category.
4. Test: Use the "Quick Test" feature to see how it handles new, unseen text before you deploy it.

2. Entity Extraction

Entity extraction is used to identify specific "entities" within a block of text. If you are processing medical records, you might want to extract "Patient Name," "Medication," and "Dosage" from a doctor’s note. Unlike classification, which labels the entire document, extraction labels specific parts of the document.

3. Sentiment Analysis (Prebuilt)

This model detects the emotional tone of a piece of text. It returns a score indicating whether the text is Positive, Negative, or Neutral. This is a foundational element for building "Voice of the Customer" dashboards in Power BI or automated escalation triggers in Power Automate.

Warning: The "Neutral" Trap Be careful when relying on sentiment analysis for high-stakes customer interactions. A "Neutral" score does not always mean the customer is happy; it often means the sentiment is simply flat or ambiguous. Always include a human-in-the-loop workflow for negative sentiments, but do not ignore the neutral ones entirely, as they may represent a lack of engagement.

Deep Dive: Prediction Models

Prediction models are the most "data-heavy" models in the AI Builder suite. They are used to forecast future outcomes based on historical patterns in your data (e.g., "Will this lead convert?" or "Is this transaction fraudulent?").

Prediction (Binary Classification)

This model answers a Yes/No question. It looks at your historical data (e.g., a table of past sales opportunities) and identifies which factors correlate with success.

• Training Data: Must be stored in Dataverse. You need a table that contains both the historical inputs and the final result (e.g., "Won" vs. "Lost").
• Configuration: You select the "Outcome" column, and the model automatically selects the most relevant "Predictor" columns.

Callout: The Importance of Historical Quality Prediction models are only as good as the data they are fed. If your historical data is missing values, or if your sales team wasn't diligent about updating the "Status" of leads in your CRM, the model will produce biased or inaccurate predictions. Ensure your data is "clean" (no duplicate records, no missing outcomes) before attempting to train a prediction model.

Comparison of AI Builder Model Types

Best Practices for Successful AI Implementation

Configuring AI Builder models is an iterative process. You rarely get the "perfect" model on the first attempt. Follow these industry-standard practices to ensure your AI projects provide real value.

1. Start with a "Minimum Viable Model"

Don't try to build a model that handles every edge case on day one. Start by building a model that handles 80% of your most common scenarios. Once that is performing well in a production environment, you can iterate by adding more training data to cover the remaining 20%.

2. Monitor Performance Over Time

AI models suffer from "model drift." As your business processes change (e.g., a new invoice template is adopted by your primary vendor), the performance of your document processing model will degrade. Establish a quarterly review process where you check the accuracy of your models and retrain them with new, representative data.

3. Maintain Data Governance

AI Builder models inherit the security permissions of the environment they are in. Ensure that the data used to train your models is handled according to your organization's privacy policies. Do not train models on sensitive personal information unless you have the appropriate data masking or security controls in place.

4. Implement a "Human-in-the-Loop" Workflow

Never fully automate a business process based on AI output without a safety net. For document processing, create a Power Automate flow that routes any document with a "confidence score" below 80% to a human reviewer. This ensures that errors are caught before they reach your downstream systems.

Common Pitfalls and How to Avoid Them

Pitfall 1: Insufficient Training Data

The most common reason for a poor-performing model is simply having too little data. If you try to train a custom model with only two or three examples, the model will be unable to generalize.

• The Fix: If you don't have enough data yet, focus on manual data collection for a few months. Use that time to clean your existing datasets so they are ready for training once you reach the required threshold.

Pitfall 2: Overfitting

Overfitting occurs when a model becomes "too smart" for its own good—it memorizes your training data perfectly but fails to handle new, unseen data correctly.

• The Fix: Ensure your training set is diverse. If you are training a document model, include documents from different times of the year, different scanners, and even slightly different layouts to force the model to learn the core structure rather than the specific pixels of the training samples.

Pitfall 3: Ignoring the Confidence Score

Every prediction and extraction model returns a "confidence score." Many beginners ignore this and treat the model's output as 100% truth.

• The Fix: Always use the confidence score in your Power Automate flows. Use conditions like: If Confidence Score > 0.85, Update Record; Else, Send Email to Manager.

Technical Implementation: A Quick Start Example

To help you get started, here is a conceptual look at how you would handle an AI Builder prediction in a Power Automate flow.

Step-by-Step: Automating a Lead Score

1. Trigger: Create a flow that triggers when a "Lead" record is created in Dataverse.
2. Action: Use the "Predict" action from the AI Builder connector.
3. Configuration:
   • Select your "Lead Conversion" model.
   • Map the fields from your Dataverse trigger to the input fields required by the model.
4. Logic:
   • Use a condition block to check the output.
   • If Prediction Result is "Converted" AND Confidence Score is > 0.90, assign the lead to a senior sales representative.
   • If Prediction Result is "Not Converted," add the lead to a nurture email sequence.
5. Update: Update the Lead record in Dataverse with the score and the prediction result so the sales team can see it.

Note: When using the "Predict" action, ensure your field mappings are exact. Data types must match—if the model expects a numeric value for "Annual Revenue," ensure the data being passed from your trigger is not a string.

Advanced Configuration: Managing Model Versions

As your organization grows, you will inevitably need to update your models. AI Builder allows you to manage versions, which is critical for maintaining business continuity.

Versioning Strategy

1. Develop in a Sandbox: Never train or update models in your production environment. Use a dedicated development environment to train and test your models.
2. Test Rigorously: Use a "Hold-out" test set—a collection of data that the model has never seen during training—to validate the model's accuracy before moving to production.
3. Deployment: Once tested, use Solution Management in Power Platform to move the model from your development environment to production. This ensures that all components, including the model and the associated Power Automate flows, are moved as a cohesive package.

Frequently Asked Questions (FAQ)

Q: Can I use AI Builder for free? A: AI Builder is a premium feature. It requires either a specific AI Builder add-on license or credits included with certain Power Apps or Power Automate plans. Check the Microsoft pricing page for the most current credit allocation details.

Q: What happens if I run out of AI Builder credits? A: If you exceed your monthly credit allowance, your models will stop functioning. It is best practice to set up alerts in the Power Platform Admin Center so you are notified before you reach your credit limit.

Q: Can I export my AI Builder model to another platform? A: No, AI Builder models are native to the Power Platform. They are designed to work within the Dataverse ecosystem. If you need to move models between platforms, you would need to rebuild them using standard data science tools like Azure Machine Learning.

Q: Is there a limit to how many documents I can process? A: The limit is governed by your credit consumption. Each page processed by a document model consumes a specific number of credits. Larger, more complex documents consume more credits than simple ones.

Key Takeaways

1. Strategic Selection: The success of your automation project begins with choosing the right model type. Distinguish between prebuilt models (for standard tasks) and custom models (for company-specific tasks) before you start.
2. Data Quality is Paramount: AI is not magic; it is mathematics. If you feed the model poor, incomplete, or biased data, the output will be poor, incomplete, or biased. Invest time in data cleaning.
3. Iterative Improvement: Treat AI model development as a lifecycle, not a one-time setup. Plan to retrain your models periodically to account for changes in your business data or document formats.
4. Safety Nets are Essential: Always implement "Human-in-the-loop" workflows. Use confidence scores to automatically route high-certainty items and flag low-certainty items for human review.
5. Start Small, Scale Fast: Don't try to solve the entire business process at once. Validate your model with a small pilot project, confirm the accuracy, and then expand to larger datasets.
6. Governance and Security: Remember that AI models operate within your existing security framework. Ensure that sensitive data is protected and that your AI implementations comply with organizational data policies.
7. Monitor the Limits: Keep a close watch on your credit consumption and model performance to avoid service disruptions or "model drift" that could negatively impact your business operations.

By following these principles, you can move from simple task-based automation to intelligent, scalable systems that provide genuine value to your organization. The goal of AI Builder is to empower you to solve problems, not to create new ones; by staying disciplined with your configuration and testing, you ensure that your AI projects remain reliable and effective assets for your team.