The human in the loop feature covers four aspects:

• Providing feedback: Users and makers can provide feedback on AI prompts responses when it isn't accurate.
• Getting notifications for reviews: Makers are notified when new feedback is available.
• Listing reviews: Makers can review the feedback sent for a given AI prompt.
• Few-shot learning: Makers can use the feedback received or create samples to drive the AI prompts' responses and increase accuracy.

Using prompts inside Copilot agents allows you to set specific logic to extract and augment information from documents. It helps you do the following:

• Save hours for each Copilot user who won't need to do all the specific data extraction, augmenting and formatting manually.
• Perform the exact same document processing logic for each Copilot user asking for it, in only one or two chat interactions.
• Reuse the prompt on multiple agents, apps, and automated flows.

To use this feature, create a prompt in Power Apps or Power Automate of Copilot Studio. Then, add an input type (for example, image or document). Supported types are PNG, JPG, JPEG or PDF. Now you're ready to include the prompt in a Copilot agent using topics (actions aren't supported yet). Finally, test the agent and publish it when you're ready.

This feature is enabled by default without any admin configuration.

In prompt builder, the model selection capability allows you to create a secure connection to a custom model deployed in Azure AI Foundry. This integration empowers you to seamlessly incorporate your own fine-tuned and optimized models—tailored to specific business needs—directly into your Power Automate flows or Copilot agents.

By leveraging models you already trained on domain specific data (for example, legal, financial, and customer service), you can design highly specialized AI actions that go beyond generic language capabilities. This allows new opportunities for smarter, context-aware automation and conversational experiences that reflect your organization’s unique knowledge and tone.

The validation station is an interactive feature in document processing that enables you to review, verify, and correct data extracted from documents before final submission or integration. After AI-driven extraction captures data from invoices, receipts, contracts, or other documents, the validation station presents this data in a user-friendly interface. It does this by highlighting fields with low confidence scores or potential errors for manual confirmation. You can quickly inspect extracted fields, make corrections as needed, and approve data, which can ensure accuracy and compliance. This feature enhances quality control by allowing human intervention in critical points, reducing errors in automated document workflows, and increasing overall trust in the document processing system. With streamlined navigation and an intuitive layout, the validation station offers an efficient way for you to confirm AI results. This improves the reliability and precision of document processing outcomes.

The validation station is reached through an approval email or from its own entry point to monitor and validate what was extracted and processed. This feature comes with the package, so there's no need to enable it manually.

The document processing agent integrated into Copilot Studio enhances business workflows by automating the collection, extraction, categorization, and validation of data from diverse document types. Leveraging AI capabilities, it seamlessly processes invoices, contracts, and receipts, which transforms traditionally manual tasks into efficient, automated actions. The document processing agent comes with pre-configured flows and validation station. This makes it easy for you to set up document processing workflows that work out of the box.

Azure Document Intelligence 4.0 brings advanced features like improved Optical Character Recognition (OCR), natural language understanding, and layout detection, allowing AI Builder to process a wider variety of document types, from structured forms to complex unstructured content. This integration allows AI Builder users to leverage Azure's latest advancements in machine learning models directly within their workflows, enhancing extraction accuracy. The new integration also supports returning table, row, and cell confidence scores.

Prompts are the instruction language of agents, whether running directly from a prompt or as a flow behind the scenes. These instructions play a critical role in how the agent performs its various steps. Prompts can also be used to perform data generation tasks in Power Automate cloud flows or Power Apps applications.

This feature allows connector data to be used in prompts, expanding beyond the current limitation of data only in Dataverse. This enhancement improves the functionality and flexibility of your agents, cloud flows, and applications, making them more efficient and effective.

To use this feature, select + Add, or type / and insert table fields of existing connections. Alternatively, you can query data with Power Fx formulas in prompt builder.

To create and use prompts in Power Automate, on the navigation menu to the left, select AI Hub > Build your own prompt, or from the Create text with GPT action when authoring a cloud flow.

You can access the Power Fx designer from prompt builder. You can define an expression and name it so it can be inserted inline in the prompt. The information resulting from the execution of the expression is added at runtime to the prompt to provide more specific responses.

You can also filter your knowledge using Power Fx expressions in addition to the current attribute and value selection. This enables you to get more control of the data to use as context of the prompt.

Draft with Copilot is a feature that allows you to write or rewrite your prompts using Copilot. Prompt builder makes it easy to trigger Copilot to help you author your prompt. Copilot can turn an intent into an effective, sophisticated prompt that follows prompt engineering best practices. Along with writing suggestions, Copilot provides optimization tips, a completeness score, and generates sample test data to help you run your validations faster.

This feature makes prompts actionable with external sources with the help of logs being allowed for validations and better exception handling. This feature also helps to overcome limitations of a native large language model (LLM). A native prompt builder experience in Microsoft Copilot Studio is built on fluent 9 (non-iFramed experience), removing the wizard for prompt actions by integrating the code interpreter elements to work with the new prompt builder experience.

Here are some tasks you can perform within this feature: You can enable and disable the toggle to generate the code from Adv. Setting. There's also a checkbox to write data to Dataverse tables in Adv. Setting. Open the Preview code to auto generate Python code and run it, with ability to edit the Python code and re-run the code. You can copy and paste the code in a code interpreter found on the internet with ability to undo code changes using an Undo button. Finally, you can write actions to Dataverse tables using the Run button when the write to Dataverse tables setting is turned on.

The prompt accuracy scoring feature in AI Builder’s prompt builder allows you to build a test suite and validate your prompt performance across different iterations of prompt development. These detailed assessments empower you to make the correct decisions about using prompts in agents, apps, and flows, moving capabilities to production, and prompt improvements. This comprehensive feedback on the effectiveness of your AI prompts helps you optimize for clarity and precision. As you create or refine prompts, the feature analyzes the prompt structure, language, and relevance to the intended task, and assigns a confidence score to each test case prediction that reflects in expected performance of the prompt. This score is generated based on factors such as specificity, complexity, alignment, and custom assertions. Thus, you can derive actionable insights to improve prompt phrasing or reduce ambiguity. By giving you a clear, quantifiable measure of prompt quality, the accuracy scoring feature streamlines the prompt engineering process, enhances model outcomes, and reduces iteration time. This enables more efficient and reliable AI interactions across use cases.

This capability empowers organizations to validate the functionality and effectiveness of generative AI prompt actions across agents, apps, and workflows. By enabling you to create systematic test datasets, define evaluation criteria, and run batch testing on prompt actions, businesses can validate outcomes and make data-driven decisions regarding the maturity or further development of these prompts. Prompt actions serve as intelligent operations within workflows, ensuring consistent and reliable outcomes. Early issue identification mitigates risks, minimizes disruptions, and enhances reliability, while accelerating deployment timelines and streamlining workflows. Ultimately, this capability drives smarter, AI-powered solutions that enhance user satisfaction, improve operational efficiency, and align with broader organizational goals.