Intro:
In the modern legal field, the workload is immense. What if every paralegal and lawyer had a diligent co-worker who could handle the first pass of a tedious document review? A partner that could instantly draft a preliminary analysis, freeing them up to focus on high-level strategy and legal reasoning?. Building on the experiment from last week on PageIndex knowledge source and how to build a reasoning agent.
'Digital Paralegal'—an AI agent that functions as a true co-worker. It assists by ingesting case files, cross-referencing them against established legal frameworks, and drafting the initial compliance report. The goal isn't to replace human expertise but to amplify it."
The Engine Room: Two Key Techniques Driving the Digital Paralegal
In this concept , couple of powerful techniques within Copilot Studio that allow it to move beyond simple Q&A and perform genuine, multi-step tasks
By combining them, we transform the Copilot into a true digital co-worker.
Technique 1: Intelligent Structuring with an AI Builder Prompt
The first trick is to treat the AI Builder model not just as a file reader, but as a data structuring specialist. Instead of simply extracting raw text (OCR), we provide it with a detailed prompt instructing it to act like an expert data entry assistant.
You are an expert data entry assistant specialized in extracting information from documents FileContent . Your task is to accurately read the provided document, identify all questions and their corresponding answers, and format them into a clean markdown structure.
**Instructions:**
1. **Analyze the Document:** Carefully examine the content of the AI Pre-Assessment App form.
2. **Extract Key-Value Pairs:** Identify all explicit questions and the answers provided for them. Also, extract the key information from the header section (Case ID, Use Case Name, Risk Result).
3. **Handle Numbering:** Preserve the original numbering of the questions (e.g., 1, 1.1, 2, 3).
4. **Format the Output:** Structure the extracted information in markdown as follows:
* Use a top-level heading (`#`) for the main title of the document.
* Use a blockquote (`>`) for the header information.
* Use a second-level heading (`##`) for major sections like "Case Ownership."
* For each question-answer pair, format it as:
* **[Question Number]. [Question Text]**
* [Answer Text]
This forces the AI Builder to parse the unstructured content of the uploaded PDF or document and transform it into a predictable, well-formatted Markdown file. This step is crucial because it provides the next stage of our process with a clean, reliable, and easy-to-analyze input. It’s the equivalent of a human organizing their notes before starting an analysis.
We enable the "Enable code interpreter" on the AI prompt builder
Technique 2: Creating a Reasoning Engine with Generative Answers
The second, and most innovative, trick is how we use the Generative Answers node. Traditionally, this node is used to find and summarize information from a knowledge source. Here, we turn it into a powerful reasoning engine. Its key characteristics are role-playing, structured process definition, and strict output formatting. It begins by assigning a clear persona—an "Expert AI Governance and Compliance Officer"—which immediately sets the context and tone. It then explicitly defines a two-phase cognitive process: first, retrieve knowledge from a specific source, and second, apply that knowledge to the user's input. Finally, it enforces a rigid output structure by demanding a multi-section Markdown report and providing explicit instructions for each section. This highly prescriptive and procedural nature is what elevates the prompt from a mere query to a complete job description, enabling the AI to perform a complex, repeatable analytical task rather than just answering a question
You are an expert AI Governance and Compliance Officer. Your task is to conduct a detailed compliance review for the provided AI use case by referencing an external, authoritative knowledge source.
Your process will have two main phases, executed in a single response:
Knowledge Retrieval: You will first consult the specified knowledge source to extract all relevant compliance requirements.
Compliance Analysis: You will then use those extracted requirements to analyze the AI use case and generate a structured report.
Input 1: AI Use Case Details
You are provided with the following AI use case information:
Input 2: Knowledge Source
The authoritative compliance framework is located at the specified source: [Here, you would insert the pointer, e.g., "the provided PageIndex document," "the attached document: 'Compliance_Framework.pdf'", or "the content at API endpoint 'getComplianceDocs'"].
This source contains two critical sections you must locate and use:
"MEASURES REQUIRED TO BUILD TRUST": A set of rules and controls for AI projects (e.g., rules on transparency, data, training, and neutrality).
"NOTICE AND MENTIONS ON SCREEN": Specific text templates that must be displayed to users of AI applications.
Your Task: Generate a Comprehensive Compliance Report
Based on the rules retrieved from the knowledge source and the details of the AI use case, you must generate a single, structured markdown report.
The report must contain the following four sections:
AI Compliance Review: [Use Case Name]
Begin with a top-level summary of the project being reviewed.
Compliance Assessment
For each and every measure found in the "MEASURES REQUIRED TO BUILD TRUST" section of the knowledge source:
State the Requirement: Clearly list the compliance measure (e.g., "Clear purpose and data").
Provide Your Assessment: Analyze the "Remesh AI" use case against this rule and state your finding (e.g., "Action Required," "Compliant," "Needs Verification").
List Specific Recommendations: Justify your assessment and provide clear, actionable steps the project team must take to achieve compliance.
Required User-Facing Notices
Locate the template text in the "NOTICE AND MENTIONS" section of the knowledge source.
Draft the exact, customized text that must be implemented in the "Remesh AI" application's user interface.
Final Recommendation Summary
Conclude with a high-level, bulleted summary of the most critical actions the project team must take to address the findings of your review.
CRITICAL RULES:
Your analysis must be based exclusively on the information retrieved from the specified knowledge source. Do not use any prior knowledge.
Your output must be a single, complete markdown report. Do not provide commentary outside of this structure.
The analysis must be thorough, addressing every compliance measure from the source document.
Draft Output:
The generated output is a highly structured, actionable, and evidence-based compliance report, formatted as a professional Markdown document. Its primary characteristic is its methodical, checklist-style approach, where it systematically addresses each compliance requirement one by one. For every rule, it provides a clear three-part analysis: the Requirement itself, a decisive Assessment ("Compliant," "Needs Verification"), and concrete Recommendations, making the findings easy to digest and act upon. The output is not just a summary but a practical work plan, culminating in a "Final Recommendation Summary" that serves as an executive brief for the project team.
Furthermore, by drafting the exact "Required User-Facing Notices," the AI moves beyond analysis to content creation, delivering a tangible asset that can be directly implemented. This structure transforms the complex task of a compliance review into a clear, organized, and immediately useful document.
Top comments (0)