CrewAI Contract clause risk assessment assistant

This project implements a CrewAI agent team designed to assist in the preliminary review of legal contracts. The agents work together to parse contract text, classify clauses, identify potential risk patterns based on predefined criteria, spot ambiguities, and generate a structured "Review Brief" highlighting areas that may warrant closer inspection by a legal professional.

🚨 CRITICAL DISCLAIMER 🚨

This tool is an AI-powered assistant ONLY. It DOES NOT provide legal advice.

• The analysis is automated and based on pattern matching and AI interpretation, which can be flawed, incomplete, or inaccurate.
• The output MUST be thoroughly reviewed and validated by a qualified human legal professional before making any decisions or taking any action based on the contract.
• DO NOT rely solely on this tool for legal assessment or contract negotiation. It is intended solely to help focus human review efforts.
• Use of this tool is at your own risk. The creators assume no liability for errors, omissions, or consequences arising from its use.

Features

• Contract parsing: Attempts to segment contract text into distinct clauses or sections.
• Clause classification (optional): Categorizes clauses by common types (e.g., Liability, Term, Payment).
• Risk pattern detection: Scans clauses for predefined keywords or structures often associated with higher risk (e.g., broad indemnities, liability limitations, vague warranties).
• Ambiguity identification: Looks for undefined critical terms, vague phrasing, or potentially contradictory statements.
• Structured reporting: Consolidates flagged items into a "Review Brief" organized for efficient human review.

Requirements

• Python 3.9+
• Google Colab Environment (Recommended for ease of use) or a local Python setup.
• OpenAI API Key: Required for the language model.
  • GPT-4 (Turbo recommended): Strongly advised due to the nuance required for legal text analysis. GPT-3.5 may struggle significantly.
  • Requires a funded OpenAI account or active trial credits. Get a key from platform.openai.com.

Setup (Google Colab)

1. Get the Notebook: Clone the repository or download the .ipynb file.
2. Open in Colab: Upload and open the notebook in Google Colab (colab.research.google.com).
3. Install libraries: Run the first cell (# @title 1. Install Necessary Libraries) to install crewai, langchain-openai, etc.
4. Configure OpenAI API Key in Colab Secrets:
   • In Colab, click the Key icon in the left sidebar ("Secrets").
   • Enable "Notebook access".
   • Add one new secret:
     • Name: OPENAI_API_KEY
     • Value: Your OpenAI API Key (starts with sk-...)
   • Ensure the toggle switch next to the secret is ON.
5. Run Cell 2: Execute the second cell (# @title 2. Import Modules...). Check the output to confirm the OPENAI_API_KEY was found and loaded. Resolve any errors before continuing.

How to Use

1. Input contract text (Cell 3):
   • Locate the cell titled # @title 3. Define Input Contract Text....
   • Replace the example text within the triple quotes ("""...""") assigned to the contract_text variable with the actual contract text you want to analyze. Try to preserve formatting like numbering if possible.
   • Acknowledge the disclaimer printed by this cell before proceeding.
2. Select LLM (Optional - Cell 4):
   • In Cell 4 (# @title 4. Select LLM...), ensure you are using an appropriate model. GPT-4 Turbo is highly recommended.
3. Run cells sequentially: Execute the remaining cells (4 through 8) in order.
   • Cell 5 defines the specialized agents for contract analysis.
   • Cell 6 defines the sequence of tasks (Parse, Classify, Detect Risks, Identify Ambiguities, Generate Brief).
   • Cell 7 creates the Crew and starts the analysis using kickoff(). This step will involve significant LLM processing and API calls. Monitor the verbose output.
   • Cell 8 displays the final generated "Review Brief". Remember to read the final disclaimer printed by this cell as well.

Workflow overview

The crew processes the contract text through these stages:

1. Contract parser agent (Task 1): Segments the raw text into potential clauses.
2. Clause classifier agent (Task 2): Assigns a category to each clause (optional step, aids focus).
3. Risk pattern detector agent (Task 3): Flags clauses matching predefined risk patterns.
4. Ambiguity identifier agent (Task 4): Flags clauses with unclear or vague language.
5. Review brief generator agent (Task 5): Compiles all flagged items from Tasks 3 & 4 into the final report.

Customization

• Risk patterns: Modify the goal description of the risk_pattern_detector agent (Cell 5) and the corresponding task_detect_risks description (Cell 6) to add, remove, or refine the specific risk patterns you want to flag.
• Ambiguity criteria: Adjust the prompts for the ambiguity_identifier agent and task to focus on different types of clarity issues.
• Input format: For real-world use, you would likely need to add tools (e.g., using libraries like pypdf or python-docx) and modify the contract_parser agent/task to handle PDF or DOCX files instead of just plain text pasted into the notebook.
• Clause classification: You could remove the clause_classifier agent and task_classify_clauses if desired, simplifying the flow but potentially making the risk/ambiguity detection slightly less focused. Adjust subsequent task contexts accordingly.

Limitations & critical disclaimer (Reiteration)

• THIS IS NOT LEGAL ADVICE. Output is AI-generated and requires expert human review.
• Accuracy is not guaranteed: LLMs can misinterpret legal language, miss nuances, or "hallucinate" findings. Parsing of complex formatting may fail.
• Limited scope: The default risk patterns are examples; they are not exhaustive and may not cover all relevant risks for a specific contract type or jurisdiction.
• Context ignored: The tool analyzes text in isolation and lacks understanding of the business context, negotiation history, or applicable laws beyond what's in the text and the LLM's general knowledge.
• API costs: Using powerful models like GPT-4 can be expensive. Monitor your OpenAI usage.