Automated Contract Processing and Rights Analysis Using Multi-Model LLM Pipeline

condé_nast · Media & Entertainment · 2025 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.4 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: Contract processing and rights analysis

Condé Nast needed to process contracts, identify rights, match templates, and support licensing decisions faster while maintaining legal accuracy, regulatory alignment, and human expert oversight.

Constraints:

• contracts span multiple brands, geographies, rights, and licensing agreements
• manual review was time-consuming and prone to human error
• source PDFs can contain handwriting, strikethroughs, multi-column formatting, and hundreds of pages
• the system cannot provide legal advice by itself
• low-similarity contracts need human template development rather than forced matching

Architecture

AWS and Condé Nast built a Step Functions-orchestrated pipeline using S3, EventBridge, SageMaker Processing, Amazon Bedrock, Claude 3.7 Sonnet, Bedrock Knowledge Bases, Lambda, clustering, spreadsheets, and visualizations. A contract upload triggers the workflow. Claude’s visual reasoning converts PDF page images to text, a second Claude step extracts metadata from a prompt-provided schema, a third step compares the contract to templates in a Bedrock knowledge base and summarizes semantic differences, humans review the results, valid outputs load into the rights/royalties system, and low-similarity contracts are clustered so experts can draft new templates for future runs.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  contract-upload-s3["contract-upload-s3"] --> eventbridge-trigger["eventbridge-trigger"]
  eventbridge-trigger["eventbridge-trigger"] --> step-functions-workflow["step-functions-workflow"]
  step-functions-workflow["step-functions-workflow"] --> pdf-text-preprocessing-job["pdf-text-preprocessing-job"]
  contract-upload-s3["contract-upload-s3"] --> pdf-text-preprocessing-job["pdf-text-preprocessing-job"]
  pdf-text-preprocessing-job["pdf-text-preprocessing-job"] --> claude-bedrock-model["claude-bedrock-model"]
  pdf-text-preprocessing-job["pdf-text-preprocessing-job"] --> processed-text-s3["processed-text-s3"]
  step-functions-workflow["step-functions-workflow"] --> metadata-extraction-job["metadata-extraction-job"]
  processed-text-s3["processed-text-s3"] --> metadata-extraction-job["metadata-extraction-job"]
  metadata-extraction-job["metadata-extraction-job"] --> claude-bedrock-model["claude-bedrock-model"]
  metadata-extraction-job["metadata-extraction-job"] --> results-spreadsheet-s3["results-spreadsheet-s3"]
  step-functions-workflow["step-functions-workflow"] --> template-comparison-job["template-comparison-job"]
  processed-text-s3["processed-text-s3"] --> template-comparison-job["template-comparison-job"]
  template-knowledge-base["template-knowledge-base"] --> template-comparison-job["template-comparison-job"]
  template-comparison-job["template-comparison-job"] --> claude-bedrock-model["claude-bedrock-model"]
  template-comparison-job["template-comparison-job"] --> results-spreadsheet-s3["results-spreadsheet-s3"]
  template-comparison-job["template-comparison-job"] --> low-similarity-clustering["low-similarity-clustering"]
  results-spreadsheet-s3["results-spreadsheet-s3"] --> human-legal-reviewer["human-legal-reviewer"]
  human-legal-reviewer["human-legal-reviewer"] --> rights-royalties-system["rights-royalties-system"]
  low-similarity-clustering["low-similarity-clustering"] --> results-spreadsheet-s3["results-spreadsheet-s3"]
  low-similarity-clustering["low-similarity-clustering"] --> human-legal-reviewer["human-legal-reviewer"]

Components

Data flow

A user uploads contracts to S3. EventBridge starts the Step Functions workflow. SageMaker Processing converts PDFs to text using Claude 3.7 Sonnet visual reasoning and saves intermediate text to S3 under the same access restrictions. A second SageMaker job extracts schema-defined metadata fields with Claude. A third job compares contract text against templates in a Bedrock knowledge base, identifies semantic differences, and writes spreadsheet results. Human reviewers validate results, Lambda loads valid data to the rights/royalties system, and low-similarity contracts are clustered with similarity scores and visualizations to support new template drafting.

Why this architecture

The pipeline separates document preprocessing, metadata extraction, template comparison, human validation, rights-system loading, and low-similarity clustering because each stage has different failure modes and review needs. Claude visual reasoning handles complex PDFs, schema prompting constrains metadata extraction, the knowledge base supports template comparison, and humans remain responsible for validation and new-template drafting. Step Functions and S3 provide orchestration and governed intermediate storage.

Models used: Amazon Bedrock · Anthropic Claude 3.7 Sonnet · Amazon Bedrock Knowledge Bases · AWS Step Functions · Amazon SageMaker Processing · Amazon S3 · Amazon EventBridge · AWS Lambda

Scale

• Condé Nast manages contracts across multiple brands and geographies.
  “spanning multiple brands and geographies”article, stated
• The source says contracts can be large hundred-page documents.
  “handle large, hundred-page documents by dividing it into smaller chunks”article, stated
• Contract analysis processing time dropped from weeks to hours.
  “Processing time for contract analysis has been reduced from weeks to hours”article, stated
• The system is designed for high-volume periods such as brand acquisitions or expansions.
  “increased workloads during high-volume periods, such as major brand acquisitions or expansions”article, stated
• Condé Nast reaches hundreds of millions of consumers across platforms.
  “reaching hundreds of millions of consumers across print, digital, and social platforms.”summary, stated

Alternatives considered (roads not taken)

Continue manual rights review and template matching

Rejected because: Manual review was slow, error-prone, and contributed to conservative rights usage and missed revenue opportunities.

Force low-similarity contracts into existing templates

Rejected because: Low-similarity contracts are routed to clustering and human template creation instead of pretending a poor match is valid.

Use AI as an autonomous legal advisor

Rejected because: The solution is scoped to parsing, comparison, and visualization, with human legal and business review validating results.

"not providing legal advice of its own. The solution uses the following key services:

• Amazon Simple Storage Service (Amazon S3) – A scalable object storage service used to store incoming contracts, reference templates, and solution outputs."

— article, stated

Patterns

prompt-chaining (workflow)

The pipeline chains OCR-like transcription, schema extraction, template comparison, and clustering rather than asking one prompt to do all rights analysis.

structured-output-contract (extraction)

Metadata extraction uses a schema prompt containing each field and description.

rag-retrieval (retrieval)

Template comparison searches possible templates stored in an Amazon Bedrock knowledge base.

human-in-the-loop-gate (legal-review)

Human reviewers validate system results and draft new templates for low-similarity contract clusters.

durable-execution (workflow)

AWS Step Functions orchestrates distributed processing stages and downstream routing.

data-quality-loop (data-foundation)

The project found preprocessing and high-quality representative contracts were essential to downstream extraction and matching accuracy.

Tradeoffs

Use AI for parsing/comparison, not legal advice

Chose AI-augmented rights analysis with human validation over fully autonomous legal decisioning.

Rationale: This preserves legal expertise and compliance while automating the bottleneck-heavy parts of contract processing.

Roll out incrementally

Chose subset of contracts for specific brands over full-scale deployment from the start.

Rationale: Phased rollout enabled rapid iteration and real-world feedback before broader deployment.

Let subject matter experts drive prompt engineering

Chose plain-English expertise encoded into prompts over traditional software cycles for every rights rule.

Rationale: Experts could express requirements directly to language models, improving accuracy and reducing development complexity.

Gotchas (what bit them)

Metadata extraction quality depended heavily on preprocessing.

Root cause: Contracts included handwritten notes, scans, multi-column PDFs, large page counts, and specialized formatting.

Resolution: Build advanced preprocessing/OCR with Claude visual reasoning and chunk large documents.

AI rights analysis still needed expert review.

Root cause: Nuanced legal and rights cases require human expertise and continuous prompt refinement.

Resolution: Keep human validation and feedback loops in the workflow.

Template matching depends on representative historical contracts.

Root cause: The system needs diverse examples spanning brands and geographies to compare new contracts accurately.

Resolution: Maintain high-quality contract archives and route low-similarity contracts to cluster-based template development.

Evaluation

The workflow is validated by human review, operational processing-time reduction, improved rights accuracy, reduced rights-violation risk, high-volume scalability, and prompt-refinement feedback from legal and rights experts.

Metrics: processing time reduced from weeks to hours · improved accuracy · risk of rights violations reduced · consistent processing during high-volume periods · human validation results · quality of representative historical contracts

Outcomes

• contract analysis time dropped from weeks to hours
• rights management experts can focus on complex cases and strategic initiatives
• the system scales during brand acquisitions and expansions without additional human resources
• contract analysis became more accessible across the organization
• the implementation produced byproducts such as plain-language rights availability support

The source reports multiple brands/geographies, hundred-page document handling, high-volume acquisition/expansion periods, and a weeks-to-hours processing-time improvement.