Transforming Mortgage Document Processing with Generative AI: A Case Study from Rocket Close

This heading encapsulates the essence of the document while highlighting the contributions and technological advancements made by Rocket Close in partnership with AWS.

Revolutionizing Mortgage Document Processing: Rocket Close and the Power of Generative AI

This post is co-written by Jeremy Little and Chris Day from Rocket Close.

In the dynamic world of mortgage services, efficiency and accuracy are paramount. Enter Rocket Close—a Detroit-based title and appraisal management company embedded within the Rocket Companies ecosystem. Recently, Rocket Close has transformed its mortgage document processing through automation, addressing challenges that threatened its operational efficiency and growth. This blog post delves into how a strategic partnership with the AWS Generative AI Innovation Center (GenAIIC) revolutionized their document processing workflow.

The Challenge: Manual Processing Bottlenecks

Processing approximately 2,000 abstract package files daily, with each file averaging 75 pages, Rocket Close faced significant operational hurdles. The intricate nature of this workload demanded extensive manual effort, averaging 10 hours per package. This labor-intensive process not only delayed workflow but also strained resource allocation and escalated costs dramatically.

Key Challenges:

• Volume Overload: Managing 2,000 abstract packages every day, each complex and document-intensive.
• Time-Intensive Workflow: Heavy manual processing resulted in bottlenecks and inefficiencies.
• Financial Impact: Manual processing costs added up to millions in annual expenditures.
• Scalability Limits: Growing demand outstripped the capabilities of manual processes.
• Quality Concerns: Human error introduced inconsistencies in data extraction.

With around 1,000 hours of manual processing effort required every day, Rocket Close urgently needed an innovative solution.

Redefining Abstract Document Processing

Abstract document packages include a variety of legal documents pertaining to property ownership, containing around 50–100 pages of diverse formats. The inconsistency in structure and quality adds layers of complexity to automated processing. These packages often include:

• Mortgage Documents: Contracts and agreements defining loan terms.
• Chain of Title Documents: Legal documentation recording property ownership transfer.
• Judgment Documents: Official records reflecting claims against property owners.
• Tax Documents: Filings that may represent potential claims for unpaid taxes.
• Legal Documents: Various court filings affecting property ownership.

Traditional processing methods struggled to keep pace due to the heterogeneous nature of these documents.

The Solution Architecture: Automation Powered by AI

Partnering with AWS GenAIIC, Rocket Close harnessed generative AI to automate its document processing workflow. Utilizing Amazon Textract for Optical Character Recognition (OCR) and Amazon Bedrock for intelligent information extraction, the teams engineered a two-stage process that achieved remarkable efficiency.

Solution Breakdown:

1. OCR Processing: Amazon Textract converts document images into machine-readable text, maintaining the structural integrity of the documents for easy reference.

2. Information Extraction: Amazon Bedrock employs foundation models to analyze the extracted text, segment the documents, and structure relevant data into a usable format.

Innovative Techniques:

• Advanced Prompt Engineering: Crafted prompts guide the AI’s behavior to enhance segmentation accuracy and data extraction.

• Domain-Specific Knowledge Integration: Industry terminology is incorporated into the AI models through data dictionaries and glossaries, boosting extraction precision.

• Evaluation Framework: Customized metrics assess extraction quality, ensuring high standards for diverse data field types.

Through these innovative strategies, Rocket Close achieved a processing efficiency that is 15 times faster than the previous manual methods.

Results and Impact: Operational Transformation

The implementation of this solution has proven transformative. Here are some remarkable outcomes:

• Speed: Processing time per package has been slashed from an average of 30 minutes to under 2 minutes.

• Cost Reductions: The automation has led to significant savings per document, culminating in potential annual savings at scale.

• Consistency and Quality: An impressive 90% accuracy rate has been maintained, greatly reducing human error and standardizing output formats.

• Scalability: The new architecture can handle over 500,000 documents annually, allowing Rocket Close to grow without increasing manpower proportionally.

Lessons Learned: Insight for Future Implementations

The project unearthed several insights pivotal for similar automated document processing implementations:

1. Prompt Engineering is Key: Tailoring prompts based on domain knowledge enhances extraction accuracy.

2. Two-Stage Architecture: Separating OCR from LLM processing optimized efficiency and output quality.

3. Importance of Domain Knowledge: Integrating specialized terminology significantly bolstered results.

4. Complex Evaluation Metrics: Developing tailored metrics is essential for accurately measuring performance.

5. Scalability Considerations: Designing solutions with scalability in mind is critical for future growth.

What’s Next for Rocket Close

Following the successful proof of concept, Rocket Close is gearing up for the production phase. Key plans include:

• Containerized Deployment: Transitioning to a scalable production environment.

• Continuous Improvement: Establishing feedback loops to adapt to evolving document patterns.

• Expansion into Other Workflows: Extending automated processing to additional document types beyond abstracts.

Conclusion

The collaboration between Rocket Close and AWS showcases the transformative potential of generative AI in document-intensive industries. By automating the intricate task of abstract document processing, Rocket Close has not only streamlined operations but also positioned itself for sustained growth and innovation. As they move towards a full production rollout, the foundation established during this project stands to further revolutionize their workflows, ensuring efficiency and excellence in service delivery.

About the Authors

Jeremy Little is the Lead Senior Solution Architect at Rocket Close, overseeing the design and implementation of innovative tech solutions in mortgage services.

Chris Day is the Vice President of Engineering at Rocket Close, leading engineering teams in streamlining title and appraisal management technologies.

(Additional author bios can be listed here if required.)