AI/ML Powered Document Workflow Transformation for Leading Australian BFSI Firm

Overview:

This customer is one of Australia's major national full-service retail stockbroking and wealth management firm. With a legacy of over 30 years, the firm provides comprehensive financial services including investment advice, wealth management, and corporate finance. To maintain its high standards of client service while scaling operations, the firm sought to modernize and automate its document processing workflows using artificial intelligence (AI) and machine learning (ML). They relied heavily on manual workflows to process thousands of client documents each month, involving time-consuming tasks like sorting, reviewing, redacting sensitive data, and rekeying information into downstream systems. This approach was labor-intensive, and struggled to scale with increasing volumes of ID proofs, account applications, statements of advice, and trust deeds.

Minfy Technologies was engaged to develop a prototype leveraging AWS Intelligent Document Processing (IDP) capabilities. The objective was to integrate AWS services with the customer’s Salesforce platform, enabling automatic document classification, Personally Identifiable Information (PII) detection and redaction, and structured data extraction. The solution was designed to handle scanned documents from multiple formats and orientations and display results via an AWS QuickSight dashboard.

‍

Customer Challenges

From a business perspective, the customer faced the following key challenges:

• Manual Document Handling: Processing over 18,000 documents a month, the firm relied heavily on manual workflows for classification, data entry, and redaction, which was time-consuming and error-prone.
• ‍Compliance and Risk Exposure: Ensuring PII was consistently and accurately redacted across diverse document types was becoming increasingly complex, exposing the firm to potential regulatory and data privacy risks. ‍
• Inconsistent Data Quality: The lack of structured data extraction limited the customer’s ability to analyze and act on critical client information stored in documents such as ID cards, account applications, and statements of advice. ‍
• Integration Limitations: Existing systems were not seamlessly integrated with downstream platforms like Salesforce, resulting in inefficiencies and delayed access to client data.

‍

Solution

Minfy delivered a functional prototype of an AI/ML-powered intelligent document processing solution using a combination of AWS services. The end-to-end pipeline was designed to automatically classify, extract, redact, enrich, and surface insights from scanned documents in various formats.

‍

Key Solution Components:

Capture and Storage: Multi-format documents (PDF, JPEG, TIFF) were ingested and stored in Amazon S3. Many of the documents are often scanned at inconsistent orientations, (upside down, sideways or skewed). This led to failed extractions and as such, Minfy implemented a preprocessing step, using Amazon Rekognition, to automatically detect and correct orientations of each image. This eliminated the need for manual correction, saving considerable time, spent by staff in manually rotating and re-uploading the documents.

Classification: Amazon Bedrock’s LLM and Amazon Comprehend were used to classify four initial document types—ID documents (e.g., driver’s license, passport), account applications, statements of advice, and trust deeds.

Extraction: Amazon Textract extracted structured and unstructured data, including key-value pairs and tables, from the documents.

PII Detection & Redaction: Amazon Comprehend and Comprehend Medical identified and redacted sensitive information like Tax File Numbers (TFNs) and identity details.

Data Enrichment: The extracted data was tagged with additional metadata to support downstream processing and analytics. One such example is entity detection and tagging — where documents containing personal identifiers like driver's license numbers or Tax File Numbers (TFNs) were automatically tagged for sensitive PII. This metadata was not only used for redaction but also to flag such documents for priority handling and compliance tracking downstream.

Dashboard & Validation: An Amazon QuickSight dashboard was built to display classification results, PII detection status, and extracted data. The GUI also allowed users to validate or override system-generated outputs, supporting continuous learning for the models.

Future Integration Readiness: Data processed through the pipeline was structured for future integration with Salesforce using APIs, with considerations for format, access, and security.

‍

Nuances of the Solution

A few technical challenges were encountered during implementation, which are summarized below, along with their workaround.

Orientation Issues with Identification Cards: Scanned ID documents came in various rotations, which hindered accurate text extraction.

Resolution: Custom logic was implemented using AWS Rekognition to detect and auto-correct orientation before processing with Textract.

Classification Errors from LLMs: Misclassifications occurred due to incomplete or incorrectly extracted data.

Resolution: A validation mechanism using JSON schema was introduced to cross-check LLM outputs and mitigate hallucinations.

Diverse Document Structures: Multiple document types appeared within a single PDF or varied significantly in structure.

Resolution: Custom prompt templates were developed to guide the LLM in consistently extracting and classifying content.

‍

Business Benefits

Minfy’s solution directly addresses the customer's operational and compliance challenges by delivering tangible results across accuracy, efficiency, and readiness for automation:

Reduced Manual Effort & Improved Operational Efficiency: With 100% classification accuracy and 90% text extraction accuracy across test documents, the solution demonstrates strong potential to replace manual document sorting and data entry processes. This directly supports the client’s goal to scale operations while reducing dependency on manual workflows.

Strengthened Compliance through Reliable PII Redaction: Achieving 97% accuracy in PII detection and redaction ensures the client can mitigate risks associated with handling sensitive client information, helping maintain adherence to data privacy and financial regulations.

Enhanced Data Quality & Decision-Making: Structured extraction of critical data from previously unstructured documents enables better downstream processing and analytics, addressing issues related to inconsistent data quality. For example, classification and redaction accuracy metrics (e.g., 97% PII detection) can now be tracked across document types to identify processing gaps and refine models.

Foundation for Seamless Platform Integration: The enriched and validated outputs are now Salesforce-ready, solving the challenge of disconnected systems and enabling faster access to accurate client information.

Scalable Model for Enterprise-wide Adoption: The success of the PoC on a targeted document set builds confidence in expanding the solution to over 77 document types, laying the groundwork for full-scale intelligent automation across the organization.