Business software provider achieves $500k annual savings with AI-powered document processing at scale

At a glance

Intelligent Document Processing with GenAI at Scale

Mantel partnered with an iconic Australian business software provider to design and rapidly deliver a low-latency, high-accuracy, AI-powered OCR solution that automatically extracts data from invoices and deliveries.

Key takeaways
  • Extensible solution for more document processing use cases
  • Highly accurate AI-powered OCR solution
  • Scalable solution that can adapt to seasonal peak demand
  • Extensible solution for more document processing use cases
Key services
  • Generative AI (GenAI)
  • AI-Powered OCR (Optical Character Recognition)
  • AWS cloud architecture: Bedrock and Textract
  • Asynchronous processing and queuing
  • Real-time monitoring

About our client

Our client is an Australian business software company providing accounting, payroll, and business management solutions for small and medium-sized businesses.

High costs and inconsistent extraction

Our client sought to optimise its high-volume operations by replacing an expensive third-party OCR solution. Processing more than 3.5 million documents per month, our client aimed to achieve a significant performance uplift: improving accuracy and strictly adhering to a Service Level Agreement (SLA) of less than or equal to 7 seconds per document, particularly for the accurate extraction of diverse data fields, a requirement the previous solution struggled to meet.

Key challenges:

  • Expensive reliance on third-party OCR.
  • High throughput (~3.5M documents/month) and strict 7-second, or less, SLA.
  • Accuracy gap when extracting complex data fields.

Building a native, AI-enabled AWS architecture

Mantel’s partnership with our client delivered a cost-efficient migration of the OCR pipeline to a resilient, AI-enabled architecture on AWS. This new system features asynchronous processing and a queuing mechanism, ensuring high speed and resilience to handle sudden increases in traffic. To guarantee stability during peak periods, the Mantel team secured higher Large Language Model (LLM) throughput quotas from AWS. 

Furthermore, real-time monitoring with alerts was introduced for proactive issue management. The rollout was secured through a robust validation and release process, which included using real production data, canary releases, and rigorous performance testing to eliminate errors and appropriately size the system.

Scalability, accuracy, and substantial savings

The implementation delivered substantial value, starting with a significant financial impact of approximately $500,000 in annual cost savings for our client. Critically, our client also gained full IP ownership of the new AI solution, making it reusable across various future document processing use cases. 

The project successfully achieved all targets and Service Level Agreements (SLAs). Key performance metrics included processing more than 350 documents per minute, maintaining an average response time of approximately 7 seconds, and reaching an accuracy of around 90%.