Challenge

A debtor finance provider serving Australia's SME sector had hit the limits of their SQL-based automation. The system couldn't handle modern payment complexity, multi-party transactions, diverse invoice formats, and unstructured financial communications. Knowledge workers spent up to 20 minutes manually reconciling each complex payment, correlating data across emails, invoices, and multiple banking systems. These bottlenecks delayed credit access for clients and created allocation errors that impacted cash flow.

The real problem wasn't processing speed alone. Traditional rule-based systems break when payment structures don't match predefined patterns. Every non-standard invoice format required manual intervention. Every multi-party payment needed human judgment to allocate correctly. The client needed a system that could understand financial context the way their experienced staff did—extracting entities from unstructured text, resolving ambiguities, and matching data across disparate systems but at machine speed and scale.

"In debtor finance, every minute of delay in payment allocation impacts our clients' working capital. Our SQL-based automation had reached its limits with the increasing complexity of modern payment structures. When we evaluated LLMs' ability to understand unstructured financial documents and correlate data across multiple systems, we saw an opportunity to fundamentally reimagine our payment processing capabilities." - Chief Technology Officer

Goals

• Automate complex payment allocations that currently require manual correlation across emails, invoices, and multiple systems, reducing processing costs while improving accuracy
• Credit client funds before office hours instead of late afternoons
• Establish production patterns for LLM-based financial operations that handle real-world document variety and maintain audit-ready accuracy

Solution

Evolve's engineering team built a two-stage system that processes unstructured financial communications through LLM-based entity extraction, followed by similarity-based entity resolution against structured ledger data. 

The architecture works like this: emails and invoices, regardless of format, flow through OpenAI models that extract financial entities: debtors, payees, invoice numbers, and amounts. These entities then match against bank ledgers and back-office systems using text similarity algorithms optimised for financial data. We deployed this on DataRobot integrated with Azure infrastructure, building automated pipelines, agentic workflows, and real-time monitoring dashboards.

The system includes intelligent caching of previously processed documents to reduce API costs and LLMOps practices for version control and performance monitoring. We designed the architecture to scale processing capacity on demand while maintaining full visibility into allocation accuracy through governance dashboards. Each component handles specific failure modes—the LLM layer deals with document format variation, the resolution layer manages entity ambiguity, and the monitoring layer catches edge cases that need human review.

We delivered this in 12 weeks, moving from initial concept to processing live transactions. The modular design means the client can extend coverage to additional document types PDFs, spreadsheets, and scanned images, without rebuilding core infrastructure.

Impact

Processing time dropped from 20 minutes to 10 seconds per debtor communication, a 120x improvement. This acceleration runs continuously, processing payments outside business hours and eliminating the manual bottlenecks that previously constrained throughput. The client now credits funds before office hours instead of after lunch, directly improving their SME clients' access to working capital.

Beyond speed, the system handles document variety that broke the previous rule-based automation. It processes emails in different formats, invoices with inconsistent layouts, and multi-party payments that require contextual interpretation. Allocation accuracy improved while processing capacity scaled to meet volume demands.

The implementation demonstrates what production LLM deployment requires in financial operations: entity extraction that handles real-world document chaos, resolution algorithms that match across data quality gaps, cost management through intelligent caching, and monitoring systems that flag edge cases. The client established these patterns in 12 weeks and can now extend them to additional financial workflows.

"What matters most isn't just the 120x faster processing, though that directly impacts our clients' cash flow. We've built a system that understands complex financial documents the way our experienced staff does, but processes them at scale. Our operations teams now have AI-powered tools that handle document variety and payment complexity that previously required manual interpretation. This proves that established financial services companies can deploy LLMs in production for critical operations when the architecture accounts for real-world constraints." - Chief Operating Officer

The entity extraction techniques deployed in this system, using LLMs for token classification instead of rule-based pattern matching, apply across industries where unstructured documents require structured data extraction. We've documented these patterns and their enterprise implementation considerations in our blog post: Accelerating knowledge processes with LLMs.