Enterprise document processing has long suffered from a fundamental flaw: systems that can't learn from their mistakes without extensive retraining. Melbourne-based Affinda claims to have solved this persistent industry challenge with the launch of its new agentic AI intelligent document processing (IDP) platform featuring "persistent model memory."

The breakthrough addresses what Affinda positions as a critical gap in the current IDP market, where organizations must choose between accuracy, speed, and adaptability. While the company claims competitors require weeks or months of fine-tuning with dozens of examples, Affinda's platform is able to recall any corrections that were required for processing individual document types and automatically apply the corrections the next time a similar document is presented.

This represents a significant departure from the approaches used by established IDP providers, according to Andrew Bird, Affinda's Head of AI, who emphasizes that the platform's ability to remember and apply corrections without human intervention sets it apart from traditional LLM-powered document processing tools that rely on static prompts.

“If a human corrects a model in the Affinda Platform for a particular document, this correction is added to the platform’s model memory and it instantly learns not to make that mistake again.  Systems that rely on fine-tuning smaller models will likely require some number of examples (usually 20+) before the model is likely to learn from its mistakes in the next training run. It should come as no surprise that an approach that leverages frontier models is becoming table stakes in IDP,” said Bird.

(A “frontier model” refers to the most advanced, state-of-the-art large language models that represent the cutting edge of AI capabilities, e.g. GPT-4o and o1 from OpenAI, Claude 4 (Opus and Sonnet) from Anthropic and Gemini Ultra/Pro from Google DeepMind.)

The technology addresses persistent challenges in digital transformation initiatives where document-heavy processes remain manual bottlenecks. Traditional approaches typically require extensive training periods and ongoing maintenance, while standalone large language models face performance limitations.

The Affinda platform operates from data centres in Sydney, Frankfurt and Oregon, offering both cloud-based and self-hosted deployment options. Affinda is ISO27001 certified and GDPR compliant, and will soon be SOC2 accredited, meeting enterprise security and privacy requirements.

Anthony England, Head of Growth, Affinda, said: “Until now, enterprises that wanted to automate their document-heavy processes were forced to compromise on accuracy, time to value or flexibility. Now they can get all three. Affinda’s platform can accurately extract any information from any document, in any format, fast. It’s world class, and our customers are blown away by how quickly they can get set up and how powerful the platform is.

“Using the Affinda platform, organizations can now build and customize document processing models themselves in just a few clicks, paying only for what they use. By removing the need for extensive DIY and in-house R&D, we’re taking the guesswork out of automating document workflows, and we’re giving time back to organizations to focus on core projects and innovation.”

The Affinda platform combines frontier large language models hosted on AWS and Azure with internally fine-tuned smaller models for pre- and post-processing. The system uses Microsoft Azure's OCR engine and Elasticsearch for vector searches in its retrieval-augmented generation pipeline.

Affinda claims organizations can build and deploy models in minutes rather than months, achieving what the vendor states is over 99% accuracy across all document types. The platform supports 56 languages and integrates with more than 400 enterprise systems.

Two enterprise customers have detailed their implementations. Insurance provider StateCover Mutual processes over 300,000 documents annually across 80 document types using the platform. Global logistics company Northline reports processing 120,000 proof-of-delivery and related documents with 82% straight-through processing rates.

"Automating validation and document handling across all 13 depots has reduced manual effort, errors and delays," said Jorg Both, Northline's head of business systems.

Regarding competitive positioning, Affinda claims its approach outperforms established players including Microsoft's Azure AI Document Intelligence and Google's recently released Document AI platform.

Microsoft's Azure service relies on fine-tuning smaller models rather than LLMs, according to Affinda, while Google's new platform uses LLM fine-tuning with what the vendor characterizes as slower learning cycles compared to its instant model memory updates.

The platform targets organizations managing compliance-heavy document workflows, offering consumption-based pricing and a free trial with access to all features.

https://www.affinda.com