How Associa transforms doc classification with the GenAI IDP Accelerator and Amazon Bedrock

This can be a visitor submit co-written with David Meredith and Josh Zacharias from Associa.

Associa, North America’s largest neighborhood administration firm, oversees roughly 7.5 million householders with 15,000 workers throughout greater than 300 department places of work. The corporate manages roughly 48 million paperwork throughout 26 TB of knowledge, however their present doc administration system lacks environment friendly automated classification capabilities, making it troublesome to arrange and retrieve paperwork throughout a number of doc varieties. Daily, workers spend numerous hours manually categorizing and organizing incoming paperwork—a time-consuming, error-prone course of that creates bottlenecks in operational effectivity and probably leads to operational delays and diminished productiveness.

Associa collaborated with the AWS Generative AI Innovation Middle to construct a generative AI-powered doc classification system aligning with Associa’s long-term imaginative and prescient of utilizing generative AI to attain operational efficiencies in doc administration. The answer mechanically categorizes incoming paperwork with excessive accuracy, processes paperwork effectively, and offers substantial value financial savings whereas sustaining operational excellence. The doc classification system, developed utilizing the Generative AI Clever Doc Processing (GenAI IDP) Accelerator, is designed to combine seamlessly into present workflows. It revolutionizes how workers work together with doc administration methods by lowering the time spent on guide classification duties.

This submit discusses how Associa is utilizing Amazon Bedrock to mechanically classify their paperwork and to assist improve worker productiveness.

Answer overview

The GenAI IDP Accelerator is a cloud-based doc processing answer constructed on AWS that mechanically extracts and organizes data from numerous doc varieties. The system makes use of OCR know-how and generative AI to transform unstructured paperwork into structured, usable knowledge whereas scaling seamlessly to deal with excessive doc volumes.

The accelerator is constructed with a versatile, modular design utilizing AWS CloudFormation templates that may deal with various kinds of doc processing whereas sharing core infrastructure for job administration, progress monitoring, and system monitoring. The accelerator helps three processing patterns. We use Sample 2 for this answer utilizing OCR (Amazon Textract) and classification (Amazon Bedrock). The next diagram illustrates this structure.

We optimized the doc classification workflow by evaluating three key features:

• Immediate enter – Full PDF doc (all pages) vs. first web page solely
• Immediate design – Multimodal prompting with OCR knowledge (utilizing the Amazon Textract analyze_document_layout) vs. doc picture solely
• Mannequin selection – Amazon Nova Lite, Amazon Nova Professional, Amazon Nova Premier, and Anthropic’s Claude Sonnet 4 on Amazon Bedrock

This complete analysis framework helped us establish the configuration that delivers the very best accuracy whereas minimizing processing inference prices for Associa’s particular doc varieties and operational necessities. The analysis dataset consists of 465 PDF paperwork throughout eight distinct doc varieties. The dataset contains some samples recognized as draft paperwork or electronic mail correspondences. These samples are categorized as doc sort Unknown attributable to inadequate classification standards. The distribution of doc varieties throughout courses is unbalanced, starting from 6 samples for Insurance policies and Resolutions to 155 samples for Minutes.

Analysis: Immediate enter

We began our preliminary analysis utilizing full PDF paperwork, the place all pages of a PDF have been used as enter to the immediate for classification. The next desk exhibits the accuracy for full PDF classification utilizing Amazon Nova Professional with OCR and picture. We noticed a mean classification accuracy of 91% contemplating the completely different doc varieties with a mean value of 1.10 cents per doc.

Utilizing full PDF for doc classification demonstrates an accuracy of 100% for Certificates of Insurance coverage and 95% for Minutes. The system accurately categorized 425 out of 465 paperwork. Nonetheless, for the Unknown doc sort, it achieved solely 40% accuracy, accurately classifying simply 8 out of 20 paperwork.

Subsequent, we experimented with utilizing solely the primary web page of a PDF doc for classification, as proven within the following desk. This strategy improved total accuracy from 91% to 95% with 443 out of 465 paperwork categorized accurately whereas lowering classification value per doc from 1.10 cents to 0.55 cents.

Aside from improved accuracy and diminished value, the first-page-only strategy considerably improved Unknown doc classification accuracy from 40% to 85%. First pages usually comprise probably the most distinctive doc options, whereas later pages in drafts or electronic mail threads can introduce noise that confuses the classifier. Mixed with quicker processing speeds and decrease infrastructure prices, we chosen the first-page-only strategy for the following evaluations.

Analysis: Immediate design

Subsequent, we experimented on immediate design to guage whether or not OCR knowledge is important for doc classification or simply utilizing the doc picture is ample. We evaluated by eradicating the OCR textual content extraction knowledge from the immediate and solely utilizing the picture in a multimodal immediate. This strategy removes the Amazon Textract prices and depends solely on the mannequin’s understanding of visible options. The next desk exhibits the accuracy for first-page-only classification utilizing Amazon Nova Professional with solely picture.

The image-only classification strategy demonstrates comparable points as the total PDF classification strategy. Though this methodology achieves an total accuracy of 93%, for Unknown doc varieties, it may classify solely 10 out of 20 paperwork accurately with 50% accuracy. The next desk summarizes our analysis of an image-only strategy.

The image-only strategy removes OCR prices however reduces total accuracy from 95% to 93% and Unknown doc accuracy from 85% to 50%. Correct Unknown doc classification is vital for downstream human overview and operational effectivity at Associa. We chosen the mixed OCR and picture strategy to keep up this functionality.

Analysis: Mannequin selection

Utilizing the optimum configuration of first-page-only classification with OCR and picture, we evaluated completely different fashions to establish an optimum steadiness of accuracy and value, as summarized within the following desk. We give attention to total classification efficiency, classification of unknown paperwork, and per-document classification prices.

Total classification accuracy ranged from 95–96% throughout the fashions, with variation in unknown doc sort efficiency. Certificates of Insurance coverage, Plat Map, and Minutes achieved 98–100% accuracy throughout the fashions. Anthropic’s Claude Sonnet 4 achieved the very best unknown doc accuracy (95%), adopted by Amazon Nova Premier (90%) and Amazon Nova Professional (85%). Nonetheless, Anthropic’s Claude Sonnet 4 elevated classification value from 0.55 cents to 1.21 cents per doc. Amazon Nova Premier achieved the most effective total classification accuracy at 1.12 cents per doc. Contemplating the trade-offs between accuracy and value, we chosen Amazon Nova Professional because the optimum mannequin selection.

Conclusion

Associa constructed a generative AI-powered doc classification system utilizing Amazon Nova Professional on Amazon Bedrock that achieves 95% accuracy at a mean value of 0.55 cents per doc. The GenAI IDP Accelerator facilitates dependable efficiency scaling to excessive quantity of paperwork throughout their branches. “The answer developed by AWS Generative AI Innovation Middle improves how our workers handle and manage paperwork, and we foresee vital discount of guide effort in doc processing,” says Andrew Brock, President, Digital & Know-how Providers & Chief Info Officer at Associa. “The doc classification system offers substantial value financial savings and operational enhancements, whereas sustaining our excessive accuracy requirements in serving residential communities.”

Consult with the GenAI IDP Accelerator GitHub repository for detailed examples and select Watch to remain knowledgeable on new releases. Should you’d prefer to work with the AWS GenAI Innovation Middle, please attain out to us or go away a remark.

Acknowledgements

We wish to thank Mike Henry, Bob Strahan, Marcelo Silva, and Mofijul Islam for his or her vital contributions, strategic choices, and steering all through.

Concerning the authors

David Meredith is Director of Worker Software program Improvement at Associa. He oversees the efforts of the Associa group to create software program for his or her 15,000 workers to make use of day by day. He has virtually 20 years of expertise with software program within the residential property administration trade and lives within the Vancouver space of BC, Canada.

Josh Zacharias is a Software program Developer at Associa, the place he’s a lead engineer for the inner software program group. His work contains architecting full stack options for numerous departments within the firm in addition to empowering different builders to be extra environment friendly specialists in growing software program.

Monica Raj is a Deep Studying Architect on the AWS Generative AI Innovation Middle, the place she works with organizations throughout numerous industries to develop AI options. Her work focuses on constructing and deploying agentic AI options, pure language processing, contact middle automation, and clever doc processing. Monica has intensive expertise in constructing scalable AI options for enterprise clients.

Tryambak Gangopadhyay is a Senior Utilized Scientist on the AWS Generative AI Innovation Middle, the place he collaborates with organizations throughout a various spectrum of industries. His function includes researching and growing generative AI options to deal with essential enterprise challenges and speed up AI adoption. Previous to becoming a member of AWS, Tryambak accomplished his PhD at Iowa State College.

Nkechinyere Agu is an Utilized Scientist on the AWS Generative AI Innovation Middle, the place she works with organizations throughout numerous industries to develop AI options. Her work focuses on growing multimodal AI options, agentic AI options, and pure language processing. Previous to becoming a member of AWS, Nkechinyere accomplished her PhD at Rensselaer Polytechnic Institute, Troy NY.

 Naman Sharma is a Generative AI Strategist on the AWS Generative AI Innovation Middle, the place he collaborates with organizations to drive adoption of generative AI to resolve enterprise issues at scale. His work focuses on main clients from scoping, deploying, and scaling frontier options with the GenAIIC Technique and Utilized Science groups.

 Yingwei Yu is an Utilized Science Supervisor on the Generative AI Innovation Middle, primarily based in Houston, Texas. With intensive expertise in utilized machine studying and generative AI, Yingwei leads the event of revolutionary options throughout numerous industries.

 Dwaragha Sivalingam is a Senior Options Architect specializing in generative AI at AWS, serving as a trusted advisor to clients on cloud transformation and AI technique. With eight AWS certifications, together with ML Specialty, he has helped clients in lots of industries, together with insurance coverage, telecom, utilities, engineering, development, and actual property. A machine studying fanatic, he balances his skilled life with household time, having fun with street journeys, motion pictures, and drone images.