GenAI-Powered Document Classification for Community Management

associa · Other · 2026 · source Agentic: no · promotional bias: med · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: community-management document classification

Associa needed to classify and organize tens of millions of community-management documents across many document types because manual categorization was slow, error-prone, and created retrieval and operational bottlenecks.

Constraints:

• Associa manages approximately 48 million documents across 26 TB of data
• the existing document management system lacked efficient automated classification
• classification had to work across uneven document classes, including an Unknown class for drafts and email correspondences
• accuracy had to be balanced against per-document inference and OCR cost
• Unknown classification mattered for downstream human review and operational efficiency

Architecture

Associa and the AWS Generative AI Innovation Center built a GenAI-powered document-classification pipeline using the GenAI IDP Accelerator on AWS. The solution uses Pattern 2: Amazon Textract OCR and Amazon Bedrock classification. The team evaluated full-PDF versus first-page input, OCR+image versus image-only prompting, and multiple Bedrock models, selecting first-page-only OCR+image classification with Amazon Nova Pro because it delivered 95% overall accuracy and 85% Unknown accuracy at 0.55 cents per document.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  document-management-system["Document management system"] --> genai-idp-accelerator["GenAI IDP Accelerator"]
  genai-idp-accelerator --> aws-cloudformation["AWS CloudFormation"]
  genai-idp-accelerator --> job-management-and-monitoring["Job management and monitoring"]
  genai-idp-accelerator --> amazon-textract["Amazon Textract"]
  amazon-textract --> amazon-bedrock["Amazon Bedrock"]
  amazon-bedrock --> amazon-nova-pro["Amazon Nova Pro"]
  amazon-bedrock --> model-evaluation["Model evaluation"]
  amazon-nova-pro --> document-management-system
  document-management-system --> human-review["Human review"]

Components

Data flow

Incoming documents enter the GenAI IDP Accelerator, which uses shared AWS infrastructure for job management, progress tracking, and monitoring. Pattern 2 extracts OCR/layout information with Amazon Textract, combines it with the first-page image in a multimodal prompt, and classifies the document with Amazon Nova Pro on Amazon Bedrock. Evaluation compares outputs against 465 labeled PDFs across eight document types plus Unknown, measuring overall accuracy, Unknown accuracy, and per-document classification cost.

Why this architecture

The team used the GenAI IDP Accelerator because it provided modular cloud document-processing infrastructure with OCR, classification, job tracking, and monitoring. They selected first-page OCR+image with Nova Pro because it improved accuracy and cost versus full PDFs, preserved Unknown-document performance versus image-only prompts, and balanced cost against higher-priced models.

Models used: GenAI IDP Accelerator · Amazon Bedrock · Amazon Textract · Amazon Nova Pro · Amazon Nova Lite · Amazon Nova Premier · Anthropic Claude Sonnet 4 · AWS CloudFormation

Scale

• Associa oversees approximately 7.5 million homeowners.
  “oversees approximately 7.5 million homeowners with 15,000 employees across more than 300 branch offices.”article, stated
• Associa has 15,000 employees across more than 300 branch offices.
  “15,000 employees across more than 300 branch offices.”article, stated
• The company manages approximately 48 million documents across 26 TB of data.
  “The company manages approximately 48 million documents across 26 TB of data”article, stated
• The evaluation dataset had 465 PDFs across eight distinct document types.
  “The evaluation dataset consists of 465 PDF documents across eight distinct document types.”article, stated
• Document classes were unbalanced, from 6 Policies and Resolutions samples to 155 Minutes samples.
  “ranging from 6 samples for Policies and Resolutions to 155 samples for Minutes.”article, stated
• The selected solution achieved 95% accuracy at 0.55 cents per document.
  “achieves 95% accuracy at an average cost of 0.55 cents per document.”article, stated

Alternatives considered (roads not taken)

Full PDF prompt input

Rejected because: Using all pages reached 91% accuracy at 1.10 cents per document and performed poorly on Unknown documents.

First-page-only input

Rejected because: This was not rejected; it was selected because it improved overall accuracy, Unknown accuracy, speed, and cost.

Image-only multimodal prompt without OCR

Rejected because: It reduced OCR cost but lowered overall accuracy and sharply reduced Unknown-document accuracy needed for human review.

Anthropic Claude Sonnet 4

Rejected because: It had the highest Unknown accuracy but more than doubled the per-document cost compared with Nova Pro.

Amazon Nova Premier

Rejected because: It had the best overall accuracy but cost about twice as much per document as Nova Pro.

Patterns

eval-driven-development (model selection)

Associa and AWS systematically evaluated prompt input, prompt design, model choice, accuracy, Unknown accuracy, and cost before selecting the production configuration.

cost-aware-routing (model operations)

Model and prompt choices were judged by accuracy and cents-per-document, leading to Nova Pro instead of more expensive Claude Sonnet 4 or Nova Premier options.

latency-optimization (document processing)

The first-page-only approach was selected because first pages contained distinctive features while later pages added noise, and the approach improved processing speed and infrastructure cost.

structured-output-contract (classification)

The model classifies each PDF into a fixed set of document types plus Unknown, making outputs operationally usable in document-management workflows.

human-in-the-loop-gate (review routing)

Unknown classification is preserved because downstream human review depends on accurately identifying documents that lack sufficient classification criteria.

Tradeoffs

Prompt input scope

Chose first page only over full PDF document.

Rationale: First-page classification improved accuracy from 91% to 95%, lowered cost from 1.10c to 0.55c, and improved Unknown accuracy from 40% to 85%.

Prompt modality

Chose OCR plus image over image-only multimodal prompting.

Rationale: Image-only was cheaper but reduced overall accuracy and especially Unknown accuracy, which was critical for downstream human review.

Model choice

Chose Amazon Nova Pro over Nova Lite, Nova Premier, and Claude Sonnet 4.

Rationale: Nova Pro balanced 95% overall accuracy and 85% Unknown accuracy at 0.55c per document, while alternatives either lowered Unknown accuracy or increased cost.

Gotchas (what bit them)

More document pages can reduce classification quality.

Root cause: Later pages in drafts or email threads can introduce noise that confuses the classifier.

Resolution: Use only the first page for classification because it contains the most distinctive document features.

Removing OCR looked cheaper but harmed Unknown classification.

Root cause: Image-only prompting relies entirely on model visual understanding and missed Unknown document distinctions needed for review routing.

Resolution: Keep OCR plus image in the multimodal prompt despite higher per-document cost.

Overall accuracy can hide class-specific failure.

Root cause: Image-only and full-PDF configurations had acceptable-looking overall accuracy while Unknown accuracy was much lower.

Resolution: Evaluate both overall classification and Unknown-document classification when selecting the configuration.

Class imbalance complicates evaluation.

Root cause: The dataset had document classes ranging from only 6 samples to 155 samples, making per-class performance important.

Resolution: Report class-level sample counts, correct classifications, accuracy, and cost rather than only aggregate accuracy.

Evaluation

Associa evaluated 465 PDFs across eight document types plus Unknown, comparing full PDF versus first-page input, OCR+image versus image-only prompting, and Nova/Claude model options on overall accuracy, Unknown accuracy, and classification cost.

Metrics: overall classification accuracy · Unknown document classification accuracy · per-document classification cost · samples classified correctly · document type sample count · first-page versus full-PDF accuracy · OCR+image versus image-only accuracy

Outcomes

• The selected Amazon Nova Pro on Bedrock configuration achieved 95% accuracy.
• Average classification cost was 0.55 cents per document.
• First-page-only classification improved Unknown accuracy from 40% to 85%.
• Image-only classification was rejected after Unknown accuracy dropped to 50%.
• Associa expects substantial cost savings and reduced manual document-classification effort.

Associa operates at very large document-management scale: 7.5 million homeowners, 15,000 employees, more than 300 branches, approximately 48 million documents, and 26 TB of data. The evaluated configuration uses 465 labeled PDFs and achieves 95% accuracy at 0.55 cents per document.