Box Extraction Agents

AI agent · Data & analytics agent · Data & analytics by Box

Automate structured data extraction from unstructured enterprise content.

Box Extraction Agents automate the retrieval of structured data from unstructured documents like contracts and invoices. They enable businesses to scale document processing with high accuracy using frontier AI models and built-in OCR.

Agent specs

The full spec card (context, pricing per token, modalities and more) on AntSpace. Sign up free to see the full spec card.

Languages
English, Japanese, Chinese, Korean

Category: Workflow Automation

Channels
API; Box CLI; Box AI Studio
Integrations
OpenAI; Anthropic; Google; AWS; Salesforce; Box Sign
Data sources
Box Content Cloud; PDF; TIFF; PNG; JPEG; Text-based files
Models
OpenAI, Anthropic, and Google models
Maximum files per request
25 files
Prompt length limit
10,000 characters
  • Extracts structured data from contracts and bank statements
  • Supports OCR for PDF, TIFF, PNG, and JPEG formats
  • Processes up to 25 files in a single batch request
  • Multilingual support for English, Japanese, Chinese, and Korean
  • Customizable via Box AI Studio and API overrides

Overview

Box Extraction Agents are specialized AI tools designed for businesses to identify and retrieve structured data from unstructured enterprise content at scale. They solve the problem of manual data entry by automatically extracting critical information from complex documents like lease agreements, bank statements, and handwritten bills of lading. The agents operate within the Box Intelligent Content Management platform, utilizing models from OpenAI, Anthropic, and Google to deliver high-accuracy metadata extraction. Developers can customize these agents through Box AI Studio or the Box AI API, applying specific extraction logic and model overrides to fit unique business workflows. By automating document intake, these agents reduce errors and accelerate end-to-end processes such as loan processing and vendor onboarding.

Document Intelligence Tasks

  • Structured Data Retrieval: The agents identify and retrieve structured data from unstructured content such as detailed contracts, bank statements, and handwritten bills of lading.
  • Customized Extraction Logic: Users can automate document processing with customized extraction agents to achieve higher accuracy and confidence in results.
  • Batch Processing: The agents can process up to 25 files in a single request, treating them as one combined input or running them in parallel for per-file results.
  • Automatic OCR Integration: The system automatically applies OCR to supported formats including PDF, TIFF, PNG, and JPEG to extract text from images.

Human and Developer Control

  • Model Overrides: Developers can override default model configurations using the ai_agent parameter to pin specific model versions or switch between providers.
  • Customizable Prompts: The prompt_template and system_message parameters allow users to steer LLM behavior without changing application code.
  • Creativity Tuning: Users can adjust temperature and other endpoint parameters to control how deterministic or creative the agent's output is.
  • Context Management Control: The system allows for the configuration of embeddings models and chunking strategies for processing long documents.

What it does

  • Identify and retrieve structured data from content
  • Automate document intake workflows
  • Extract metadata from files
  • Process handwritten bills of lading
  • Perform OCR on images and PDFs

Use cases

  • Insurance firms automate the extraction of claim details from handwritten forms and policy documents to speed up processing.
  • Legal teams retrieve key dates and financial terms from large volumes of lease agreements for database entry.
  • Finance departments automatically pull vendor names and totals from PDF invoices into searchable metadata fields.
  • Logistics companies process handwritten bills of lading to digitize shipping records without manual data entry.

Specifications

Channels
API; Box CLI; Box AI Studio
Integrations
OpenAI; Anthropic; Google; AWS; Salesforce; Box Sign
Data sources
Box Content Cloud; PDF; TIFF; PNG; JPEG; Text-based files
Human control
Developers control model behavior via ai_agent overrides including prompt templates, system messages, and temperature settings.
Models
OpenAI, Anthropic, and Google models
Languages
English, Japanese, Chinese, Korean, Russian, Ukrainian, Bulgarian, Serbian, French, Spanish
Maximum files per request
25 files
Prompt length limit
10,000 characters
Single file text limit
2 MB
Maximum images per request
5 images
Maximum image resolution
1024 x 1024 pixels
OCR Support
Automatic for PDF, TIFF, PNG, JPEG
Model Providers
OpenAI, Google, AWS

How it works

  1. File Retrieval: The agent reads file content from the provided items array or optional content parameter.
  2. Representation Generation: Box converts documents into text representations and applies OCR automatically for image-based files.
  3. Model Routing and Generation: The request is routed to the default or overridden LLM to process the prompt against the content.

FAQ

What file formats are supported for structured extraction?

Box Extraction Agents support PDF, TIFF, PNG, and JPEG formats for structured extraction with automatic OCR.

How many files can I process at once?

You can process up to 25 files in a single request for extraction tasks.

How can I control the agent's behavior?

Developers can use the ai_agent parameter to override default models, customize prompt templates, and adjust LLM parameters like temperature.

Added by AntSpace from public sources (box.com).

Related

More AI agents by Box