Recording and Summary: Trusting AI Extraction - Confidence Scores, Governance and Human in the Loop Design

Thanks ​@jrobbins ​@Scott Picanco Box for yesterday’s overview and demo of metadata extract and confidence scores.  Please see below for the recording, summary and deck. 

Slide deck here

1. The Value of Structured Data in an Unstructured World
​@Scott Picanco Box began by highlighting that while content is the lifeblood of modern organizations, 90% of this data is unstructured, making its value difficult to tap into. He explains that metadata (structured data about data) is the missing link that makes AI truly useful. Applying metadata effectively leads to better decision-making, powers smarter agentic workflows with tools like Box Automate, and dramatically accelerates content search and discovery.

2. Challenges with Traditional Data Extraction
​@Scott Picanco Box  outlined the problems with current data extraction methods. Manual processes are slow, error-prone, and lack visibility, while legacy Intelligent Document Processing (IDP) and OCR tools are often expensive, inefficient, create data silos by separating data from content, and struggle with variable document types and formats.

3. Introducing Box Extract: An Agentic AI Solution
Box Extract is a platform-native solution that combines the latest AI models, advanced OCR, and agentic reasoning to accurately extract structured data from complex content. ​@Scott Picanco Box  detailed the governed, multi-step pipeline that every document goes through:

• Digitize: Advanced OCR makes content machine-readable, including handwritten text.
• Understand: The system identifies the document type and selects the appropriate extraction logic.
• Plan: Agentic reasoning is used to understand the document's meaning and context to plan and refine the extraction.
• Validate: Confidence scores and human-in-the-loop processes are strategically employed to ensure high accuracy at scale.
• Deliver: The extracted, trusted data is delivered as metadata directly alongside the content in Box.

4. Trusting AI: The Role of Guardrails and Governance
​@jrobbins discussed the necessity of building trust in AI outputs. He emphasized that guardrails are crucial to mitigate legal, compliance, and financial risks associated with bad data. He identified three common types of guardrails:

• AI Confidence: Flagging outputs where the AI indicates ambiguity.
• Formatting & Validation Rules: Checking data against strict formats or external sources like an approved vendor list.
• Deterministic Rules: Automatically routing documents for review based on their content, such as an invoice total exceeding a specific dollar amount.

5. Live Demonstration: Operationalizing a Human-in-the-Loop Workflow
​@jrobbins  provided a live demo showing how to build a complete, trustworthy extraction process for invoices using Box tools:

• Extract Agent Setup: He showcases a new feature that uses AI to automatically generate a metadata template and extraction prompts from a few sample documents.
• Box Automate Workflow: He builds a workflow that triggers on file upload, runs the extraction, and then applies a series of guardrails: a custom AI agent to check for low confidence, an HTTP request to an external system to validate the vendor, and a conditional step to check the invoice amount.
• Human Review Process: Based on the guardrail outcomes, the workflow routes the invoice for straight-through processing or to a dedicated 'review' folder, creating a task for a human reviewer.
• Reviewer App: He demonstrates a custom Box App that serves as a dashboard for reviewers to see flagged documents, understand why they were flagged, make corrections, and approve or reject them, which then continues the automated workflow.

Key Insights & Upcoming Features
The team discussed the roadmap for Box Extract, highlighting upcoming features designed to enhance trust and usability, including AI-powered metadata template generation, bounding boxes to visualize the source of extracted data, field-level confidence thresholds, and advanced accuracy metrics for prompt optimization. They also confirm that highly requested features like sub-folder (cascading) extraction and support for metadata taxonomies are in active development and coming soon.