Samsara reduces Bill of Lading review effort by over 60% with Claude Sonnet on Amazon Bedrock

Project Overview

Samsara's connected operations platform helps organizations run safer, more efficient physical operations. Bills of Lading (BoLs) are a critical shipment control document, but they're still largely paper: BoL images are captured during field workflows, and operations teams key the data in by hand from documents that mix print and handwriting across dozens of carrier formats, slowing billing and limiting real-time visibility.

Samsara partnered with AWS and Minfy Technologies, an AWS Premier Tier Consulting Partner, to modernize this workflow with generative AI. Minfy designed and deployed a document intelligence pipeline pairing Amazon Textract with Anthropic Claude Sonnet 4.6 on Amazon Bedrock, turning driver-captured BoL images into structured, reviewable shipment data within seconds.

‍

‍

The Challenge

BoL automation isn't a standard OCR problem — the documents are operationally critical, visually inconsistent, and often captured in the field under imperfect conditions. Key priorities included:

• Cutting the manual effort of reviewing BoL images and entering shipment data.
• Reliably extracting fields from printed, handwritten, and poor-quality documents.
• Returning results fast enough for mobile drivers, with more detailed review available in the web application.
• Managing extraction risk through confidence indicators and human

‍

The Solution

Samsara uses Claude Sonnet 4.6 on Amazon Bedrock, alongside Amazon Textract, as the intelligence layer of a two-phase BoL pipeline. Amazon Textract reads document structure and handwriting; Claude interprets that output in context to identify shipment attributes, locate them on the page, and judge whether each extraction is trustworthy — returning results within seconds on mobile, then enriching web results in the background with field highlights and confidence scores.

Operations teams stay in control throughout: an administrative validation interface shows the original document alongside extracted fields and confidence scores, so reviewers can approve or correct values before data is finalized.

Implementation details

• Event-driven architecture on AWS: driver-captured images upload via a secure API into Amazon S3, tagged with shipment and client metadata.
• Minfy engineers worked directly with Samsara's product and engineering teams to design the two-phase pipeline and specialize Claude's prompts for BoL fields. ‍
• Phase 1: Textract output is passed to Claude Sonnet 4.6 with specialized field prompts, returning extracted values and coordinates immediately. Phase 2 (web): a second Claude pass fuses coordinates with Textract's bounding boxes and scores each field 0.0–1.0 with a justification. ‍
• Product: Anthropic Claude via Amazon Bedrock. Model: Claude Sonnet 4.6 — API model string anthropic claude-sonnet-4-6.

‍

How Claude supports Samsara's BoL workflow

• Contextual field extraction — identifies BoL-specific attributes even as labels and handwriting vary across carriers. ‍
• Layout-aware interpretation — reasons over full document context, not isolated text snippets. ‍
• Field localization — returns coordinates so reviewers can verify a value against its source on the page. ‍
• Confidence scoring with justification — focuses review time on low-confidence fields instead of every document. ‍
• Extensible pattern — uses reusable prompts grounded in BoL-specific semantics, so it can extend to other logistics documents without a new rules engine.

‍

Business benefits

• 60%+ less manual effort: automating interpretation cut BoL review and data-entry workload substantially. ‍
• Up to 89% field-level accuracy: 89% on printed documents, 76–78% across handwritten and poor-quality captures. ‍
• Near real-time data: structured fields available within seconds of capture. ‍
• Scalable, reusable pattern: built to grow with shipment volume and extend to other logistics documents.