What challenges did United Airlines face before AI implementation?

Automate passport/document processing. Reduce costly manual annotation for extracting passenger identity fields and detecting potentially fraudulent documents. Improve model performance while minimizing labeling effort.

What was the impact of the AI solution?

Reduced manual labeling workload by labeling only images that maximize model improvement. Recurring cost reduction through active learning and elastic endpoint scaling. Reusable model-agnostic pipeline for other document-processing use cases.

Source

ExpandedProductionEvidence: Low40/100

United Airlines builds active learning document-processing pipeline with SageMaker Ground Truth and Amazon Textract

Use case typeIntelligent document processingUpdated Jun 13, 2026

United Airlines built an in-house passport information verification workflow to automate document processing and reduce costly manual annotation. The solution uses an active learning framework on AWS with Amazon Textract, Amazon SageMaker Ground Truth, AWS Step Functions, AWS CDK, and SageMaker endpoints.

Organization: United Airlines
Industry: Logistics
Location: United States
Published: September 2023

Reported outcomes

Strategic outcomes

New product / capabilityBuilt automated passport verification workflowCost efficiencyReduced manual labeling workloadNew product / capabilityCreated active learning labeling pipelineScale & capacityEnabled reusable document-processing pipeline

Why do we believe this deployment?Customer identity, provider attribution, maturity, and source checks

Customer: United Airlines
Provider: AWS
Maturity: Production
Linked source: AWS Machine Learning Blog

Labeled data is used to train a document understanding model deployed as a SageMaker endpoint

Customer identity supportedSource describes one deploymentMaturity supported

Primary read

Use case focus

Showing 3 of 3

1Document Processing Automation
2Intelligent Document Processing
3Active Learning

Automate passport/document processing.
Reduce costly manual annotation for extracting passenger identity fields and detecting potentially fraudulent documents.
Improve model performance while minimizing labeling effort.

Amazon Textract extracts text bounding boxes from passport images.
An uncertainty-sampling auto-labeling pipeline runs periodic inference and triggers human labeling jobs in Amazon SageMaker Ground Truth.
Labeled data is used to train a document understanding model deployed as a SageMaker endpoint.
The workflow is implemented end-to-end with AWS CDK and AWS Step Functions.

Technologies

Amazon Textract Amazon SageMaker Ground Truth AWS Step Functions AWS CDK Amazon SageMaker

Reduced manual labeling workload by labeling only images that maximize model improvement.
Recurring cost reduction through active learning and elastic endpoint scaling.
Reusable model-agnostic pipeline for other document-processing use cases.

Architecture

An active learning document-understanding pipeline on AWS: Amazon Textract extracts text bounding boxes from passport images; an auto-labeling workflow runs periodic inference, uncertainty sampling, and human labeling in Amazon SageMaker Ground Truth; labeled data feeds model training; the trained LayoutLM-based model is deployed as a SageMaker endpoint. The pipeline is orchestrated with AWS Step Functions and provisioned with AWS CDK, with elastic endpoint scaling for inference.

Sources & evidence1

Evidence: Low40/100Evidence strength