What challenges did Travelers Insurance face before AI implementation?

Travelers needed to classify large volumes of service-request emails that often contained ambiguous language and PDF attachments. Manual processing consumed significant time and was difficult to scale for routine policy-service requests.

What was the impact of the AI solution?

The classifier reached 91% accuracy, up from 68% before prompt engineering. The automation can save tens of thousands of manual processing hours and redirect staff to more complex work.

Source

Travelers Insurance: Amazon Bedrock + Amazon Textract email classification automation

Use case typeIntelligent document processingUpdated Jun 13, 2026

Travelers Insurance receives millions of service-request emails each year, many with ambiguous content and PDF attachments. The company and AWS Generative AI Innovation Center built an AI-based email classification workflow to automate routing of these requests into 13 service request categories.

Organization: Travelers Insurance
Industry: Insurance
Location: United States
Published: January 2025

Reported outcomes

91%

accuracyQuality & accuracy

68%accuracy

Strategic outcomes

New product / capabilityAutomated email routing into service categoriesSpeed & agilityAccelerated routine request processingScale & capacityEnabled handling of high email volumesEmployee experienceFreed staff for complex work

Catalog median for quality & accuracy deployments: +90% across 282 reported metrics. Compare benchmarks →

Primary read

Use case focus

Showing 3 of 3

1Email classification
2Document processing automation
3Workflow automation

Travelers needed to classify large volumes of service-request emails that often contained ambiguous language and PDF attachments.
Manual processing consumed significant time and was difficult to scale for routine policy-service requests.

The solution used Anthropic Claude models on Amazon Bedrock as a foundation-model classifier with prompt engineering and few-shot examples.
Email body text was extracted, PDF attachments were split into pages, and Amazon Textract was used to extract text, entities, and table data from the page images.
The email text and Textract output were combined into a single prompt to classify each message into one of 13 categories.
The team iterated on prompts, condensed categories, and improved instructions to raise accuracy and produce explainable outputs.

Technologies

Amazon Bedrock Amazon Textract Anthropic Claude 3

The classifier reached 91% accuracy, up from 68% before prompt engineering.
The automation can save tens of thousands of manual processing hours and redirect staff to more complex work.

Architecture

Raw emails were ingested, body text extracted, and any PDF attachments were rendered into page images. Amazon Textract processed the page images to extract text, entities, and table data. The extracted attachment text was combined with the email body text and sent to Anthropic Claude on Amazon Bedrock for classification into 13 service-request categories.

Implementation partners2

AWS GenAI Innovation Center Anthropic

Sources & evidence1

Groundedness: 5/5

Primary source

AI-generated summary. Verify important details with the linked sources before relying on this case.

Explore related AI use cases

Customer evidence