Back office ops · Production

Travelers Insurance classifies service request emails with Amazon Bedrock and Anthropic Claude, achieving 91% accuracy

The problem

Travelers receives millions of emails a year requiring classification of complex, sometimes ambiguous service requests into categories such as address changes, coverage adjustments, and payroll updates — a volume of manual processing that warranted automation.

Workflow diagram · grounded in source
1
Email ingestion
trigger
“The raw email is ingested into the pipeline. The body text is extracted from the email text files.”
2
PDF attachment parsing
integration
“If the email has a PDF attachment, the PDF is parsed. The PDF is split into individual pages. Each page is saved as an image.”
3
OCR text extraction via Textract
ai_action
“The PDF page images are processed by Amazon Textract to extract text, specific entities, and table data using Optical Character Recognition (OCR).”
4
Text cleaning and combination
integration
“The text is then cleaned of HTML tags, if necessary. The text from the email body and PDF attachment are combined into a single prompt for the large language model (LLM).”
5
Claude email classification
ai_action
“Anthropic's Claude classifies this content into one of 13 defined categories and then returns that class”
6
Prediction output and analysis
output
“The predictions for each email are further used for analysis of performance.”
Reported outcome

Through prompt engineering with Anthropic's Claude models on Amazon Bedrock, the team achieved 91% classification accuracy — up from 68% without prompt engineering — with the system positioned to save tens of thousands of hours of manual processing.

Reported metrics
Classification accuracy with prompt engineering91%
Initial classification accuracy without prompt engineering68%
Claude Instant classification accuracy90%
Manual processing hours savedtens of thousands of hours
Reported stack
Amazon BedrockAmazon TextractAnthropic's Claude
Source
https://aws.amazon.com/blogs/machine-learning/how-travelers-insurance-classified-emails-with-amazon-bedrock-and-prompt-engineering?tag=soumet-20
Read source ↗

Frequently asked questions

What did this team achieve with this AI workflow?

Through prompt engineering with Anthropic's Claude models on Amazon Bedrock, the team achieved 91% classification accuracy — up from 68% without prompt engineering — with the system positioned to save tens of thousand…

What tools did this team use?

Amazon Bedrock, Amazon Textract, Anthropic's Claude.

What results were reported?

Classification accuracy with prompt engineering: 91%; Initial classification accuracy without prompt engineering: 68%; Claude Instant classification accuracy: 90%; Manual processing hours saved: tens of thousands of hours (source-reported, not independently verified).

How is this back office ops AI workflow structured?

Email ingestion → PDF attachment parsing → OCR text extraction via Textract → Text cleaning and combination → Claude email classification → Prediction output and analysis.

Related back office ops cases