
Energy Management Firm Cuts Utility Bill Processing Cost to 6% of Manual Baseline
Protagona partnered with a national energy management firm to replace manual utility bill entry with a fully automated, serverless AI pipeline on AWS, processing tens of thousands of bills per month across hundreds of document formats.
Industry
Energy
Teams & Services
Generative AI Architecture, Data Engineering, DevOps
Tech & Tools
Amazon Bedrock, Amazon Textract, AWS Lambda, Amazon S3, Amazon SQS, Amazon SNS, Amazon API Gateway, Amazon DynamoDB, AWS Glue, Amazon Athena, AWS Lake Formation, Amazon QuickSight, Amazon CloudWatch, Anthropic Claude 3, Anthropic Claude 3.5 Sonnet, Python, Terraform, GitHub Actions
Key Data Points
The Vision
This energy management firm serves school districts, colleges, local governments, and healthcare organizations, having optimized more than 25,000 buildings and delivered cumulative customer savings exceeding $5 billion. As the client base expanded, so did the volume of utility bills flowing through operations. Leadership saw an opportunity to turn bill processing into a competitive advantage: faster client onboarding, more accurate data, and energy consultants freed to focus on analysis rather than manual entry.
The Goal
Protagona aimed to replace manual utility bill entry with a fully automated, serverless pipeline on AWS. Success meant processing approximately 58,000 bills per month across a wide range of formats, extracting roughly 30 data fields per bill with validated accuracy, and achieving a measurable reduction in both cost and processing time relative to the manual baseline.
The Challenge
The core difficulty was reliable extraction across a highly inconsistent document landscape. Utility bills arrive from hundreds of providers, each with its own layout, field labeling conventions, and formatting quirks. A pipeline that performs well on one format can fail silently on another, producing plausible but incorrect data that erodes trust over time. Handling this variety without a separate template for every provider required a fundamentally different approach than traditional OCR. The pipeline also had to meet production-grade reliability standards at volume. Failed documents needed to be captured, logged, and resubmissible without manual intervention. Accuracy had to be measurable against a defined threshold across a rolling window, and the system had to be observable enough that a small team could diagnose issues quickly.
The Solution
Protagona designed a fully serverless, event-driven pipeline that moves a utility bill from raw document to validated, queryable data without human involvement in the standard path. When a bill lands in an S3 bucket, a Lambda function preprocesses each page, rotating, converting, and resizing it to optimize extraction quality. Amazon Textract then pulls approximately 30 structured fields per bill, covering billing dates, usage, demand, cost, and rate codes. SQS queues coordinate a chain of Lambda stages for processing, deconsolidation, and AI-assisted validation. Amazon Bedrock, running Anthropic Claude models, applies format-specific prompts to locate and validate fields Textract alone cannot resolve. When Claude 3.5 Sonnet became available mid-project, the team adopted it: its broader format comprehension and Bedrock tool use capability handled a wider range of layouts at lower cost. A vision path covers bills that fall outside standard OCR reach entirely. Failed documents route to dead-letter queues and a dedicated failure bucket, with SNS alerts enabling resubmission. Validated results land in a data lake, with API Gateway exposing Lambda endpoints for downstream systems and reviewing staff. The full deployment runs through Terraform via GitHub Actions.

