CarbonTrail cuts generative AI costs by 88% for sustainable emissions intelligence using Amazon Bedrock

Overview

New Zealand-based CarbonTrail built an AI-powered emissions measurement platform and CarbonAPI on AWS, running a document-analysis pipeline on Amazon Bedrock and workloads on AWS Inferentia to process bank-scale invoice and transactional data. The platform achieves an 87% reduction in processing time and 88% lower cost than a comparable GPT-4-with-embeddings approach, and reduces low-confidence classifications by up to 40%. The Bank of New Zealand (BNZ) uses CarbonTrail's CarbonAPI to progress toward measuring emissions across its SME customers.

The challenge

CarbonTrail's mission is to help businesses measure and reduce emissions, with a goal of avoiding 1 billion tons of CO2 by 2050, but traditional methods — manual processes and broad spend-based estimates — were too slow, inaccurate and resource-heavy to scale. Banks such as the Bank of New Zealand (BNZ), which set a target to measure the emissions of 50% of its SME customers, are further challenged by regulatory requirements that demand processing hundreds of thousands of customer records across multiple systems, a scale manual approaches could never achieve.

The solution

CarbonTrail built an AI-powered emissions measurement platform on AWS that runs a document-analysis pipeline on Amazon Bedrock to process unstructured invoice and receipt data, integrating generative AI models such as Llama through Amazon Bedrock foundation models. To increase efficiency, CarbonTrail runs workloads on AWS Inferentia for higher throughput with lower emissions per token, and uses AWS Fargate and AWS Lambda for containerized, serverless, elastic operations. The platform is hosted in-region, supporting Amazon VPC peering for banking clients to meet data sovereignty requirements. CarbonTrail also introduced CarbonAPI, a developer-facing service that extends the same invoice-level emissions measurement to partners and fintechs.

Reported business value

CarbonTrail's platform delivers an 87% reduction in processing time and 88% lower cost than a comparable GPT-4-with-embeddings approach, and reduces low-confidence classifications by up to 40%. The Bank of New Zealand (BNZ) uses CarbonTrail's CarbonAPI to progress toward its goal of measuring emissions across its SME customers.