AWS: Fine-tuning Amazon Nova models raised email data extraction accuracy to 94.77%

The AWS Machine Learning Blog has published a case study on fine-tuning Amazon Nova models with Amazon SageMaker AI to extract data from corporate email. Accuracy rose to 94.77%, and processing costs dropped by 50%.

What is Amazon Nova

Amazon Nova is AWS's family of multimodal foundation models, introduced at re:Invent in late 2024. The lineup includes text models in several sizes (Micro, Lite, Pro, and Premier), plus Canvas and Reel for image and video generation. All of the models are available through Amazon Bedrock. AWS pitches them as a cheaper, faster alternative to third-party models, with comparable quality on routine enterprise tasks such as text classification, summarization, and data extraction.

Why an off-the-shelf model gets emails wrong

Without additional training, a general-purpose model often confuses fields with similar meanings in incoming mail: a shipping address and a billing address, for example, or an order number and a support ticket number. Every company phrases requests and formats emails its own way, and a generic model doesn't know these quirks. Fine-tuning through Amazon SageMaker AI targets exactly this problem. The model trains on real emails from a specific organization and learns its particular data patterns: number formats, field placement, and industry-specific terminology.

• Email data extraction accuracy reached 94.77%
• Processing cost per email dropped by 50%
• Fine-tuning was done with Amazon SageMaker AI
• The model learned to distinguish structurally similar fields and recognize a specific company's patterns

Who benefits from accuracy above 94%

Extracting data from incoming mail is routine work for support teams, logistics companies, and insurance and finance departments that manually sort orders, invoices, requests, and claims every day. At around 95% accuracy, most of these emails can be handed to the model for fully automated processing, leaving people to handle only complex or disputed cases. Halving the processing cost makes automation worthwhile even at high email volumes. The same approach applies beyond email to any semi-structured document, including waybills, questionnaires, and feedback forms.

How fine-tuning works in SageMaker AI

Amazon SageMaker AI is AWS's managed service for the full machine learning lifecycle, from data preparation and training to deployment and production monitoring. To fine-tune a language model like Nova, a company needs to assemble a set of labeled examples (real emails with the fields already tagged by hand) and pass it to SageMaker AI, which then tunes the model's weights for that specific task. Unlike training a model from scratch, fine-tuning needs far less data and compute, because the model already understands language. It only needs to be adjusted to the specifics of one narrow task.

What it means

The AWS case study shows the practical value of fine-tuning for narrow enterprise tasks. Nova, a relatively small model by industry standards, can compete on accuracy with more expensive general-purpose models after fine-tuning on a company's own data, while costing noticeably less to run. For businesses, this means the path to cheaper automation of routine work runs not only through buying a more powerful model, but through fine-tuning an already available one on their own data.

The AWS case study shows the practical value of fine-tuning for narrow enterprise tasks. Nova, a relatively small model by industry standards, can compete on accuracy with more expensive general-purpose models after fine-tuning on a company's own data, while costing noticeably less to run. For businesses, this means the path to cheaper automation of routine work runs not only through buying a more powerful model, but through fine-tuning an already available one on their own data.