AWS: Fine-Tuning Amazon Nova Models Increased Email Data Extraction Accuracy to 94.77%

AWS Machine Learning Blog published a case study on fine-tuning Amazon Nova models using Amazon SageMaker AI for extracting data from corporate email: accuracy increased to 94.77%, and processing costs decreased by 50%.

What is Amazon Nova

Amazon Nova is a family of foundational multimodal models from AWS, presented at the re:Invent conference in late 2024. The lineup includes text models of various sizes — Micro, Lite, Pro, and Premier — as well as models for image and video generation called Canvas and Reel. All models are available through Amazon Bedrock service and are positioned by AWS as a cheaper and faster alternative to models from third-party vendors while maintaining comparable quality on typical corporate tasks such as text classification, summarization, and data extraction.

Why ready-made models make mistakes with emails

Without additional training, a general-purpose model often confuses similar fields in incoming email — for example, shipping address and billing address, order number and support ticket number. Each company formulates requests and formats emails differently, and a standard model does not know these particulars. Fine-tuning through Amazon SageMaker AI solves this exact problem: the model learns from real examples of emails from a specific organization and memorizes its own data templates — number formats, field placement, specific industry terminology.

• Data extraction accuracy from emails reached 94.77%
• Processing costs per email decreased by 50%
• Fine-tuning was performed through Amazon SageMaker AI service
• The model was trained to distinguish between structurally similar fields and recognize patterns specific to a company

Who will benefit from over 94% accuracy

Extracting data from incoming email is a routine task for support departments, logistics companies, insurance and financial divisions that manually process orders, invoices, requests, and claims daily. With accuracy around 95%, most of these emails can be passed to the model completely automatically, leaving humans only complex or disputed cases, and the halved processing cost makes automation justified even with large volumes of incoming correspondence. Similar approach applies not only to email but to any semi-structured documents — invoices, questionnaires, feedback forms.

How the fine-tuning process works in SageMaker AI

Amazon SageMaker AI is a managed AWS service for the full lifecycle of machine learning models: from data preparation and training launch to deployment and production monitoring. To fine-tune a language model like Nova, a company needs to collect a set of labeled examples — real emails with manually marked fields — and pass it to SageMaker AI, which launches the process of fine-tuning model weights for this specific task. Unlike training a model from scratch, fine-tuning requires significantly less data and computing resources because the model already understands language — it only needs to be adjusted for the specifics of one narrow task.

What this means

AWS's case study shows the practical value of fine-tuning for narrow corporate tasks: a relatively small by industry standards Nova model, after fine-tuning on a company's own data, can compete in accuracy with more expensive universal models while being significantly cheaper to operate. For business, this means that the path to savings in automating routine work lies not only in purchasing a more powerful model but in fine-tuning an already available one on its own data.

AWS's case study shows the practical value of fine-tuning for narrow corporate tasks: a relatively small by industry standards Nova model, after fine-tuning on a company's own data, can compete in accuracy with more expensive universal models while being significantly cheaper to operate. For business, this means that the path to savings in automating routine work lies not only in purchasing a more powerful model but in fine-tuning an already available one on its own data.