How Amazon’s New Bedrock and SageMaker Features Cut AI Agent Costs and Speed Up Customization

Efficiency is a core challenge when deploying AI agents at scale; large‑model inference is costly and resource‑intensive, yet agents spend most of their time on routine tasks such as schedule lookup and document search that do not require the most powerful models.

Amazon addresses this gap with two new customization capabilities announced at re:Invent 2025: the reinforced fine‑tuning feature in Amazon Bedrock and a serverless model‑customization capability in Amazon SageMaker AI. Both aim to let developers replace heavyweight foundation models with smaller, task‑specific models, reducing cost, improving response speed, and increasing accuracy.

Amazon Bedrock reinforced fine‑tuning

Bedrock’s reinforced fine‑tuning (RFT) automates the entire fine‑tuning pipeline, requiring no PhD‑level expertise. Developers select a base model, point it to logs or upload a dataset, choose a reward function (AI‑based, rule‑based, or a preset template), and the service runs the fine‑tuning end‑to‑end.

According to benchmark tests, fine‑tuned models achieve an average 66% accuracy improvement over the original base model, and in a Salesforce case study the accuracy gain reached 73%.

When the feature launched, it supported the Amazon Nova 2 Lite model, with plans to add more models later.

Amazon SageMaker AI serverless model‑customization

SageMaker AI now compresses the model‑customization cycle from months to days by offering two modes:

Agent‑guided mode: developers describe requirements in natural language; an AI agent orchestrates data synthesis, training, and evaluation.

Autonomous mode: developers retain full control, managing parameters and techniques without handling infrastructure.

The service supports a suite of advanced customization techniques, including Reinforcement Learning from AI Feedback (RLAIF), Reinforcement Learning with Verifiable Rewards, Supervised Fine‑Tuning, and Direct Preference Optimization. It also works with open‑weight models such as Llama, Qwen, DeepSeek, and GPT‑OSS.

Customer examples illustrate the impact: Salesforce agents saw both accuracy and efficiency rise sharply after applying Bedrock RFT; Collinear AI reported weeks saved in their model‑optimization pipeline by using SageMaker AI’s serverless workflow.

Quotes and resources

"The reinforced fine‑tuning feature lets us achieve up to a 73% accuracy boost without the overhead of massive models," says Phil Mui, Senior Vice President of Software Engineering at Salesforce Agentforce.

"With SageMaker AI’s serverless customization, we shortened our experiment cycle from weeks to days, allowing us to focus on high‑quality training data rather than infrastructure," adds Soumyadeep Bakshi, Co‑Founder of Collinear AI.

For more details, see the Amazon Bedrock customization page (https://aws.amazon.com/bedrock/customize/) and the Amazon SageMaker AI model‑customization page (https://aws.amazon.com/sagemaker/ai/model-customization).