Top 5 AI trends in March 2025: AI agents, smarter supply chains, and more 

March 2025 has brought a fresh wave of AI innovations that go beyond theoretical breakthroughs and translate directly into real-world impact. From OpenAI’s newly released tools for building AI agents to Microsoft’s external knowledge integration approach, this month’s highlights underscore how artificial intelligence is evolving in ways that matter today, not just tomorrow. Whether it’s optimizing supply chains, pushing the limits of robotics, or making AI systems smarter and more trustworthy, these developments are shaping the landscape of technology and business alike. 

Below, we delve into five top AI trends from March 2025 and explore the practical ways they’re changing industries—from retail to robotics—and setting the stage for what’s ahead. 

1. OpenAI Releases New Tools for Building AI Agents

OpenAI released a suite of tools aimed at helping developers and enterprises build useful and reliable AI agents. These tools include the Responses API and an open-source Agents SDK, designed to simplify the creation and management of AI agents capable of performing complex, multi-step tasks. 

Why OpenAI’s new tools for building AI agents matter: 

By providing these building blocks, OpenAI enables businesses to develop customized AI solutions tailored to their specific needs. This democratizes access to advanced AI capabilities, fostering innovation and accelerating the integration of AI into various sectors. The tools also emphasize safety and reliability, addressing critical concerns in AI deployment.

2. Microsoft’s KBLaM: External Knowledge Integration for LLMs

Microsoft has introduced KBLaM (Knowledge Base-augmented Language Model), a new approach that encodes and stores structured external knowledge within an LLM itself without the need for retraining. This approach allows for dynamic updates of knowledge without requiring model fine-tuning or managing multiple RAG-based retrieval tools. By integrating external knowledge directly into LLMs, KBLaM enhances the model’s ability to retrieve and verify information, improving accuracy and reducing the occurrence of hallucinations. 

Why KLBLaM matters: 

KBLaM represents a major step forward in integrating structured knowledge into LLMs. Results from the research team show that KBLaM significantly reduces response time and memory usage compared to RAG. It also allows more information to be stored in context compared to traditional methods for in-context learning. This approach has significant potential for applications that rely on dynamic and consistently updated information. Microsoft Research is still working on scaling KBLaM and testing its performance on more complicated tasks, however, the initial results demonstrate a promising alternative to traditional RAG and fine-tuning approaches which are often complex and costly to implement. 

3. Neudesic’s AI-Powered Supply Chain Control Tower

Neudesic has introduced an AI-powered Supply Chain Control Tower to help retailers overcome operational challenges, improve efficiency, and drive cost savings. Neudesic’s supply chain control tower integrates AI, IoT, and analytics into a single, intelligent platform, providing end-to-end visibility and automation across the supply chain. Unlike conventional solutions that collect data without clear insights, this system actively analyzes, predicts, and automates to drive better outcomes. In addition, the supply chain control tower can be integrated into existing retail technology rather than serving as a replacement. It’s modular and customizable, so it does not require significant time or financial investment to implement. 

Why Supply Chain Control Tower matters:  

By unifying and standardizing data across retail systems, the supply chain control tower enables real-time monitoring of inventory, logistics, and store operations. AI-driven analytics uncover patterns and trends, allowing for advanced forecasting and proactive issue resolution. This leads to optimized inventory levels, reduced stockouts, and enhanced operational efficiency. Retailers looking to streamline their operations, reduce costs, and gain a competitive edge can leverage this technology to unlock new growth opportunities and ultimately improve customer experience.

4. Google’s Gemini Robotics brings AI into the physical world

Google DeepMind announced Gemini Robotics, advanced models that integrate language, vision, and action to enable robots to understand and interact with the physical world. These models enable robots to perform complex tasks across various environments. 

Why Gemini Robotics matters:  

World Foundation Models and robotics are becoming a greater focus area for companies like NVIDIA and Google that are pushing the boundaries of physical AI. Gemini Robotics demonstrates Google’s investment in physical AI by creating models focused on adapting to novel situations, understanding diverse instructions, and manipulating objects with fine-grain motor capabilities. Integrating Gemini’s robust language understanding with motor capabilities allows people to interact with robots in a very natural way in a variety of domains.  The investment in physical AI, like Gemini Robotics, holds the potential to revolutionize industries such as manufacturing and healthcare, by enabling robots to take on tasks that require both complicated instructions and advanced motor skills.

5. Overcome Failing Document Ingestion & RAG Approaches with Agentic Knowledge Distillation

Most Generative AI use cases today involve using RAG to find critical information outside of a language model’s training data and use this information to respond to user requests. Despite RAG’s popularity, most implementations fall short of user expectations, especially when the types of tasks are complex and require understanding information from multiple sources. The Agentic Knowledge Distillation + Pyramid Search approach aims to address the limitations of RAG by leveraging the full capability of the model at both ingestion time and during inference time when generating a response.  

Why Knowledge Distillation + Pyramid Search matters 

Agentic Knowledge Distillation works by preserving only the most critical information from documents during the ingestion process and storing varying levels of information in a knowledge pyramid which is accessible at inference time. This knowledge pyramid contains highly valuable information from the document dataset with very little noise. Some of the key benefits of this approach include using fewer tokens at inference time, reducing the cognitive load of the model so it can focus on what matters most for addressing the user task, and a stronger ability to respond to questions about information that exists in multiple documents. For more details on this approach check out Tula Masterman’s article on Towards Data Science here.   

Looking ahead 

This month’s advancements in AI reinforce a clear trajectory: artificial intelligence is no longer just a tool for automation—it’s becoming an essential collaborator in research, business, and even the physical world. As AI systems grow more capable, intuitive, and embedded in enterprise and consumer applications, the focus shifts from innovation alone to strategic implementation. Organizations that adapt to this shift will be best positioned to harness AI’s full potential, while those who lag behind risk being left in the wake of rapid technological evolution. 

With AI continuing to accelerate across industries, the question is no longer whether businesses will adopt AI, but how quickly they can do so in a way that maximizes its value.