Extreme Connect 2025
May 19-22
Join us in Paris, France.
Learn More
It seems that wherever you look in IT and beyond, companies are talking about their new AI systems. But not all AI is created equal. Some companies take the easy route by bolting AI onto existing systems and calling it innovation. At Extreme Networks, we’re taking a different approach. We built Extreme Platform ONE and its AI capabilities as a true AI-centric platform, designed from the ground up to reimagine how IT teams interact with their networks, and more.
To get the full story of how this platform came to life, I sat down with Markus Nispel, our AI leader. Our conversation covered the tough decisions, technical hurdles, and lessons learned in building a system that isn’t just another AI assistant, but a transformational shift in network operations.
From the very beginning, the team had a choice: should we bolt AI onto our existing systems, or build new with an AI-first mindset? The first option was tempting, and it certainly would have been faster and easier. But it also would have been superficial, limiting AI’s ability to truly transform every experience we deliver—our mission for our AI activities. We quickly realized that without building a platform for the era of pervasive AI, we’d be capping what’s possible in the future.
Instead, we took a step back and applied our ARC Framework, which we use to evaluate (AI-driven) innovation. The first step was acceleration—using AI to optimize and speed up existing workflows to make humans more efficient. That alone would have been valuable, but we saw an opportunity to go further. Replacement became the next goal: redesigning how IT teams interact with our systems. This would make AI not just a tool but an integral part of daily operations, rethinking workflows and processes with AI capabilities in mind, and freeing them from time-consuming manual tasks to focus on higher-value work.
Then came the real ambition: Creation of new business value and deliver new, differentiated experiences. AI isn’t just about doing the same things faster; it’s about enabling entirely new ways of working that wouldn’t be possible otherwise. We think of AI as a transformational inflection point, a general purposed technology (GPT), just as mobile and cloud technologies before it. Businesses like Google, Uber, and Amazon created whole new markets or categories, but they were only possible because of those new innovations. So we asked: what new IT services, what new ways of managing networking and security, could only exist becaue of AI? How can we harness AI to create new business value in a way that simply wasn’t possible without it?
That’s when it became clear: this couldn’t just be an add-on. It had to be reimagined.
Before we started building, we had to tackle some fundamental questions. First, what does an AI system need to be truly useful? Data was the obvious answer, but raw data alone isn’t enough. The real challenge was making sense of it: connecting disparate sources, structuring it properly, and ensuring accuracy so that AI-generated insights could be trusted. And that is true for both structured and unstructured data across the enterprise.
Then came the issue of architecture. AI models and tools are evolving rapidly, and we didn’t want to be locked into a single approach that could become outdated in a year. That led to a key design principle: modularity. Extreme Platform ONE would be built in a way that allowed us to swap in new AI models, integrate emerging machine learning tools, and continuously refine its capabilities without needing a complete overhaul. It also had to be flexible and adaptable to new data sources and types, leading to a lakehouse approach within the AI core of the platform.
Just as important as the backend was the way users would interact with AI. It wasn’t enough to make AI powerful—it had to be intuitive. That’s where another design principle emerged: user-first integration. The AI interaction models had to fit seamlessly into IT teams’ existing workflows, enhancing their work rather than forcing them to adapt to an unfamiliar system.
It meant building a platform that has unified intelligence at the core while presenting multiple ways of interacting with it to meet users where they are.
One of the biggest challenges in building AI Expert initially with its conversational interaction model was that AI itself is evolving at an unprecedented pace. We needed to create a stable, scalable system while knowing that the underlying technology would keep shifting. We started with a simple but powerful idea: what if you could talk to your documentation?
The first iteration of AI Expert was a retrieval-augmented generation (RAG) system, designed to pull insights from manuals, knowledge bases, and past incident reports. But we quickly realized that a language model alone wasn’t enough. AI needed structured knowledge, not just free-text search, to deliver high-quality answers.
The real test came when we had to decide when AI was good enough to ship. We weren’t aiming for perfection—AI is inherently probabilistic, and chasing absolute accuracy would mean never launching at all. Instead, we set clear benchmarks for quality and usability. Once AI Expert met those targets, we moved into a tech preview phase, putting it in the hands of real users and all our employees as well as trusted partners to refine its capabilities in the field.
One thing became clear during this process: a modular approach was the right decision. As AI technology evolved, we were able to swap in better models and adjust our approach without needing to rebuild the platform. The flexibility of AI Expert’s architecture ensured that it wasn’t just a product for today, it was a foundation for ongoing innovation leading to AI canvas and an agentic system underneath
One of the most interesting insights from the tech preview was that conversation alone wasn’t enough. Users wanted AI to do more than just provide information; they wanted it to take action. That realization led to a major evolution of AI Expert and AI canvas: agentic AI. Instead of just answering questions, AI Expert needed to interact with live network data, interpret real-time events, and make intelligent recommendations. It had to move from being a passive source of information to an active participant in network operations.
This shift also influenced how users interacted with AI. Initially, AI Expert functioned primarily as a conversational, “chat-based” system, but we found that complex troubleshooting often required a more structured approach. That led us to introduce a canvas-based model, allowing users to visualize issues, create and explore different AI-generated insights, and make decisions collaboratively with AI rather than just querying it like a search engine. This also allows them to build complex dashboards, which can be shared beyond IT to unlock insights for stakeholders throughout the company.
As trust and adoption increases, we are going to see more and more tasks being fully automated, with or without a human in the loop.
One of the defining reasons we built AI right into the platform rather than an add-on was the need to work with live data. AI-driven operations require more than just historical analysis; they need real-time awareness. Trying to bolt AI onto existing network tools would have created performance bottlenecks and fragmented insights. Instead, we designed it to consume streaming data at the base layer, treating real-time analysis as a foundational capability rather than an afterthought.
That decision meant AI could process events as they happened, providing instant recommendations and automating responses where appropriate. The result was a system that could not only diagnose issues but also predict and prevent problems before they escalated.
Early in the development process, we intended to build everything ourselves. But as we scaled up, it became clear that our expertise was in differentiation, not in reinventing the AI research being done by industry leaders.
That’s why we made the strategic decision to partner with Microsoft, leveraging their AI services for foundational model development and AI safety and security. Meanwhile, we focused our efforts on what made our solution unique: its deep integration with Extreme Networks’ ecosystem and its real-time operational intelligence.
This partnership allowed us to stay ahead of the curve while keeping our development efforts focused on what truly mattered: delivering value to customers rather than getting bogged down in building models from scratch.
Perhaps one of the most unexpected insights from building a highly accurate AI system was the importance of metadata. AI performs exponentially better when it has access to high-quality metadata, which applies to both structured and unstructured data sources. This realization changed how we approached product development. Rather than treating metadata as an afterthought, we began embedding AI into the development process itself, ensuring that data was properly tagged and structured from the moment it was created. AI wasn’t just a tool for operations—it became a cross-functional asset, improving everything from documentation to network analytics, and taking inspiration from the data mesh architecture across the organization to make this a success.
As AI continues to advance, we’re in a position to integrate the latest innovations seamlessly. For this reason, Extreme Platform ONE isn’t just a product, it’s an evolving platform.
From deeper automation to even more sophisticated real-time intelligence, we’re only scratching the surface of what AI can do in network and security operations. The foundation we’ve built ensures that it will continue to adapt, expand, and redefine the way IT teams manage their environments.
And, most importantly, it’s proof that when AI is done right—not as a bolt-on feature, but as a fundamental shift in how we approach technology—it has the power to transform everything.
Continue the journey with Extreme Platform ONE with our six-part deep-dive webinar series. Register now to explore innovative AI features that will transform your IT operations.