“SaaS is dead,” Satya Nadella sparked quite a conversation. He was pointing to the paradigm shift from traditional workflows to ones powered by intelligent agents. Believe it or not, this shift is happening, and it’s time to embrace a new paradigm where ecosystems of intelligence seamlessly collaborate to make faster, smarter decisions.
We all use some form of Artificial Intelligence in our daily lives. A lot of it comes from general-purpose models like ChatGPT and Llama, which are trained on vast amounts of internet data. While that’s beneficial and powerful, it also raises concerns.
Imagine if a Large Language Model (LLM) in a healthcare setting started making recommendations based on something someone said on Reddit!
That’s why we’re seeing a shift towards building Foundation Models trained on specific domain data, hence the term Vertical Foundation Model (VFM).
I’m witnessing this shift firsthand and wanted to share some of my observations and insights as someone building LogLM, a leading foundation model for cybersecurity. In this series of seven blog posts, I’ll explore why VFMs are crucial for the next generation of business processes. The industry is already moving in this direction, with Netflix’s foundation model for recommendations and BloombergGPT for finance leading the way.
The Paradigm Shift
I still remember logging into Salesforce or NetSuite for the first time — clicking through endless, static forms to record every detail by hand. These vertical SaaS platforms did an incredible job of unifying data and human workflows into a single “system of record,” but they often felt rigid and impersonal. Today, I expect more: systems that not only gather my inputs but actually understand who I am, reason about what I need next, and adapt on the fly. That’s why we’re moving beyond traditional vertical SaaS toward “vertical foundational intelligence” — an evolution that transforms static, form‑based tools into truly intelligent partners in how we work.
Evolution from Human-Centric to AI-Driven Systems
We all would have logged into those early vertical SaaS dashboards — clicking through static GUIs to manually enter every piece of data, from customer details to order statuses. Those systems did a remarkable job of consolidating information into a single repository, but they often felt like rigid silos: each workflow trapped behind its own menu, unable to adapt when priorities shifted or new edge cases emerged.
Fast‑forward to today, and I’m amazed by platforms that not only serve human operators but also coordinate with intelligent machine agents. In my own work, I’ve watched how the Model Context Protocol (MCP) has transformed disparate services into a cohesive orchestra — automatically routing tasks, triggering data updates, and even calling AI‑driven microservices to handle complex logic. It’s like having a digital co‑pilot that knows when to hand over to you and when to take charge.
Perhaps the most profound change, though, is in how we treat the data itself. Rather than relying solely on fields typed in by users, I now see systems that leverage embeddings to build a “system of meaning” — a living map of context that learns and refines itself over time. In practice, this means my analytics dashboards can recommend next‑best actions, detect anomalies before they become problems, and continually improve their own understanding of my domain. Fundamentally, what began as human‑centric software has evolved into truly AI‑driven intelligence.
Emergence of Vertical Foundation Models (VFMs)
When Geoffrey Hinton’s students first trained a deep neural network across multiple GPUs, they demonstrated the power of harnessing parallel physics in practice — and set the course for today’s compute revolution. As an engineer, I’ve watched our entire computing stack transform: GPUs have given way to data centers rebuilt for massive parallelization, and the breakthrough transformer architecture has unlocked that raw power for language understanding.
But powerful hardware is only half the story. The real leap from static, human‑centric platforms to dynamic, AI‑driven ecosystems comes from two forces: protocols like the Model Context Protocol (MCP) and the rise of foundation models. Born out of Natural Language Processing (NLP) research, foundation models stand on two pillars: transfer learning and scale. Transfer learning is the magic that lets us adapt a single pretrained model to countless tasks, but only scale — training on truly massive corpora makes these models so impressively capable. I’ve spent weeks watching self‑supervised pretraining ingest oceans of text, teaching the model a deep, general understanding of language. Then, with a few lines of fine‑tuning code, that same model learns to excel at my specific use case — cutting adaptation time from months to hours. However, these models still lack a true understanding of the specific domain and will often fall back on their general knowledge (sometimes hallucinating) when faced with tough, domain‑specific questions.
Instead of relying on fine‑tuned generalists, industry has started building models trained on large corpora of domain‑specific data, highlighting the trend toward Vertical Foundation Models (VFMs). Bloomberg GPT, a 50 billion parameter model trained on financial data from scratch, Netflix built a foundation model for recommendation systems, our very own, LogLM, a foundation model for Cybersecurity. Each is crafted to understand its domain down to the smallest detail, enabling agents to act with precision and insight that wasn’t possible before. Now, these systems do far more than our old “system of record” tools ever could.
Graphic User Interfaces (GUI) to Language User Interfaces (LUI)
Another fundamental shift in human-machine interaction that we are seeing is the transition from GUIs to language-based interfaces (LUI). Remember going through endless, form-based screens just to get a simple status update — sigh, it felt like jumping through hoops. Now, with the rise of foundation models, I can simply get my next movie recommendation by asking, “Recommend me a classic sci-fi movie,” and get an instant context-aware answer, probably Star Wars (yours might be different 😆). Or if you are too much into your work, you can simply ask, “Find similar past issues to TICKET‑2001” (A nod to 2001: A Space Odyssey 😜 )
These Language User Interfaces (LUIs) transform interactions into natural conversations that adapt to who I am and what I need. Under the hood, state‑of‑the‑art reasoning engines like OpenAI’s O1, Claude 3.7, and DeepSeek‑R1 power these experiences, not only explaining their recommendations but also surfacing relevant documents via Retrieval‑Augmented Generation (RAG). It feels less like using software and more like collaborating with an intelligent teammate.
Impact on Business Workflows
Those days when teams spent hours triaging edge-case tickets by hand, switching between spreadsheets, dashboards, and email threads, to piece together a single customer issue and root-cause analysis. Today, Vertical Foundation Models let us collapse that chaos into a single, domain‑aware workflow. By tapping into vast, specialized knowledge bases, VFMs can spot those tricky outliers and suggest precise next steps, freeing us from repetitive tasks and letting us focus on higher‑value work. Better yet, we can plug in intelligent agents to carry out routine actions — everything from sending compliance reports to updating incident logs — while the Modern Context Protocol (MCP) quietly orchestrates data flows and task handoffs behind the scenes. And because feedback loops are built in, each interaction sharpens the model’s understanding, ensuring the system continually adapts to our evolving needs. In short, VFMs, agentic workflows, and MCP aren’t just incremental improvements — they’re a fundamental redesign of how we build and use applications.
What’s Next?
Vertical Foundation Models are already here, bringing deep domain understanding and enabling more informed decision-making. However, before delving deeper, let’s understand general-purpose models and highlight some of their challenges.
In the next blog post, I’ll examine general-purpose foundation models and their common pitfalls, and explain how Vertical Foundation Models will transform industry by industry. I will break down everything in simple terms, no jargon — in fact, I hate jargon.
— Follow along for more!
