SaaS is a dead man walking. The entire multi-trillion dollar software-as-a-service model, built on passive tools that wait for your command, is being rendered obsolete. Its replacement is active, autonomous, and already here. It’s called agentic AI.
For the past decade, the game was to build a slick UI, solve a workflow problem, and capture users in a recurring revenue loop. You, the user, did all the work. Your CRM, your analytics platform, your email tool — they were all just glorified spreadsheets, dumb databases you had to prod into action. That era is over. The next £100bn opportunity for UK founders lies in building systems that do, not systems that wait to be told.
This isn't just another tech trend. It's a fundamental platform shift, as significant as the move from desktop to cloud. We're moving from a world of applications to a world of agents. I'm not just an observer in this; I'm a builder. At Creative Marketing Group, and through my work as <a href="/kasim-javed">Kasim Javed</a>, I’m building the agentic products (like our own Autoemails system) designed to replace the broken, labour-intensive marketing workflows we've all been forced to endure. The map of this new world is what I call the Agentic Stack.
To understand the opportunity, you need to understand the architecture. The Agentic Stack is the hierarchy of technologies that enables autonomous AI. It’s composed of four distinct, interdependent layers. Founders and investors who fail to grasp this structure will be left building castles in the sand.
Layer 1: The Foundation (Infrastructure & APIs)
This is the bedrock. The raw intelligence. The large language models (LLMs) and multi-modal models provided by a handful of colossal, capital-intensive players. Think OpenAI (GPT-4), Anthropic (Claude 3), and Google (Gemini). These are the utility companies of the agentic age. They provide the electricity, but they don't build the appliances.
For a UK founder, competing here is financial suicide. The capital required to train a frontier model from scratch runs into the hundreds of millions, soon to be billions. The game here isn’t to build a rival LLM, but to be a savvy, strategic consumer of their APIs. Your competitive advantage comes from being model-agnostic, ready to swap out one provider for another as performance and cost dictate. We, for example, build our systems to be compatible with multiple foundation models, arbitraging them for the best combination of speed, intelligence, and cost for a specific task.
Layer 2: The Connective Tissue (Orchestration & Tooling)
If Layer 1 is the power plant, Layer 2 is the national grid. This is the realm of orchestration frameworks and developer tools that allow you to chain model calls, connect them to data sources, and give them access to other tools. The dominant names here are open-source projects like LangChain and LlamaIndex.
These tools provide the plumbing required to build complex agentic workflows. They handle prompt engineering, state management, and the crucial ability for an AI to interact with external APIs (like checking the weather, booking a flight, or accessing a customer database). This layer is rapidly evolving, but it’s fundamentally about enablement. It provides the standardised parts that developers use to build more sophisticated applications further up the stack. While there are commercial opportunities here, the real value is captured in the layers above.
This is where things get interesting. Layer 3 is where the concept of an agent truly comes to life. It's about creating frameworks that allow multiple AI agents to collaborate to solve complex problems. Think of it as building a digital project team. Frameworks like CrewAI and Microsoft's Autogen are pioneering this space.
Instead of a single AI call, you might have a 'Researcher' agent that gathers information, a 'Writer' agent that drafts a report, and a 'Critic' agent that refines the output. This multi-agent approach allows for a level of sophistication and reliability that a single monolithic call simply cannot achieve. At our agency, we don’t just ask an AI to