A new model every week. A new platform every month. A new category that has become “AI-first” every quarter.
Look at today’s AI landscape and the first reaction is usually overwhelm. There is no overview. There is no obvious winner. There is no standard. And the pace of change is so high that every choice you make now seems outdated six months later.
The logical conclusion appears to be: wait. Pick the most familiar name. Go for the all-in-one platform from your existing vendor.
That is exactly the wrong conclusion.
Fragmentation is not a temporary problem the market will solve. It is the structural reality of a technology ecosystem under construction — and how you deal with it decides whether AI becomes a strategic advantage or an expensive mistake.
How fragmented is it, really?
The numbers are sobering. Every major SaaS vendor has added an AI layer to its platform: Salesforce with Agentforce, ServiceNow with AI agents, SAP, Oracle, Workday — all shipping domain-specific AI functionality. The result: every application category now has an AI layer, but those layers do not talk to each other. Nearly three quarters of enterprise applications exist in isolation.
At the infrastructure end, hundreds of models, orchestration frameworks, memory solutions and evaluation tools compete for attention. The agentic AI stack now has seven recognisable layers — foundation model, agent runtime, orchestration, tooling, observability, governance. Each layer has dozens of competing solutions, each with its own APIs, pricing and roadmap.
67% of organisations say they are actively steering away from dependency on a single AI vendor. That sounds sensible. But 45% of those same organisations say vendor lock-in has already prevented them from adopting better tools. Intent and practice drift apart structurally.
The real risk is not the chaos — it is the seductive simplification
The most dangerous response to fragmentation is not passivity — it is the reach for the most familiar name.
Microsoft, with Azure OpenAI and Copilot, represents the deepest enterprise AI lock-in currently on the market. That is not necessarily a criticism — it is a description of a strategic choice with consequences. Build your AI workflows on a proprietary orchestration layer and you accept switching costs that grow exponentially with every new agent workflow. API dependency means your architecture bends to the design choices of a single vendor, not to the needs of your organisation.
The opposite extreme is just as costly. Assemble tools at random, without architectural principles, and you build an AI zoo rather than an AI system. Integration, maintenance and governance costs rise. Value lags.
Organisations running more than twelve AI tools show measurably declining returns — not despite the tools, but because of the disconnection between them.
Fragmentation as a design premise
The turning point in thinking is this: stop seeing fragmentation as a problem to solve. Start seeing it as a design condition to architect against.
What does that mean in practice?
It means choosing a layered, modular architecture over a monolithic platform. An architecture that separates the three core functions of an AI system — reasoning, acting, remembering — into independent layers that can be replaced without rebuilding the rest. Not because you plan to switch vendors constantly, but because you do not want to be trapped when a better option appears.
It means recognising that you do not need one model — you need a strategy for model use. Which model for which task type? Which orchestration layer handles routing? Which memory solution secures the context your organisation is accumulating? These questions have no universal answers — they have your answers, grounded in your processes and your risk tolerance.
And it means treating the EU AI Act and adjacent governance requirements not as compliance overhead but as an argument for a clear stack. Vendor-locked AI systems make consistent oversight of decisions structurally harder.
The three layers that are always there
However complex the landscape becomes, every robust AI implementation eventually needs three layers. Not as an abstract model, but as a concrete architectural choice:
The model layer — which model does the reasoning? This is not one answer for everything. It is a deliberate choice per task type, with criteria for quality, cost and control.
The harness layer — the orchestration layer that connects the model to your data, tools and processes. This is the control plane of your AI system: it handles routing, validation, fallback, and governance. Without a strong harness, a model is an isolated text generator — however capable.
The memory layer — the persistent storage layer that retains context, knowledge and decision history across sessions. This is what turns an AI system into institutional memory. Without a memory layer, every conversation starts over and the organisation learns nothing.
These three layers are the structuring choice in a fragmented landscape. Not technical details for IT to sort out later, but strategic design decisions that determine the value of your AI investment.
What this asks of you as a decision-maker
Fragmentation in the AI landscape does not require technical expertise from leaders. It requires architectural thinking. The question is not “which tool do we pick?” but “which layers do we build, and what principles govern each one?”
That thinking sounds abstract, but it is urgent. The organisations that define their AI architecture clearly now — with deliberate choices per layer — are building a foundation that scales. The organisations that rush to deploy a platform without that foundation are building a problem that will be expensive to correct later.
Success with agentic AI is 20% technology and 80% change management and strategic choice. The chaos in the landscape is real. But the chaos is also the arena where organisations with a clear architectural view make the difference.
What follows
The next three posts cover each of the three layers in detail. We start where the most misunderstanding exists: the model.